Renewable and Sustainable Energy Reviews 160 (2022) 112128

Contents lists available at ScienceDirect

Renewable and Sustainable Energy Reviews

journal homepage: www.elsevier.com/locate/rser

Data-driven probabilistic machine learning in sustainable smart energy/

smart energy systems: Key developments, challenges, and future research
opportunities in the context of smart grid paradigm
Tanveer Ahmad a, *, Rafal Madonski a, Dongdong Zhang b, Chao Huang c, d, Asad Mujeeb e
a
Energy and Electricity Research Center, International Energy College, Jinan University, Zhuhai, Guangdong Province, 519070, China
b
School of Electrical Engineering, Guangxi University, Nanning, 530004, China
c
School of Computer and Communication Engineering, University of Science and Technology Beijing, Beijing, 10083, China
d
Shunde Graduate School, University of Science and Technology Beijing, Guangdong, 528399, China
e
Department of Electrical Engineering, Tsinghua University, Beijing, 100084, PR China

A R T I C L E I N F O A B S T R A C T

Keywords: The current trend indicates that energy demand and supply will eventually be controlled by autonomous soft­
Data-driven probabilistic machine learning ware that optimizes decision-making and energy distribution operations. New state-of-the-art machine learning
Energy distribution (ML) technologies are integral in optimizing decision-making in energy distribution networks and systems. This
Discovery and design of energy materials
study was conducted on data-driven probabilistic ML techniques and their real-time applications to smart energy
Big data analytics and smart grid
systems and networks to highlight the urgency of this area of research. This study focused on two key areas: i) the
Strategic energy planning and smart
manufacturing use of ML in core energy technologies and ii) the use cases of ML for energy distribution utilities. The core energy
Energy demand-side response technologies include the use of ML in advanced energy materials, energy systems and storage devices, energy
efficiency, smart energy material manufacturing in the smart grid paradigm, strategic energy planning, inte­
gration of renewable energy, and big data analytics in the smart grid environment. The investigated ML area in
energy distribution systems includes energy consumption and price forecasting, the merit order of energy price
forecasting, and the consumer lifetime value. Cybersecurity topics for power delivery and utilization, grid edge
systems and distributed energy resources, power transmission, and distribution systems are also briefly studied.
The primary goal of this work was to identify common issues useful in future studies on ML for smooth energy
distribution operations. This study was concluded with many energy perspectives on significant opportunities
and challenges. It is noted that if the smart ML automation is used in its targeting energy systems, the utility
sector and energy industry could potentially save from $237 billion up to $813 billion.

overload allowing the whole grid to suffocate or be damaged. Grid op­

1. Introduction erators have to compensate energy providers for making changes to their
supplies (i.e., power system infrastructure), saving German consumers,
Today, while countries seek to restructure their energy strategies and for example, about $553 million a year [4].
make cleaner energy more dependent, one major challenge remains [1]. Besides, electric utilities generate excessive carbon dioxide emissions
Both wind and solar power are, by definition, intermittent nature of when their excess electricity is dissipated [5]. All this contributes to the
sources of electricity [2]. The power output of a solar panel or wind complexities of accurately forecasting the industry’s health status of
turbine is never constant; it is determined by external variables such as energy distribution infrastructure. In general, keeping the demand and
cloud cover intensity, solar radiation, and wind speed, all of which are supply of energy balanced will turn into a constant operational and
uncontrollable [3]. When wind and solar farms generate less electricity, technological struggle. This leads us to the potential that ML has and to
grid operators must switch to traditional power plants to balance energy the major impact it could have on the energy spectrum as a whole [6].
supply and demand. Alternatively, after 90% of their electricity needs Even though ML is in its early stages of deployment, it could funda­
for the day have been met on windy and sunny days, operators must mentally transform how we communicate with resources [7]. Its effect
decrease gas-fired and coal power plant production to prevent a power could be far-reaching in renewable energy distribution and forecasts and

* Corresponding author.
E-mail addresses: [email protected], [email protected] (T. Ahmad).

https://doi.org/10.1016/j.rser.2022.112128
Received 23 March 2021; Received in revised form 5 December 2021; Accepted 10 January 2022
Available online 5 March 2022
1364-0321/© 2022 Elsevier Ltd. All rights reserved.
T. Ahmad et al. Renewable and Sustainable Energy Reviews 160 (2022) 112128

Abbreviations HEM Home energy management

EV Electric vehicles
ML Machine learning ESS Energy storage system
TWh Terawatt-hours RES Renewable energy source
REmap Renewable energy map US The United States of America
PV Photovoltaics AEPC Active building energy performance contracting
ROI Return on investment CO2 Carbon dioxide
DOE Department of Energy USD The United States dollar
IBM International Business Machines IoT Internet of Things
ABB ASEA Brown Boveri SCADA Supervisory control and data acquisition control system
BP British Petroleum RTUs Remote terminal unit
AI Artificial intelligence GIS Geographic information system
SVM Support vector machine CLV/LTV Consumer lifetime value
CNN Convolution neural networks TEASER Energy analysis and simulation for efficient retrofit
DBNN Deep belief neural networks RL Reinforcement learning
DBM Deep Boltzmann machine EDLC Electric double-layer capacitors
RNN Recurrent neural networks SMES Superconducting magnetic energy storage
DRL Deep reinforcement learning V2G Vehicle-to-grid
LoRa Short for long-range TESS Thermal energy storage system
PCC Point of common coupling

the implementation of smart grids. sustainable future in Sect. 6. Recent progress on properties and discov­
The aim and objective of this study: Right now, ML is considered a “hot ery of ML is described in Sect. 7. Finally, Sect. 8 concluded the con­
topic” in a variety of research fields, and it is currently the most rapidly ducted study.
developing sector in high-tech. Data processing and interpretation are Fig. 1 shows the results of an analysis of search queries on ML in the
becoming increasingly important, even for engineers, due to energy subject area, article type, and area of research in engineering and energy
digitalization. Smart grids are a term used to describe various new data- distribution. The analysis is divided into three major parts: including i)
based services in renewable energy supply, marketing, storage, and document by subject area on ML models (e.g., published research arti­
usage. This study covers recent advances and fundamental ML tech­ cles in all areas of research but with the use of ML models), ii) document
niques in energy distribution. ML assists in the fast and efficient pro­ by subject area on ML in engineering (e.g., published research articles in
cessing of this data in energy distribution (e.g., the energy distribution is the engineering field); iii) and document by subject area on ML in en­
the last stage of energy delivery, it carries out the energy from trans­ gineering and energy (e.g., published research articles in the engineer­
mission systems, conventional grid or smart grid infrastructure, indi­ ing field, particularly in energy). Overall, the United States is leading in
vidual consumers, transmission and distribution infrastructure, energy publishing research articles in all fields. China is number one in the
devices and materials, big data analytics, etc.). To start, we identify the World for publishing research articles in the engineering and energy
numerous potential challenges that ML tries to solve, review recent fields. Fig. 1 shows the data clusters result from Scopus.
advances in the field, and analyze ML’s effect on the energy sector. The
introduction of different ML model classes to solve such complex
problems is briefly examined. Eight various aspects of ML in core energy 1.1. Energy conversation: moving 100% towards clean and renewable
technologies are discussed. Five use cases of ML in energy distribution energy
are briefly covered. The used software for ML models in energy distri­
bution, challenges, ML opportunities towards a smart and sustainable Globally, a total of 376 TW-hours (TWh) of renewable energy was
future, and recent progress on discovery and properties of ML models produced in 2018, a rise of 6.1% relative to the previous year (2017) [8].
aim to strengthen the core theme of this study. Wind and solar production rose by 11% and 28%, respectively, in 2018
The Framework and Structure of this Study: The description of the [8]. In 2018, Asia accounted for much of the increase in renewable
performed study has been arranged into eight sections. The paper opens energy production, increasing generation by a total of 219 TWh. Asia’s
with a general introduction of the study, including four subsections. The volume of world renewable generation also rose, increasing to 40% of
process of energy conversion (supporting shift towards 100% renewable overall renewable generation. North America and Europe together hold
energy) is discussed in the first subsection. The main idea and concept of 40% of the shares, led by Eurasia (5%) and South America (12%) [8].
ML, historical overview of ML methods, list of ML models and their The United States’ clean energy initiative is becoming more asser­
usage in the smart energy systems, as well as the role of ML in the dis­ tive: more groups target a rise in renewable energy use and carbon
tribution of energy are described in Sects. 1.2 to 1.5, respectively. Then, emissions decrease [9]. Many states have recently declared a target to
Sect. 2 follows the core concept of energy technologies using ML, which increase their current renewable portfolio requirements to incorporate a
discusses the use of ML in various areas. The use of ML for energy dis­ 100% renewable energy standard (e.g., New Mexico, Washington,
tribution utilities is elaborated in Sect. 3, with its five subsections Nevada, and Colorado) [10]. A rising number of power companies are
dedicated to accurate energy price forecasting, keeping the merit order now making enormous commitments to reduce greenhouse emissions
of accurate energy prices, predicting consumer life values, probability and increase sustainable energy usage [11]. At least ten publicly listed
assessment of winning consumers, and making good offers to the energy utility companies have declared 100% decarbonization targets, and a
consumers. Sect. 4 highlights various uses of ML tools for energy dis­ large portion of these utilities have established goals to minimize carbon
tribution systems and introduces different software used by organiza­
tions that deal with ML techniques. Sect. 5 describes the analysis of the
existing challenges of ML in the context of energy distribution systems.
We then go on to identify ML opportunities towards a smart and

2
T. Ahmad et al. Renewable and Sustainable Energy Reviews 160 (2022) 112128

Fig. 1. ML search results in engineering and energy distribution were examined in terms of subject, article type, and research area.

emissions by much more than 80% below 2005 levels by 2050 [12]. reducing carbon emissions that is becoming a bigger priority today. The
Fig. 2 demonstrates1 the proportion of the various renewable energy United States Department of Energy (DOE) collaborated with Interna­
sources in overall final energy use worldwide in 2010 as well as for 2030 tional Business Machines (IBM) to create a Watt-Sun software, which
with the additional increase arising from the renewable energy map measures vast quantities of weather data obtained from a large range of
(REmap) opportunities [13]. The primary source of green energies data sources and sites [16].
would be bioenergy, which can produce both heat and power and Another key advantage of moving 100% towards clean, renewable
vehicle fuels. By 20% of renewable energy usage in REmap 2030, many energy is the energy transition process will generate additional 11
natural resources of liquid, solid, and gaseous biomass constitute 61% of million jobs in the energy sector by 2050 (see Fig. 3). The transformation
the total [13]. However, as previously reported, most of the transition will massively improve overall jobs in the energy sector by proactive
would be the move from conventional to modern technologies and fuels policies. The change to green energy sources would produce more en­
[14]. By 2030, REmap strategies would boost both the total volume and ergy employment than the fossil fuel industry would lose. The renewable
share of wind power usage, while wind power deployment would sur­ energy map case will lead to a loss of 7.4 million workers from fossil
pass hydropower. Solar-PV will make up a significant portion of overall fuels by 2050, with a net increase of 11.6 million jobs generated by 19
power production as well. Though all REmap strategies are deployed, million new jobs for renewables, energy conservation, grid improve­
solar thermal heat can produce almost ten times more electricity than ment, and energy flexibility. Education and training policies would
today’s building and industrial sectors [13]. require the expertise and value-creating capabilities of these industries
There is enormous potential for the usage of ML for power companies to fulfill the needs for the human capital of renewable energies and
and investors to deploy more reliable and profitable processes that energy-efficient industries in the rapid growth process. A change that
improve return on investment (ROI) and help the energy transition [15]. would lead to just and fair social and economic effects will prevent
Although there are still various implementations of ML, it can speed up resistance. One of the most significant possible gains is to transform the
the phase of positive transition in the energy industry and business socioeconomic environment.
Countries are just starting to realize the promise that ML brings and
are integrating them into their policies. For example, China (adopting a
nationwide smart grid using the current integrated technologies in AI),
1
Renewables: RISE Regulatory Environment, which countries have an Australia (spending $25 million on ML and heading out to focus on
enabling environment for investment in renewable energy? 25 Jun 2017, https energy in comparison to other countries), the United Kingdom (spending
://www.seforall.org/data-stories/renewables-rise-regulatory-environment.
$119 million into the ML and robotics development programs in extreme
Accessed: 03/12/2021.

3
T. Ahmad et al. Renewable and Sustainable Energy Reviews 160 (2022) 112128

Fig. 2. An overview of global renewable energy usage in 2010 and the renewable energy map 2030 projections are divided into technologies and sectors.

climates, for application on ML and robotics innovations that are widely execute the algorithms, compare the outcome with the desired re­
utilized in different conditions like offshore gas and oil fields), Italy sults it could have generated, and make changes to the algorithm’s
(establishing 41 million smart meters for a project that began in 2001. weights and prejudices as needed, repeating the process before the
Consequently, Italy seems to have the most developed smart grid in algorithm returns the right or accurate answer most of the time.
the country), and the United State (the energy department proposed $30 4) Improving and using the model: Finally, the method is used with
million in support for ML) spending much to incorporate the ML tech­ updated information and, if possible, to see if it becomes more ac­
nologies in the energy industry.2 Also, large energy companies (e.g., curate and more effective over time. New data will come from
ABB, Shell, Exxon Mobil, British Petroleum (BP), VIA, Stem, Pranav, wherever the problem to be solved calls for it.
Schneider Electric, Google, and DeepMind, etc.) are focusing on ML
models. For example, integration between an ML technique and a major The ML models fall into four main classes, including i) supervised
hydroelectric utility resulted in a 10% reduction in regular maintenance ML; ii) unsupervised ML; iii) semi-supervised ML, and iv) reinforcement
costs and an improvement of 2% in ABB company output.2 BP company ML. Table 1 shows a detailed overview and class functions of probabi­
used an amazon cloud storage system was built with lubricant enterprise listic ML models in energy distribution. Further, detailed analysis of
planning applications, resulting in a 40% quicker response time.2 different machine models, main categories/subcategories, and real-time
implementation with codes are given in Ref. [19].

1.2. Introduction to machine learning

1.3. Historical overview of ML models
ML is a subset of artificial intelligence (AI) that applies mathematical
methods to data to learn the machine how to enhance its performance According to Fig. 4, the first stage started in the mid-1950s and
and accuracy without the support of external guidance [18]. ML focuses continued until the early 1960s. Before that, the idea of ML was devel­
on applications that benefit from their previous experiences and oped, with the primary focus on computer translation of logical thought.
improve forecasting and decision-making over time. We use probabi­ There was rapid development in symbolic theory, information engi­
listic ML for two reasons. Firstly, it’s the most successful approach to neering, and expert systems that arose as essential fields of study. Since
making decisions better under uncertainty. Secondly, probabilistic 1960, ML study had phases of success accompanied by periods of stag­
modeling is the main language for other scientific and engineering nation. After completing the concrete evidence of all theorems in the
fields, facilitating its utility as a unifying framework [18]. book Principia Mathematica, Herbert Simon and Allen Newel developed
How does ML work? To construct a ML program, there are generally the Logic Theorist, an automated theorem proving device. People slowly
four following steps: discovered.
That rational thinking capacity was not enough to achieve computer
1) Prepare and select the training dataset: Training data is a compilation of intelligence. Consequently, the analysis of AI inevitably reached its
data that accurately represents the ML algorithm used to solve the second level [20]. The second stage started in the 1970s and continued
issue it was developed to address. until the 1980s [20]. Scholars began to consider whether the informa­
2) Select a method to be applied to the training set: Algorithm classes vary tion could be gathered and converted to a computer machine. In the
depending on whether they operate with unlabeled or labeled data 1990s, the third phase started. The ML discipline arose to cope with the
and on whether they solve a problem that is unlabeled or labeled. information acquisition bottlenecks. Specifically, in 2006, ML joined the
3) Training the algorithms to build the model: Training the algorithms is an age of cognitive intelligence, which focuses on collecting big data,
iterative procedure. First, we set variables to some values, then we modern theoretical techniques, more efficient machines, and autono­
mous learning. These latest AI generations, such as cross-border inte­
gration, deep learning, SI, autonomous control, and human-computer
2
Machine Learning and Renewable Energy, https://www.pangea-si.com collaboration, all carry out new features. The success of mathematical
/machine-learning-and-renewable-energy/, Accessed: 03/12/2021. and computational learning theory has led to ML models’ constant

4
T. Ahmad et al. Renewable and Sustainable Energy Reviews 160 (2022) 112128

Fig. 3. The energy transition process will generate

additional 11 million jobs in the energy sector by
2050 (*For 2016, there are no employment estimates
for grid enhancement and energy efficiency. **Grid
enhancement jobs apply to transmission and distri­
bution grids and energy flexibility jobs, provided in
the operation, development, and maintenance of
infrastructure to allow the renewable energy inte­
gration of renewable energy sources into the grid.
***Both employment in the fossil fuel sector,
including mining, refining, and consumption, are
included) [17].

availability; representative ML techniques include deep learning, rein­ memory cells that can keep valuable information for the long term. It
forcement learning, decision trees, and support vector machine (SVM)— shows greater efficiency in dealing with load based on long-term de­
the history of ML from 1957 to today as seen in Fig. 4 (right side). In the pendencies, high volatility, and uncertainties [24,25]. Fully Bayesian
energy industry, the word ML is commonly used in conjunction with AI inference on the other hand can be used to model selection choosing
and is significant. On the other side, ML and AI are two separate entities both predictive and evidence frameworks. In this regard, results from
because ML covers a portion or part of AI, but not all of it. some studies have proved that the evidence framework outperforms the
predictive method, which shows data overfitting [26].
The probabilistic ML models for smart energy systems also include
1.4. List of machine learning models and their usage in smart energy
Binary and Bernoulli distributions, Multinomial and Categorical distri­
system
butions, and Univariate Gaussian distribution as described in Table 1.
The binomial distribution is a probability distribution that summarizes the
Several ML models listed in Table 1 are vastly used in smart energy
likelihood for a particular value among one or two independent values
systems. Here we will briefly elaborate on each model in detail.
under a given set of parameters. Binary and Bernoulli distribution models
Supervised ML, Unsupervised, and Reinforcement Machine learning: In
have played a key role in the probability distribution of the smart energy
the context of ML techniques, these terms are brought up with different
power system. Researchers have adopted them for the Plug-in Electric
definitions according to their use in different areas.
vehicle PEV model to determine the probability distribution of EV
Supervised learning: It has got much attention in image classification,
charging patterns at different times of use [27,28]. Recently, most
speech recognition, and machine translation. Power systems and smart
studies have adopted grey Bernoulli models to reduce randomness and
energy systems play one of the key roles in data and load forecasting by
uncertainty in the energy systems; hence, making predictions requires
providing input/output pairs to the ML algorithm. The ML algorithms do
fewer operational data and analysis, especially in predicting the
a function mapping on those pairs of input to their desired output which
long-term development trend over data [29].
is also termed as a regression process. The key example is supervised
Logistic Regression: One kind of supervised ML used to model condi­
vector machines (SVM) learning models used to analyze the electricity
tional probability. It is a univariate model, but it can be multiclass lo­
data generated by distributed generation and conventional generators
gistic regression. At the multi-stage decision process, the decision tree
[21].
method has been commonly adopted to make a binary decision at each
Unsupervised learning only requires feeding input into the ML algo­
stage. Logistic regression, binomial and multivariate regressions
rithm without any corresponding output (no dependency on output).
brought different applications for smart energy system prediction such
Here, the clustering-based technique clusters the data into similar
as power black outage prediction, load forecasting, etc. [30].
groups to hunt for hidden data groups or patterns of data. Grid-based
Common Univariate Distribution (CMD): CMD in ML is usually dis­
clustering for energy systems can be incorporated for optimal energy
cussed separately by probability studies. Some of them are Student t
distribution from the nodes. This energy clustering ensures the fast
distribution, Beta distribution, Cauchy distribution, Gamma distribu­
processing of unsupervised electricity data [22].
tion, Laplace distribution. The beta distribution is often proposed in PV
Reinforcement learning RL: RL doesn’t operate over a fixed data set
irradiance and Wind speed calculation and forecasting by calculating its
and has gained considerable interest in the energy domain. In the smart
alpha and beta distribution [31]. Cauchy distribution is widely used in
energy domain, some studies have compared the RL approach with the
power system harmonic assessment, prediction-based models, wind
Model predictive control (MPC) method for uncertainty and forecasting;
power uncertainty calculation, along with the real-time dispatch of wind
however, increasing complexities in the power system due to integration
farms [32]. Weibull and Gamma distribution are also some of the methods
of renewable energies, has resulted in power system security along with
commonly used for wind speed calculation in distributed generations
bringing more uncertainties. RL did great in dealing with these smart
(DGs).
energies based problems due to its sequential decision-making power
Optimization Algorithms: Power system, in general, offers a variety of
taking account of risk measures and uncertainties [23].
optimization technique for heterogeneous problems such as non-linear,
Bayes Concept ML learning: Bayes theorem is a concept in the applied
large-scale, subject to uncertainties, discrete and continuous. There are
ML which provides the relationship between data and model. Bayesian
certain optimization methods used in the modern power system which
methods can quantify that uncertainty, and deep learning models follow
are addressed and mentioned by our study in Table 1. Those are; First-
the Bayesian paradigm. In terms of smart power systems, the Bayesian
order and Second-order methods, Stochastic gradient descent, Con­
methods could be used to predict net load forecasting. The aggregated
strained optimization, proximal Gradient method, Bound optimization,
net load forecasting in the smart energy power systems can be achieved
and Derivative and blackbox free optimization. The first-order optimi­
by combining Bayesian theory and deep (LSTM) long short-term mem­
zation method is often used to categorize numerical optimization
ory. LSTM is a recurrent neural network specially designed containing

5
 T. Ahmad et al.
Table 1
An overview and classes functions of probabilistic ML models in energy distribution.
Types of ML in energy distribution

Supervised learning Unsupervised learning Reinforcement learning

Classification based Regression-based Generalization and Clustering-based Self-supervised Evaluating unsupervised Value-based Policy-based Model-based
supervised learning supervised learning overfitting unsupervised learning learning reinforcement reinforcement reinforcement learning
learning learning learning
Probabilistic inference
Bayes’ rule Bayesian concept learning Bayesian ML
Inverse issues The Monty Hall issues Discrete concept learning Continuous concept Fully Bayesian Scalar input and binary Plug-in Binary input and Scaling up
learning approach output approximation scalar output
Probabilistic models
Binomial and Bernoulli distributions Multinomial and Categorical distributions Univariate Gaussian distribution
Sigmoid logistic Binary logistic Log-sum-exp trick Softmax function Multiclass logistic Cumulative distribution Regression Probability density Half-normal
function regression regression function function
Common univariate distributions Multivariate Gaussian distribution Mixture models
Student t distribution Beta distribution Cauchy distribution Gamma distribution Laplace distribution Mahalanobis distance Conditionals and Gaussian mixture Gaussian scale
marginals of an MVN models mixtures, and mixtures
of Bernoullis
Optimization algorithms
First-order methods Second-order methods
Descent direction Momentum methods Step size and learning Convergence rates Natural gradient Trust region methods Newton’s method Quasi-Newton and other
rate descent Broyden–Fletcher–Goldfarb–Shanno methods
Stochastic gradient descent Constrained optimization
SGD for fitting linear Iterate averaging Preconditioned SGD Lagrange multipliers Linear programming Mixed-integer linear The KKT conditions Quadratic –
regression programming programming
Proximal gradient method Bound optimization Derivative and blackbox free optimization
Projected gradient Proximal operator for Proximal operator for The EM algorithm EM for a GMM EM for an MVN with Simulated annealing Grid search and Model-based blackbox
descent L1-norm regularize quantization missing data optimization random search optimization
6

optimization
Information theory
Entropy Relative entropy
Entropy for discrete Joint entropy Cross entropy Conditional entropy Perplexity Differential entropy for Forward vs. reverse KL divergence KL divergence and
random parameters continuous random KL between two MLE, and Non-
parameters Gaussians negativity of KL
Bayesian statistics
Conjugate priors Non-informative priors
The beta-binomial The multivariate The Dirichlet- Beyond conjugate The Gaussian- Jeffreys priors Reference priors Invariant priors –
algorithm Gaussian-Gaussian multinomial algorithm priors Gaussian algorithm

Renewable and Sustainable Energy Reviews 160 (2022) 112128

algorithm
Hierarchical priors Empirical priors Bayesian model comparison
Hierarchical binomial Hierarchical Gaussian Hierarchical binomial Hierarchical Gaussian Bayesian model Occam’s razor, and Posterior predictive Bayes model averaging Bayesian hypothesis
algorithm algorithm algorithm algorithm selection Bayesian hypothesis checks testing
testing
Approximate inference techniques –
Laplace Markov Chain Monte Online inference using Grid approximation Variational
approximation Carlo approximation assumed density filtering approximation
Bayesian decision theory
Bayesian decision theory Bandit problems
Classification Precision-recall ROC curves Probabilistic Contextual bandits Exploration-exploitation Optimal solution, and Markov decision Upper confidence
problems curves and regression prediction problems tradeoff regret processes bounds, Thompson
problems sampling, and simple
heuristics
Linear discriminant analysis and logistic regression
Gaussian discriminant analysis Binary logistic regression
Quadratic decision Nearest centroid Logistic regression Fisher’s linear Nonlinear classifiers, Iteratively reweighted
boundaries classifier and discriminant analysis and MAP estimation least squares
(continued on next page)
 T. Ahmad et al.
Table 1 (continued )
Types of ML in energy distribution

discriminative vs. Linear decision Linear classifiers, and Perceptron algorithm,

generative classifiers boundaries, Naive maximum likelihood and stochastic
Bayes classifiers estimation gradient descent
Multinomial logistic regression Robust logistic Bayesian logistic regression
Nonlinear and linear Gradient-based Maximum entropy regression, and Bi- Laplace Approximating the Variational inference MCMC approximation Online inference using
classifier, and optimization, and classifiers tempered loss approximation posterior predictive assumed density
hierarchical bound optimization filtering
classification
Linear regression
Standard linear Ridge regression Robust regression Lasso regression Bayesian linear regression
regression Student distribution Laplace distribution Huber loss Estimating ρ(w ↑, σ2 ) Estimating ρ(w, σ2 ↑D) Sparsity-promoting
with Gaussian prior with Gaussian- priors, Hierarchical
Gamma prior priors, and empirical
bayes
Generalized linear models
The exponential family Generalized linear Probit regression
MLE for the Log partition function models Multinomial probit Bayesian inference Latent variable Ordinal probit Maximum likelihood –
exponential family, is cumulant models interpretation regression estimation
and exponential generating function
dispersion family
Deep neural networks
Multilayer Backpropagation Training neural networks Regularization Feedforward networks
perceptrons Forwards pass, and Backwards pass, and training the learning Exploding and Early stopping, and Bayesian neural Mixtures of experts Radial basis function
computation graphs automatic differentiation rate, variable banishing gradients, dropout networks, and Sparse networks
initialization and batch DNNs
normalization
Neural networks for electric images
7

Adversarial neural networks Solving different discriminative vision tasks with Neural network basics
CNNs
Whitebox gradient- Defenses based on Blackbox gradient-free Deep dream and Real-world Objective detection and Human pose Convolution in 1d and Boundary strides and
based attacks robust optimization attacks neural style transfer adversarial attacks image tagging estimation and image 2d, and convolution as conditions, pooling
segmentation matrix-vector layers, and
multiplication normalization layers
Neural networks for sequences
Recurrent neural networks 1d CNNs Transformers Efficient transformers
Vec2Seq sequence Seq2Vec sequence Backpropagation through 1d CNNs for sequence Self-attention, RNNs and CNNs, Low-rank and kernel Recurrence and Fixed non-learnable
generation classification time, gating, and long- classification and positional encoding, Comparing transformers methods, and memory methods localized attention

Renewable and Sustainable Energy Reviews 160 (2022) 112128

term memory sequence generation and multi-headed learnable sparse patterns
attention attention patterns
Nonparametric models
K nearest neighbor Learning distance metrics Kernel density estimation
classification Convex, and linear Deep metric learning, and Ranking losses, and Speeding up ranking Kernel regression Density kernels Parzen window KNN and KDE
methods classification losses different training loss optimization density estimator classification
tricks for DML
Kernel methods
Mercer kernels Gaussian processes Scaling GPs to large datasets Support vector machines
Mercer’s theorem, Some popular Mercer Noise-free and noisy GPs for classification Random feature Sparse variational SVMs for regression, Multi-class Kernel ridge
Hilbert space, kernels, Kernel ridge observations approximation, and inference Sparse vector classification with regression, and
representer exploiting machines SVMs, relevance Converting SVM
theorem parallelization vector machines outputs into
probabilities
Forests, trees, boosting, and bagging
Regression and Ensemble learning Bagging tree Random forests Boosting tree Interpreting tree
classification tree Stacking ensemble Bayes mode averaging LogitBoost, and gradient Forward stagewise ensembles with feature
boosting additive modeling
(continued on next page)
 T. Ahmad et al.
Table 1 (continued )
Types of ML in energy distribution

AdaBoost and importance and partial

exponential loss, and dependency
least squares boosting
Learning with fewer labeled examples
Transfer learning Model-agnostic meta- Semi-supervised learning Active learning
Fine-tuning Domain adaptation learning and few-shot Self-training and Label propagation and Deep generative models, Entropy minimization Decision-theoretic Information-theoretic
supervised pre- and unsupervised pre- learning with matching pseudo-labeling consistency and a combination of approach, and batch approach
tuning training networks regularization semi-supervised and self- active learning
supervised learning
Dimensionality reduction
Principal components Factor analysis Autoencoders Manifold learning
analysis EM model for PPCA Generative model, Nonlinear factor Bottleneck Denoising autoencoders, The manifold Isomap, Kernel PCA, Multi-dimensional
and FA, and Mixtures of factor analysis and factor autoencoders, sparse autoencoders, and hypothesis, and local and maximum scaling, Laplacian
exponential family analyzers analysis models for contractive Variational autoencoders linear embedding variance unfolding eigenmaps, and t-SNE
factor analysis paired data autoencoders
8

Clustering
Hierarchical K means clustering Clustering using mixture models Spectral clustering Biclustering
agglomerative The K-means++ Vector quantization Mixtures of Mixtures of Gaussians Laplacians encode the Normalized cuts Nested partition Basic biclustering
clustering algorithm, and the K- Bernoullis clustering for models
medoids algorithm eigenvectors
Recommender systems
Explicit feedback Implicit feedback Leveraging side Exploration- –
Collaborating Autoencoders Matrix factorization Bayesian Factorization Neural matrix information exploitation tradeoff
filtering personalized ranking machines factorization
Graph embeddings

Renewable and Sustainable Energy Reviews 160 (2022) 112128

Shallow graph embeddings Graph neural networks Deep graph
embeddings
Distance-based: Distance-based: non- Unsupervised Outer product-based: Spectral Graph Message passing GNNs Non-Euclidean graph Spatial graph Unsupervised and
Euclidean methods Euclidean methods, embeddings and outer matrix factorization convolutions convolutions convolutions semi-supervised
supervised product-based: Skip-gram methods embeddings
embeddings methods
T. Ahmad et al. Renewable and Sustainable Energy Reviews 160 (2022) 112128

Fig. 4. A new AI generation: a technology perspective for ML application to smart energy and control systems and the evolution of ML from 1957 to today [20].

methods using a first derivative algorithm. The second-order method [41], e) data updating, and so on. Many reasons can be attributed to the
also called the Newton method is an algorithm that uses the second growing interest in ML: 1) better computation power [42]; ii) hybrid
derivative in a scalar case. These optimizations play a vital role in the transactional/analytical processing systems and big data [43], and iii)
optimal power flow (OPF) problem of the energy power system. OPF is advanced and new ML techniques [15]. How can ML help the energy in­
an optimization tool used for power system operation, analysis, and dustry? Technology developments have enabled us to leverage the
energy management. Linear programming, Mixed-integer linear pro­ following approach in the energy industry: for example, i) ML and grid
gramming, KKT conditions, and Quadratic programming are all bringing management [44]; ii) solving demand response [45,46]; iii) predictive
the application to energy and smart grid system in a variety of ways that maintenance [47]; iv) energy source exploration [48]; v) save costs
include but are not limited to generation and expansion planning of when reducing electricity use [49]; and so on. Table 2 shows ML’s high
power systems, operation of different kinds of conventional and modern benefits and low complexity use case in core energy distribution func­
power system, optimal power flow modeling and heuristic techniques, tion and real-time examples.
demand response and risk, uncertainty measures [33,34]. Pervasive sensing empowers extensive data, making it easier to uti­
Bayes decision theory uses the Bayes theorem to find the conditional lize ML-related approaches [52]. In this scenario, statistical
probabilities. It is based on tradeoff quantification among various de­ techniques-based ML algorithms may rapidly identify failure sensors and
cisions based on Bayes theorem probability and the cost associated with extract their data from analytics. Fig. 5 (a) illustrates a primary grid
the probability. It deals with classification problems, ROC curves, connected with microgrids (also known as “dispatchable electricity re­
probability predictions, Markov decision process, Thompson samplings, sources”). Energy production plants (renewable, conventional, etc.),
etc. In modern energy systems, including various kinds of renewable power transmission lines, and power distribution networks are
energy resources, the degree of uncertainty emerges as a huge problem. constantly regulated by utilizing cutting-edge technology such as ML
Bayesian decision theories stand out to effectively tackle the degree of models. Using the base stations in urban areas, sensor data can be
uncertainties [35]. transmitted via cellular communication networks. In certain rural pla­
Deep Neural Networks: In the Smart grid and energy systems, the ces, satellites or LoRa could transmit sensor data.
energy load forecasting and uncertainty brought by distributed gener­ Further, the sensor data is sent to the supervisory control and data
ations remains a discussing topic for researchers. Deep neural network acquisition system so that it can remotely track, automate the whole
(DNN) consists of many ML-based methods to handle energy load and network, preserve network reliability, create alerts, and evaluate
uncertainty problems differently. Convolutional Neural Network (CNN), network topology. For billing purposes, the metering in residential,
Artificial Neural Network (ANN), and some other methods (refer to commercial, and manufacturing regions sends electricity usage and
Table 1) have done a great job in dealing with the mentioned problems production data to the utility [53]. Additional data is frequently
in modern energy systems [36]. required to control the highly intermittent output of renewable energy.
This information is forwarded to the demand response unit to better
1.5. How is ML changing energy distribution? manage demand and supply. The processing and storage of sensor data
are important for sensors to gather and report their findings to analytics
In reality, renewable energy utilities (e.g., solar, wind, nuclear, and tools. Fig. 5 (b) demonstrates the deployment of the sensors on distri­
hydro) have significantly gained from the recent increase in powerful AI, bution lines [54], transmission lines, transformers, underground cables
ML predictive models, and data science [37]. The predictive models [55], smart meters, circuit breakers, protective relays [56], etc., in a
have reduced their expenses, better forecast potential occurrences, and substation. In addition, a detailed analysis of pervasive sensing, ML in
maximized their portfolio’s return. ML models perform the different energy systems, and cyber-enabled grids is presented by Philip et al.
kinds of tasks in energy applications, for example, a) data storing [38], [52].
b) visualizing the data [39], c) data analysis [40], d) capturing data

9
T. Ahmad et al. Renewable and Sustainable Energy Reviews 160 (2022) 112128

Table 2 2. Machine learning in core energy technologies

ML high benefits and low complexity use case in core energy distribution
function: real-time examples. ML is creating new opportunities for cutting-edge research in energy
Sr. Area of implementation of Real-time applications at the glance distribution [57]. This section covers seven different core areas of en­
# ML algorithms ergy distribution, including i) advanced energy materials; ii) energy
1 Yield optimization GE renewable energy uses ML to develop on a systems and energy storage devices; iii) energy efficiency in the in­
cloud-based platform virtual wind farms that dustry; iv) energy demand response; v) strategic energy planning under
imitates an actual, physical design. This model uncertainty; vi) large-scale integration of renewable energy; and vii) big
works on a real wind trend and measures
data analytics in the smart grid environment. Table 3 explains the ML
energy to maximize output on a single turbine
scale or an individual level. It projects a 20% technologies in core functions and technical use-cases in energy distri­
growth in electricity efficiency to save $100 bution systems [51].
million over a 100 MW farm’s lifespan [50]. This visualizes four significant aspects: power transmission and dis­
2 Energy forecasting U⋅S.-based Vermont Electric Power Company tribution use cases in core areas, energy services and supply, generation,
is developing hyperlocal weather prediction
systems through integrated data science and
exploration, production core use cases, and energy trading use cases in
ML techniques. In both solar and wind farms different energy distribution systems. Detailed analysis of each section is
in Vermont, the application of the weather listed below:
model resulted in a decreased average solar
energy prediction error by 6% and wind errors
2.1. ML for advanced energy materials
by 9%. Mr. Kerrick Johnson, VELCO’s Vice-
President of Policy and Marketing, said, we
will save $1 million in fee payments per 1% Recent advances in ML have raised the likelihood that data-driven
reduction in load through improved resource materials research can revolutionize breakthroughs and offer new par­
orchestration. For example, as a transmission
adigms for producing energy materials. We can anticipate substantial
service in the short-term load prediction, we
also recognize operational advantages [51].
improvements in materials science [7]. Moreover, recent developments
3 Energy storage Greensmith Energy uses ML to control energy in data-driven materials engineering often propose that ML technologies
storage devices and wider energy can improve the design and production of advanced energy materials in
environments, a multinational energy storage addition to the discovery and implementation of such materials. For
firm. It supplies the Spanish island Graciosa,
illustrate, the National Institute of Standards and Technology (NIST)
for example, with its real-time applications.
With an incorporated 6 MW/3.2 MWh of maintains 65 separate databases with 67,500 different measurements.
energy store control facility, a modern Furthermore, during the last decade, more than 1.7 million research
‘Graciosa Hybrid Renewable Power Plant’ can studies have been published on both solar cells and batteries [58].
deliver 1 MW of solar and 4.5 MW of wind Computational techniques are being established to help create
power to the local power grid, minimize the
island’s fossil fuel demand and reduce
structures and measure different electronic properties and other features
emissions of greenhouse gases substantially [69]. Another noteworthy example is the materials initiative, which
[51]. utilizes supercomputers to simulate all characteristics of the materials.
4 Grid behavior interface The “Gridsense” technology is a Swiss leading Currently, there are more than 700,000 materials with forecast prop­
energy corporation that uses ML to evaluate
erties identified [60]. In 2016, Organization for Economic Co-operation
criteria such as grid requirements, demand,
and generation of electricity, environment and Development member countries spent over $16.6 billion on energy
predictions, and pricing for electricity. This material research and development activities, while in 2000 was less
makes it possible for consumers to learn than $10 billion [60]. In October, the United Kingdom unveiled its
better. Using this information, GridSense renewable development plan, which would spend more than $3.3 billion
optimizes customer and generator power use
and decreases power grid peak loads, balance
on low-carbon material technologies between 2015 and 2021 [58].
loads, and stabilizes power delivery [51]. However, the enormous ability to transform this form of data into in­
5 Energy trading In energy trading, British Petroleum (BP) uses dustrial and business applications is yet to be recognized [60]. ML al­
ML automation. ML automation consolidates gorithms designed to discover trends in large data sets — can
trading floor data utilizing automation
significantly accelerate energy-materials discoveries. However, there
robotics to imitate repeated procedures, CIO
for enforcement, legal, risk, and finance at are always challenges. Materials cannot be described universally. Spe­
British Petroleum energy supply & trading. It cific physical properties, such as the elements’ composition, shape, and
changes the position of analysts and liberates conductivity, are essential for numerous applications. An abundance of
them up more freely to concentrate on more scientific evidence on materials is hard to come by, and the feasibility of
important tasks. So, they can spend time
analyzing and asking about the context of
computational testing of theories relies on predictions and simulations
these data instead of collecting data sets [51]. that could be unreliable under the actual situation.
6 Complaints management Exelon created a channel-agnostic chatbot Different ML models for energy materials modeling are visualized in
operated by ML, a US-based power and gas Fig. 6 [61]. The first key step in every ML research is to assess the ob­
utility to address consumer problems on
jectives and forecast goals of the ML algorithms, which is normally
challenges like billing and outages. Exelon
reduced the number of consumers and gained achieved with the assistance of domain experts. This is arguably the
more insight into their consumption needs most critical step since the goal must be unambiguously specified and
[51]. theoretically learnable from available knowledge, such as composition
The implementation of ML algorithms includes yield optimization, energy and crystal/molecule structure, material information, determining
forecasting, energy storage, energy trading, grid behavior interface, and quantities or objects experimentally, etc. In materials engineering, su­
customer complaints management. Detailed analysis of each approach is given pervised learning seems to be the most general method of ML, in which a
in Section 2 and Section 3. model has trained the mapping function between input feature variables
(i.e., composition and molecular/crystal) and output labels or values (i.
e., characteristics like bandgaps, energies, and so on). The parame­
ters/weights of the algorithm are iteratively modified during training to
mitigate model losses in response to training results. Both ML models
aim to make low-cost, relatively reliable predictions that can be used

10
T. Ahmad et al. Renewable and Sustainable Energy Reviews 160 (2022) 112128

Fig. 5. Pervasive sensing (a) in the power grid and (b) in the substation side [52].

Table 3
ML technologies in core functions and technical use-cases in energy distribution.
Sr. Description Power transmission and Energy services and supply Generation, exploration, and Energy trading core use
# distribution use cases in core production core use cases cases
areas

1 Use of AI in natural 30% 34% 31% 27%

language
2 Use of AI in machine 14% 14% 16% 6%
learning
3 Use of AI in computer 33% 24% 30% 36%
vision
Intelligent automation advantages
Sr. Key areas of smart automation
# Description Smart automation benefits as Overestimated intelligent automation Underestimated intelligent automation –
predicted as well as expected advantages in the energy sector advantages in the energy sector
1 Advantages of energy 14% 39% 47% –
cost savings
2 Consumer satisfaction 16% 36% 48% –
advantages
3 Incremental and new 17% 37% 45% –
revenue advantages
The energy and utility sector are driving significantly with the use of intelligent automation
Sr. Sector by category Improved consumer Decrease the different number of steps Increase the total number of consumer Personalized products/
# experience through faster and processes for purchase and queries availability by being open longer hours services for consumers
response
1 Energy and utilities 81% 78% 74% 67%
2 All sectors 60% 61% 30% 48%
Compared to other sectors, the energy sector brings huge benefits from smart automation
Description Increase in energy operations Data accuracy improvements Improvement in workforce agility Increase productivity in
quality staff members
1 Energy and utilities 40% 37% 33% 32%
2 All sectors 30% 30% 20% 26%
Description Turnaround faster time for Consistency in data improvements Less requirement of resources to Accurate and better
service requests accomplish the designed task compliance with
regulatory
3 Energy and utilities 32% 32% 28% 27%
4 All sectors 21% 20% 18% 21%

instead of more costly computational, experimental, or human-driven must include adequate funding to meet their commitments. In summary,
approaches [62,63]. As a consequence, ML models make it possible to: a substantial increase in investment is expected in ML-enabled materials
science worldwide. Experimentalists, ML researchers, and algorithm
1) Providing fast forecasts of characteristics and novel materials to programmers can work together to help with troubleshooting more
speed up the discovery of new materials. quickly [64]. The training samples or, more generally, computed—are
2) At greater time/length scales, precise simulations of complex mate­ obtained in the second step. Featurization is converting experimental
rials are feasible. data into numerical properties (usually in the form of tensors or vectors)
3) Further, it enhances the interpretation and characterization. that can be used to differentiate between various materials [7].

Collaboration between energy sciences and ML communities is

required. One of the keys aims of the energy and renewable energy 2.2. ML toward advanced energy systems and storage devices
materials research task, administered by the mission innovation global
partnership, is to help for implementing the ML approaches. Mutual A more comprehensive range of storage technologies has been built
government commitments support this partnership, and governments to address the daily energy needs of the utility grid. It includes a battery
(e.g., flow batteries, capacitors, advanced chemistry batteries, and a

11
T. Ahmad et al. Renewable and Sustainable Energy Reviews 160 (2022) 112128

Fig. 6. A framework for developing ML models for energy materials. This workflow consists of five major steps: defining an objective, data processing, data quality
improvement, model creation, and implementation [61].

large number of electrochemical storage solutions) [65], hydrogen (e.g., flexible storage availability, resulting in improved demand for both the
excess energy production is converted into hydrogen via electrolysis and consumer and the grid. Since renewable energy sources are intermittent,
stored) [66], thermal (e.g., captured cold and heat to create on energy ML may help address the problem of catching the intermittent variations
offset or demand needs) [67], mechanical storage (e.g., gravitational of generation phases.
energy, harness kinetic energy or other innovative technologies) [68], To conclude, the demand for ML in energy storage is now greater
and pumped hydropower (e.g., using water to build large-scale energy than ever [76]. Incorporating it with energy technologies would affect
reservoirs) [70]. the planet’s future. The convergence of ML and energy storage is a
As technology advances, energy storage systems and devices with logical as well as a practical starting point. However, the system is only
increased capacity, better efficiency, longer life, and a more intelligent in its development stage in terms of the possible effect on the energy
management approach are anticipated [70]. Advanced control methods storage system. This synergy has the power to transform the environ­
depend on certain indicators’ instantaneous status, so developing those ment and open up many possibilities while simultaneously enhancing
structures needs a trade-off among many parameters. ML will signifi­ sustainability.
cantly speed up simulations, catch dynamic structures to boost predic­
tive precision and make smarter decisions based on specific data
2.3. ML tools for energy efficiency
information. Fig. 7(a) demonstrates the use of ML in different energy
storage devices and systems. Primus Power,3 a pioneer in low-cost,
ML’s energy effect will rapidly become unsustainable unless new
grid-scale electrical energy storage systems, has developed a flexible,
methods are created. But how can we allow the responsible use of it?
stable distributed flow battery device that can be used for different
Manufacturers rapidly utilize ML to increase throughput and reduce
storage applications (see Fig. 7 (b)). Fig. 7 (c) represents the overall
energy usage (see Table 4). The EU-funded FUDIPO initiative is making
distributed concepts of ML in advanced energy systems. It is divided into
substantial progress in incorporating ML into various critical process
three main classes including: 1) energy domain [71]; 2) energy com­
sectors on a broad scale to gain significant energy and resource savings
ponents [72]; and 3) energy applications [73,74]. Saqib et al. [75] have
[77]. An example of the energy performance of active buildings con­
presented a detailed analysis of these classes. Continuing to grow the
tracting techniques toward resilience and smart buildings is visualized
percentage of renewable energy providing electricity to the grid as part
in Fig. 8. 3As an electrification-focused on basic demand response to
of the energy transition phase ensures that the amount of produced
differential pricing is often the best approach to a thriving business
power can be impacted by conditions, making it harder to monitor. A
scenario, ML can play a significant role as aggregator sites of implicit
smart grid must be built and combined with energy storage technologies
demand response in buildings. There is a lot of ML research that aims to
to effectively control and dispatch electricity to enhance grid reliability
improve predictive algorithm accuracy, but scientists and engineers are
and renewable energy usage. The bulk of this can be done by storing
recently getting more active in improving energy efficiency.
excess wind and solar resources in an energy storage facility. Such de­
Energy4 use impacts millions of schools, households, hospitals, and
vices serve various industrial and corporate purposes, including peak
industrial facilities, each imposing its energy conversion strategies [78].
shaving, contract capacity optimization, and backup control. This form
Consumers in the US spend billions of dollars/years on energy conser­
of energy conservation is commonly recognized as the most adaptable
vation hoping that these investments would compensate for themselves
green energy approach possible (see Fig. 7 (d)).
by lower electricity prices in the future [80]. Many studies have been
ML algorithms will benefit the energy storage industry greatly. ML-
conducted on energy efficiency optimization algorithms so far. The most
enabled energy storage can assist in the processing and reviewing data
prominent categories include: i) simulation-based optimization methods
and provide insights into maximizing power consumption and fore­
[81], and ii) surrogate-based optimization methods [82]. In the
casting future faults by simulations. Incorporating battery-based intel­
simulation-based optimization method, the objective function (e.g.,
ligent storage into a renewable energy system almost inevitably
energy usage as well as thermal comfort) is controlled by an optimiza­
improves the economic benefit. ML-enabled storage can allow real-time
tion method, and the decision parameters are controlled by an optimi­
zation tool to develop the objective function iteratively. There are

3
Roberta D. Angiolella, Active building Energy Performance Contracting
4
(AEPC) models towards smart flexible buildings, https://www.construction21. Jim Kring, Grid-Scale Energy Storage Powered by LabVIEW, https://blog.
org/articles/h/active-building-energy-performance-contracting-aepc-models- jki.net/news/grid-scale-energy-storage-powered-by-labview, Accessed: 03/12/
towards-smart-flexible-buildings.html. (Accessed: 03/12/2021). 2021.

12
T. Ahmad et al. Renewable and Sustainable Energy Reviews 160 (2022) 112128

Fig. 7. (a) ML in energy storage devices [68]; (b) grid-scale energy storage solutions [75]; (c) use of ML in advance energy systems; (d) real-time energy storage
applications.

Table 4
ML designed to help record huge improvements across the value chain in the energy industry [187].
Sr. Description Applicable Produce Project Provide Promote
# technologies

ML can produce values across the value chain

1 Retail Automate the store operations and Anticipate energy demand Tailored, enriched, and Services, and products
optimize the merchandising energy trends, however, automating convenient customer at low prices, with the
devices and products at micro space and optimizing supplier experience right target and the
contracting and negotiation right message
2 Power utilities Improve electricity generation yield, Enhance energy supply and Provide accurate energy Day-of-day dynamic
optimize preventive maintenance, demand, measure the reliability consumption insights, tariffing as well as
prevent electricity theft, and of integrated energy production automate the consumer match producers of
decrease energy waste assets, and automate the control services with the use of virtual different consumers in
of energy demand-side response agents real-time
3 Manufacturing Reduce the error at different states, Enhance the device efficiency Enhance the skills for pilot Optimize the energy
devices improve the process tasks, reduce and yield, and automate training and maintenance pricing, forecast of
the energy material delivery time, supplier requirements and engineer maintenance services,
and limit the different product anticipate different parts of and refine sales-leads
rework requirements prioritization
The ML techniques can help capture a significant number of chains across the value chain
1 Power utilities – 20% energy production increased The key aim of reducing the There is a huge saving in –
with the use of ML models and smart 10% in-country energy usage is energy bills, for illustrating
sensors to optimize the energy by applying the deep learning $10-$130 by applying ML
yields, improvement in EBIT models to forecast the energy models to automatically
10–20% by applying ML to increase supply and demand [79]. switch the energy supply deals
the predictive maintenance, [79].
automate the line fault forecasting
and enhance the capital productivity
[79].

13
T. Ahmad et al. Renewable and Sustainable Energy Reviews 160 (2022) 112128

Fig. 8. Techniques for active building energy performance contracting in the direction of smart, flexible buildings.

different types of simulation-based optimization methods including: i) Fig. 9). It is an improvement over the 2017 report, which reported that
optimization methods [83]; ii) derivative-free methods; iii) annual carbon dioxide emissions would be higher in 2050, showing that
gradient-based methods; iv) global search (meta-heuristic) methods; v) nationally determined contributions and the steadily improving cost and
local search methods; vi) population-based methods; and vii) single efficiency of renewable energies affect long-term energy planning and
solution-based methods. Surrogate-based optimization has a lot of po­ scenarios. Severe additional cuts, furthermore, are required. Yearly
tentials, particularly for high computational systems. However, the few energy-related carbon dioxide emissions would also decrease by 2050
experiments that have been performed so far have not looked at how to from 35 gigatons (in the comparison case) to 9.7 gigatons, reducing even
build surrogate models effectively, have neither they thoroughly more than 70%, to reach a climate target of restricting global warming
exploited their benefits in terms of facilitating optimizing improve­ to 2 ◦ C [17].
ments. The usage of both surrogate-based and simulation-based opti­ According to international renewable energy agencies’ study, energy
mization approaches in buildings is still a hot topic of research. Both conservation and green energies, along with deep electrification of end-
approaches, though, have a high computational cost when searching for uses, will provide over 90% of the required reduction relating to energy
near-optimal solutions.5 carbon dioxide emissions. For particular production processes, pollu­
According to the reference case (which considers existing and ex­ tion, fossil fuel switching (to natural gas) and carbon capture, and
pected policies, like nationally determined contributions), energy- sequestration in manufacturing can be used to make up the difference.
related carbon dioxide emissions will rise slightly per year until 2040, Nuclear power generation will remain unchanged from 2016. Simulta­
then drop slightly by 2050 to approximately match current levels (see neously, by 2050, substantial efforts must be made to cut carbon emis­
sions from industrial processes and land use to less than zero. Without
success in those areas, the temperature target will not be met. ML can
help in this regard. ML improve the efficiency of power system infra­
structure, helping to integrate the renewable energy with the smart grid,
improving the efficiency of conventional power plants through auto­
mation of power grid infrastructure, and enhancing the performance of
buildings, transport, district heat, and industrial sector, which are the
key part of energy distribution.

2.4. ML techniques for energy demand-side response

ML is one of the most important parameters for demand-side man­

agement and market response since it allows consumers to adjust the
market mechanisms in actual environments, thanks to ML enabling
Fig. 9. Energy efficiency and renewable energy can render over 90% of the technologies [84,85]. Smart homes are using digital technology to help
decrease related to energy carbon dioxide emissions [17]. with demand management. The ML techniques link devices, including
rooftop solar PV, local battery storage, smart meters, and home appli­
ances to the internet, allowing for data collection and sharing [86].
Fig. 10(a) demonstrates the synergies between developments for
demand-side management integration [87]. It includes four essential
5
Delta Energy Storage System: A Versatile Power Modulation Strategy, https: parts, including i) enabling technologies (e.g., renewables power to heat
//blog.deltaww.com/en/energyinfrastructuresolutions-en/products/delta-ene of residential sector, batteries behind the meter, ML/AI, and big data
rgy-storage-system-a-versatile-power-modulation-strategy, Accessed: 03/12/ applications; and electric vehicle for smart charging); ii) market design
2021. (e.g., net billing schemes and time-of-use tariffs); iii) business models (e.

14
T. Ahmad et al. Renewable and Sustainable Energy Reviews 160 (2022) 112128

Fig. 10. ML for (a) synergies between developments for demand-side management integration [87]; (b) 40% energy demand reduction using AI [87]; (c) multi-agent
based demand-response [88]; (d) in the sustainable development scenario, from 2018 to 2040, there is a huge potential for demand response.7

g., energy-as-a-service); and vi) system operations (e.g., renewable includes real-time energy pricing, rates based on time of use, and
power production and advanced prediction of parameters). critical-peak pricing. The incentive-based energy demand response in­
According to a new study, the numbers of “connected homes” or cludes emergency energy demand response programs, control of the
“smart homes” increased by 31% from 17 million to 29 million in the direct load, ancillary market services programs, curtailable/interrup­
three years between 2015 and 2017, reflecting a compound annual tible rates, capacity market arrangements and programs, and buyback
growth rate of 31% [87]. In a pilot project in Sweden that used market programs/demand bidding [91].
mechanisms and ML models for demand response, peak energy usage There are different kinds of demand response challenges including: i)
fell from 23% to 19% of overall electricity demand; 17% of peak demand advanced metering investment (e.g., meters that record consumer en­
was moved to non-peak hours.6 A 1% decrease in power prices for a ergy usage based on time-of-day or more regular billing blocks are
utility result in a 0.66% decrease in peak demand for that system on required for all time-based rates); ii) customer desire/inertia for
average.6 According to the American Council for an Energy-Efficient simplicity (e.g., most consumers (mainly residential consumers) would
Economic, demand-response programs are being used to reduce power be suspicious of programs that involve initiative, such as when the
demands by 10% or more. Fig. 10(b) explains the energy reduction for program’s proposed framework is not transparent); iii) reduction in
cooling using AI-enabled ML models. Con Edison’s Community Power demand (e.g., the ability to sell gaming participants can provide or state
Pilot Project provides solar electricity at a reduced cost to 350 New York that they would reduce as they have already shut down energy re­
City Housing Authority households. Customers could save 15% on their quirements for the day is one with the main problems of most assessment
utility bills, or about USD 80. A detailed analysis of these facts and approaches); iv) voluntary and mandatory participation price-based
figures using AI-enabled ML in demand is given in Ref. [87]. programs (e.g., experience shows that the degree of consumer engage­
Fig. 10(c) demonstrates the demand response structure for multi- ment and overall load declines is moderate where the participation in
agent coordination [88]. Buildings can be seen exchanging informa­ price-based services were voluntary); v) fair time-based and simple
tion with the electricity grid, as their price and demand of electricity, pricing (e.g., customers informed of everyday pricing and market in­
while providing some kind (cooperative or competitive) dynamical creases by different methods achieve better results and are comfortable
communication and/or sharing of information between them. Fig. 10(d) with the programs); vi) peak load demand during off-peak hours (e.g., in
visualizes the sustainable development scenario; from 2018 to 2040, the demand response), the key strategy is to change demand throughout
there is a potential for demand response. Despite its tremendous po­ peak load hours or high energy rates. There could be a scenario when a
tential, demand-side resilience success is erratic across the world. large number of consumers are likely to change their energy demand in
Further, there are two main demand-response groups, including i) response to the energy price, contributing the grid to reliability issues);
price-based energy demand response [89]; and ii) incentive-based en­ and vii) the intermittent nature of renewable energy sources (e.g., the
ergy demand response [90]. Price-based energy demand response renewable energy sources are highly unpredictable and intermittent.
Renewable energy sources (e.g., solar and wind) will improve their
participation in profitable electricity markets by joining demand
6
IEA, Demand Response, https://www.iea.org/reports/demand-response, response by a virtual power plant) [91].
Accessed: 03/12/2021.
7
IEA, Demand Response, https://www.iea.org/reports/demand-response,
Accessed: 03/12/2021.

15
T. Ahmad et al. Renewable and Sustainable Energy Reviews 160 (2022) 112128

2.5. ML in smart manufacturing in the smart grid paradigm reliability and performance, and boost overall productivity [74]. Fig. 11
(a) depicts the development of smart manufacturing. ML is used to
The modern technological revolution is known as Industry 4.0 [92]. stimulate autonomous smart manufacturing schemes, in which key
A large amount of data is produced by connecting any machine and performance indicators and adaptive specifications are smart
operation to the internet using network sensors. ML is used to analyze manufacturing, and input data framework is the key objective. The au­
produced valuable information and data about smart manufacturing. thors argue that conventional manufacturing processes in transition can
Different kinds of ML models are used in smart manufacturing in the understand specific criteria from four different viewpoints: objective,
smart grid paradigm. The most prominent models include: 1) Bayesian practical requirements, technical requirements, standards, and market
networks [93]; 2) k-nearest neighbors [94]; 3) SVM [95,96]; 4) requirements. Fig. 11(b) depicts the goals and conditions of smart
instance-based learning [97]; 5) ensemble methods [98]; 6) artificial manufacturing using ML [107]. The key objective of smart
neural networks [99]; 7) multiple logistic regression [100]; 8) decision manufacturing in the energy sector includes i) autonomous lean oper­
tree [101]; 9) gradient boosted [102]; 10) self-organizing map [103]; ation (e.g., the primary goal of creating autonomous lean smart
11) k-means [104]; 12) bag of words; 13) random forest [105]; 14) manufacturing is to improve the production system’s performance and
locally weighted learning; and 15) additive models [106]. autonomy); ii) sustainable value added (e.g., the goal is primarily con­
The development of smart manufacturing is linked to technical ad­ cerned with smart manufacturing’s long-term value and viability of
vancements and the changing needs of business models and stake­ smart devices); and iii) win-win partnership (e.g., the ultimate goal of
holders. Smart energy manufacturing systems collect real-time energy smart manufacturing in autonomous service environments is to produce
consumption data, improve decision-making accuracy, increase plant continuous and better communication in which the whole smart

Fig. 11. ML in smart manufacturing (a) the advancement of intelligent manufacturing systems; (b) requirements and objectives of smart manufacturing [107]; (c)
industrial AI architecture [106]; (d) industrial ML structure [106].

16
T. Ahmad et al. Renewable and Sustainable Energy Reviews 160 (2022) 112128

manufacturing lifecycle is more successful than the number of its indi­ grid service challenging. They become energy-efficient and reliable as
vidual elements). Smart manufacturing’s ultimate goal in autonomous ML analyzes the large amounts of data obtained from their processes
service environments are to produce continuous and better communi­ daily, elevating the ML in energy design management and market. For
cation in which the whole smart manufacturing life cycle is more suc­ example, Table 6 demonstrates that ML transforms the customer and
cessful than the number of its different pieces or elements of different operations services using ML models. It is further shown that the new
smart energy devices. No wonder integrating and bridging a variety of investment in distributed energy generation by 2040 would be $10.4
emerging technology helps smart manufacturing accomplish autono­ Trillion [113–115]. The total investment in consumer empowerment,
mous lean operations and long-term value-added. Industry 4.0 has three grid modernization, and industrial size, and energy internet is noted 700
key dimensions: it promotes the computerization or digitization of Million, $400 Billion, and $14 Trillion, respectively [113–115]. The ML
manufacturing industries, reduces delivery times, provides greater in the energy management industry is expected to rise at a 19.8%
reliability, adapts to consumer needs with limited batch sizes, and ex­ compound annual growth rate from $4439.1 million in 2018 to $12,
pands downstream service offerings. 200.9 million in 2024, up from $4439.1 million in2018.9 In 2018, the
Fig. 11(c) demonstrates the ML structure in the industrial sector. The utility group led the industry by the end customer, as ML technologies
first layer, which starts at the bottom and works its way up, involves were introduced by several organizations, including Duke Energy Cor­
gathering information from the different energy systems devices. Data poration and Dominion Energy Inc., to fill the difference between power
from manufacturing equipment, data about the duration/summation/ demand and supply.
faults of output, and data knowledge about the logistics procedure are Incorporating the internet of things (IoT) in energy management is
essential to building a well-performing and robust industrial ML the most important development in the ML in the energy industry. IoT
framework. The different hardware components used to acquire the enables remote access, program automation, data analytics, and intel­
data, such as actuators, sensors, and embedded ML units, cannot be ligent tracking [116]. It employs smart meters and sensors mounted in
ignored. The next layer creates coordination links between industry both assembly lines and equipment to notify consumers of the energy
sectors and ensures effective data acquisition. It is also in charge of used volume. Most utilities believe ML would have a significant effect on
storing the data gathered by the first tier, securing it with the right their businesses, but they have yet to incorporate it into their key stra­
software and procedures, and extracting useful functionality from a pre- tegies. The smart grid may be improved in terms of control and per­
processing method if necessary. The industrial ML architecture in formance for energy performance strategy. ML forecast load demand
Ref. [108] as six levels of the pyramid shown in Fig. 11(d). The industrial and energy supply, enabling fully automated the world’s largest thermal
sector is included in the first layer. The ML models, health, and human solar farm in California, for example, to avoid millions of dollars in
performance, the resilient factory, productive energy systems, industrial damages due to mismatches between forecasted demand and real gen­
ML-based energy systems, and worry-free transportation are embedded eration [117]. Furthermore, linking small, localized networks such as
ML devices. The second layer is designed to meet the demands of In­ micro-grids to the cloud allows them to exchange and create data and
dustry 4.0. The third layer emphasizes the complications/challenges usage, allowing them to forecast load better. In reducing energy man­
that an ML system could face. The fourth layer demonstrates how to agement expenses, trading operations can be done automatically using
construct industrial ML systems using enabling technologies. The pro­ ML-based algorithms [117].
gramming methods, deep learning or ML methodology, and platforms The uncertainty of the modern energy landscape poses serious
used to build the framework comprise the fifth layer. Finally, the sixth challenges for utilities and independent power producers. Traditional
layer refers to the potential effects of an AI framework. Thanasis et al. market models, which are marked by consistent supervision, vertical
[106] have been conducted a detailed study on ML in innovative integration, guaranteed long-term return on investment, and non-
manufacturing industries. intermittent generating developments, are being overwhelmed and
outpaced. New innovations in ML that are affordable, autonomous, and
2.6. ML for strategic energy planning under uncertainty digitized would take their position, requiring new business opportu­
nities in the process.
Decision-makers have long been interested in energy forecasting,
management, and operation, especially regarding energy demand and 2.7. ML support for large integration of large-scale renewable energy
resource allocation [109,110]. Researchers have adopted various
decision-making techniques—and eventually implemented by manag­ According to a survey, renewable energy production must rise eight
ers—to assist decision-makers in selecting optimal choices, designing times larger than the fastest pace to reduce global warming or heating
energy strategies, and supporting competitive alternatives to energy in (see Table 7). The IREA claims $131tn will be required in investment in
uncertain settings [111]. Strategic energy planning aims to use the ML clean energy over three decades (see Table 8). The IREA said that im­
models to increase awareness and identify the overall market needs for mediate steps are needed to respond to increasing energy demand,
incorporation into the basic and advanced energy infrastructure, as well which would entail the total expenditure of $131tn in renewables
as to reach an agreement on the best integration strategies and industry by205010. The agency IREA anticipates that the global energy con­
practices focused on these strategies (see Fig. 12)8 [112]. Moreover, sumption would decrease to 4% by 2050, although the gas peak would
Table 5 shows the real-time implementation of ML models in different fall to 6% in 2025 and coal would fall to 2% by the mid-century.9 The
energy systems. increase in power consumption indicates that just over half of all energy
One of the key ML in the energy management industry drivers is AI’s used by 2050 will be renewable electricity, contrasted with 21% by
usage to improve the smart grid performance [106]. A grid is a network 2018.9 Approximately two-thirds of the energy consumed by fossil fuels
of cables, transformers, and different parts of infrastructure components have been consumed in recent decades but will be limited to 10% by
that carry energy over long distances. It may be for a whole country or a 2050.9
transcontinental grid, which distributes electricity throughout conti­
nents and countries. The grid’s power comes from several resources,
including solar panel systems, fossil-fuel-based power plants, hydro­ 9
AI in Energy Management Market, https://www.reportlinker.com/p05
electricity plants, wind power plants, and nuclear power plants, making 842955/AI-in-Energy-Management-Market.html?utm_source=GNW, Accessed:
03/12/2021.
10
The Guardian, https://www.theguardian.com/environment/2021/mar/15/
8
EPRI Smart Grid Demonstration Initiative, https://smartgrid.epri.com/doc/ renewable-energy-growth-must-speed-up-to-meet-paris-goals-agency-says:
EPRI%20Smart%20Grid%20Demonstration%202-Year%20Update_final.pdf. Accessed: 03/12/2021.

17
T. Ahmad et al. Renewable and Sustainable Energy Reviews 160 (2022) 112128

Fig. 12. Smart grid energy resources and energy management.

ML can revolutionize renewable energy: To develop a significant rise in production data, enabling ML to forecast different capacity levels,
renewable energy’s intermittent nature while maintaining the grid sta­ thanks to the implementation of sensor technologies [120]. More reli­
bility, ML would be needed in the low-carbon transition (see Fig. 13)11 able forecasts of intermittent renewable energy over shorter timescales
[37]. enable utilities and energy investors to predict their performance better
and compete in the balancing and wholesale markets–all while avoiding
1. ML has the power to unlock the enormous green energy potential. If penalties. In the meantime, ML algorithms based on massive volumes of
organizations do not accept it, it may risk falling behind the curve. weather data will help grid operators make the most use of their power
2. ML’s strong analysis capability can boost market forecasting and system networks and utility power grids by adjusting swift power system
resources management. operations to change the climate circumstances/conditions at any time
3. ML’s automation capabilities will help organizations achieve oper­ [76]. Improved short-term renewable energy prediction will lead to
ational success in various ways. greater unit engagement and dispatch.
performance, improving flexibility and reducing the need for oper­
The energy market is confronted with major global challenges that ational reserves. According to Brian-Case, a Chief Digital Officer at
need immediate action. Policy initiatives to a net-zero energy future, General Electric Renewable Energy, unexpected disturbances in the
like the endorsement and implementation of the Paris Agreement, energy industry will cost 3.01–8.01% of capability and USD 10 billion in
necessitate a rapid transition to a low-carbon economy [118]. As poli­ the annual lost-generation cost.12 ML-based algorithms analyze indus­
cymakers scale up renewables and move dependence on fossil fuels, trial data to render machine health forecasts and suggest measures to
major disruptions to the energy market are anticipated [119]. Grid op­ increase asset quality, such as wind and solar farms.
erators, engineers, and customers are trying to harness ML as we step The challenges of implementing ML through the industry: The poor na­
into the 4th Industrial Revolution, opening the way for a seamless shift ture of energy consumption data, customer regulatory, and mistrust
to more green energy [92]. The capacity of ML to have better prediction challenges could all be obstacles for the new energy technologies. There
tools enables enhanced asset management and market forecasting, is concern that is depending too heavily on ML might expose energy
although the automation technology drives operational excellence, networks to cyber-attacks in today’s information age. The hacking of 30
resulting in a strategic edge and increased efficiency for stakeholders. substations in Ukraine in 2015 served as a wake-up call, leaving 230,000
How do ML techniques transform renewable energy? ML could improve residents in the darkness for 6 h.9 One year later of this incident, a quiet
several aspects of business, from market forecasting to asset mainte­ more minor attack on a transmitting station in Kiev happened. Audit,
nance. Forecasting power levels has become critical to ensuring a reli­ data bias, and continuing algorithm testing are problems ML frame­
able and productive system when growing megawatts are fed into the works must consider when planning implementations from a perfor­
utility grid from intermittent variable renewable energy resources. This mance perspective [63]. Data must be extracted and rendered
is attributed to the lack of baseload production from different sources computer-readable, ensuring that accuracy is maintained from begin­
like coal, which renders power grid inertia by high concentration spin­ ning to end. Frequent data checking is needed for trustworthy ML to
ning machineries like gas and steam turbines, while renewables take up ensure that techniques stay in effect over time and that devices do not
a larger share of the grid. Power grids would be unreliable and vulner­ deviate significantly from the original techniques as they learn. Skep­
able to blackouts if system inertia is not present. Wind and solar gen­ ticism has been raised that depending too enormously on ML might
eration will now have a massive amount of wind and solar power expose energy networks to cyber-attacks. Advanced ML automation

11 12
Energy systems integration, University of Twente, https://www.utwente. Brian Case, Chief Digital Officer at GE Renewable Energy, https://www.ey.
nl/en/et/tfe/research-groups/TE/research/research/Energy_Systems_Integrat com/en_dk/power-utilities/why-artificial-intelligence-is-a-game-changer-f
ion/, Accessed: 03/12/2021. or-renewable-energy, Accessed:03/12/2021.

18
T. Ahmad et al. Renewable and Sustainable Energy Reviews 160 (2022) 112128

Table 5 Table 6
ML for smart grid paradigm in energy distribution. ML transforming the customer and operations services.
Sr. Description Main areas ML play an integral role Category %
# Improvements

1 Modeling and optimization 1) Google’s Deepmind is the leading in Increased the operational efficiency 78%
the area of modeling and optimization Reduced false positives rate 68%
in energy distribution, ii) Reduced operational cost because of process improvement 75%
reinforcement-learning algorithms are Greater regulatory/legal compliance at lower cost 70%
used for energy modeling, iii) Reduced churn 66%
probabilistic models used for solar Enhanced customer satisfaction 73%
irradiance and wind power Reduced the customer complaints 72%
forecasting, iv) IBM proposed Japan got success enhancing the turbine efficiency and reducing 20%
sophisticated energy generation and the operating and maintenance cost
weather forecasting in the United McKinsey’s UtilityX achieved replacement and maintenance cost 10–25%
States from 15 min up to 30 days in savings
advance. Billing, metering, & security 9%
2 Consumer-facing services 1) Integration of storage and Decentralized energy trading 33%
microgeneration with virtual power Carbon trading and green certificate 7%
plants, ii) smart home energy Energy gird management 8%
management, iii) ML used to identify ML opportunity and revolution
customers’ lifestyles, iv) social media New investment on distributed energy generation by 2040 $10.4 Trillion
and fintech, and v) storage or demand Consumer empowerment 700 Million
response, vi) virtual power plants (e. Grid modernization $400 Billion
g., artificial neural networks to Industrial size and energy internet $14 Trillion
decrease supply forecasting errors).
3 Investment and markets ML agents are largely used for
investment models and market industry concerning time [121]. ML’s capacity to incorporate a massive
interactions. rise in variable renewable energy into a predictable and secure grid will
4 Security and maintenance i) ML used to forecast and optimize the
maintenance efforts and schedules, ii)
be needed if the low-carbon transition accelerates.
physical and cyber security measures
to calculate the energy assets, iii) 2.8. ML for big data analytics in smart grid
previous maintenance cycles, iv)
failure modes, v) component lifetimes,
vi) outages cost to optimize
ML helps power utilities in all big data cycle steps, including col­
replacement and maintenance, lecting, preserving, and retrieving different forms of data from various
timetables (e.g., used decision tree sources [122]. These include pattern, data, content, action, decision,
and artificial neural networks), vii) risk, and goal management [123]. ML algorithms improve the potential
deep learning models are used to
of IoT platforms and big data analytics to add value to both of these
identify the energy consumption trend
and flow, and viii) rendering early business segments. There are three categories of IoT data, according to
warning of disruption. the author: (1) raw information or unstructured data [84], (2) metadata
5 ML make the smart grid smarter i) Sensors and ML make by-minute [43], and (3) valued-added or transformed data [124]. ML can help
and decrease the requirement for changes to optimize production assist three of these data categories to categorize, define, and make
power utilities to add new power energy efficiency through adjustment
decisions. By 2024, the global demand for ML in IoT and big data would
plants to wind change, ii) ML-enabled
forecasting anticipates the demand surpass $24 billion. By 2024, embedded ML in the service of IoT and
and supply peaks and increase the objects would be worth $4.6 billion globally.13 Asia Pacific would lead
usages of different intermittent the overall demand for IoT, ML, and AI in big data, followed by North
renewable energy sources, iii) smart
America. By 2024, ML in industrial machines would be worth $415
wired combined with ML to enable the
real-time power system dispatching as million, with collaborative robots growing at a 42.5% compound annual
well as optimize the current building’s growth rate. ML will be a critical AI technology in realizing the full
asset portfolios and current grid loads, potential of IoT, big data, specifically in edge computing platforms.
iv) insect size robots and drones used Fig. 14(a) represents the three-basic structure of sources of data in
to forecast the power system failure
the power grid, including 1) structured data [7]; 2) unstructured data
and inspect the power system
infrastructure without interrupting [125]; and 3) semi-structured data [126]. The dimensions and format
energy generation, v) ML and smart structure varied in terms of many data sources in the smart grid. The
meter enables power utilities to offer data include the distribution stations and electrical information from
different kinds of services which based
distribution switch stations, non-electrical information like marketing,
climate conditions, energy usage, and
other factors, vi) different virtual
smart energy meters, and regional economic and meteorological data.
agents used to automate the call The processing and collection of these data are critical for power sub­
centers which is based on service system operation, plant scheduling, marketing business conduct, and
history, and further ML offers different critical control equipment repair. These kinds of data come from
kind of early warning of bad debts, vii)
different energy sources, for illustrate, smart metering infrastructure,
fields workforce receive real-time
updates as enable to decrease the micro-phasor determining unit, phasor measurement users, remote ter­
response-time to decrease the large minal units, smart appliances, geospatial network topology, climate
impact of outage.

13
Research and Markets, Artificial Intelligence in Big Data Analytics and IoT
systems, on the other hand, are disconnected from information tech­
Markets, 2019–2024: Focus on Data Capture, Information and Decision Support
nology structures and have no network links between them, making
Services, https://www.globenewswire.com/news-release/2019/05/21/1833
them far more challenging to penetrate. In summary, with advanced 672/0/en/Artificial-Intelligence-in-Big-Data-Analytics-and-IoT-Markets-2019-
market forecasting, advanced asset control, and organizational excel­ 2024-Focus-on-Data-Capture-Information-and-Decision-Support-Services.html,
lence by automation, ML has revolutionized the renewable energy Accessed: 03/12/2021.

19
T. Ahmad et al. Renewable and Sustainable Energy Reviews 160 (2022) 112128

Table 7
Renewable energy map 2030, an overview of energy transaction indicators for deploying AI-enabled ML technologies [188].
Description Unit 2000 2012 Renewable Renewable Renewable Reference Compound Compound Indicator for
energy map energy map energy map case 2030 annual growth annual growth renewable
2020 2030 reference (%) rate 2012–2030 rate 2000–2012 energy map
(%/y) (%/y) 2030

Technology indicators
Pumped hydro GWe 689 1004 1350 1600 06 1508 0.6 3.2 –
Hydropower GWe 689 150 225 325 06 306 4.4 N/A –
excluding
pumped storage
Wind offshore GWe 17 06 50 231 242 68 22.5 N/A 300000 of 5
Wind onshore GWe 17 283 600 1404 56 900 9.3 26.4 MWe wind
plants
Concentered solar GWe 0 03 15 83 62 52 12.5 7.6 830 of 100 MWe
power solar plants
Ocean GWe – 01 03 09 519 02 17.3 – –
Solar photovoltaics GWe 08 100 400 1250 184 441 15.1 23.5 12.50 million of
100 kWe solar
plants
Electric vehicles Million – 0.20 25 160 133 69 45.80 N/A 10% of the total
passenger car
fleet
Biomass, EJ/year 28 27 20 12 − 58 29 − 4.3 − 0.0 –
traditional
Battery storage GWe N/A 2.0 25 150 105 73 27.1 N/A 5% of total
parameter
renewables
capacity
Biomass, advanced EJ/year – 01 04 04 88 02 8.4 10.4 270 million 5
for cooking kWth cookstoves
Number of heat Million N/A 04 15 40 58 25 13.3 N/A –
pumps
Biomass heat from EJ/year 01 03 04 14 129 06 9.8 10.2 –
cogen
Heat pump GWth N/A 50 177 474 58 300 13.3 N/A –
Biomass pellets for EJ/year 0.1 01 02 03 49 02 5.80 48.6 16 million 20
heating kWth household
boilers
Geothermal heat EJ/year 0.2 0.5 0.7 1.2 86 0.6 4.3 9.6 –
Biomass pellets for EJ/year 0.1 01 02 03 49 02 5.8 48.6 16 million 20
heating kWth household
boilers
Share in industry % – 01 09 99 968 03 41.8 – –
Biomass chips logs EJ/year – 05 05 06 49 04 1.0 6.4 31 million 20
for heating kWth household
buildings boilers
Share in buildings % 100 99 91 67 − 31 97 10.50 − 31 –
Biomass boilers EJ/year 04 04 05 07 0 07 3.4 − 1.0 0.7 million 1
industry MWth industrial
including biogas boilers
Solar thermal area Million 157 446 1162 4029 163 1532 13.0 11.3 –
m2
Biofuels transport Billion 18 105 214 650 127 287 10.7 15.9 15% of
liters/ worldwide
year transport fuel
consumption
Total biomass use EJ/year 43 51 61 108 37 79 4.3 1.4 20% of total
primary energy
supply
Financial indicators
Fossil fuel subsidies USD – 544 – – – – – – –
billion/
year
Net incremental USD – – – 133 – 0.9% – – –
system cost billion/
year
Subsidies need USD – 101 – 265 – 58% – – –
billion/
year
Net incremental USD – – – 265 – 1.7% – – –
investment billion/
requirements year
Regional indicators based on renewable energy map 2030
Global – modern + (%) – – – 30 – – – – –
access
(%) – 09 – 27 – 14 – – –
(continued on next page)

20
T. Ahmad et al. Renewable and Sustainable Energy Reviews 160 (2022) 112128

Table 7 (continued )
Description Unit 2000 2012 Renewable Renewable Renewable Reference Compound Compound Indicator for
energy map energy map energy map case 2030 annual growth annual growth renewable
2020 2030 reference (%) rate 2012–2030 rate 2000–2012 energy map
(%/y) (%/y) 2030

Worldwide –
modern
renewable
energy excluding
traditional
biomass
Global – modern + (%) – – – 34 – – –
access + energy
efficiency
Global – modern + (%) – – – >36 – – – – –
access + energy
efficiency +
“renewable
energy+”

information, social media, traffic information, data acquisition, and

Table 8 supervisory control, asset inventory, smart plugs, network switches,
The road map of future energy generation and production.
transformers, programmable thermostats, and field measurement de­
Generation source Modern energy storage and Nature of load vices as visualized in Fig. 14(b). Five different characteristics of big data
conversion technologies
in the smart grid environment, including 1) value; 2) volume; 3) ve­
Solar power, Photovoltaic Batteries Building load, racity; 4) velocity; and 5) variety, are shown in Fig. 14(c) [127]. As a
Wind power, panels flow industries load, and result, big data analytics would be essential for the effective operation of
Biomass, Photochemical Alkaline Ion transportation load (e.
Hydropower, conversion batteries g., electric buses and
potential utility power grids and for implementing appropriate business
and Nuclear Wind turbines CO2 storage vehicles) strategies for key stakeholders (i.e., system operators, electric utilities,
Power and capture aggregators, and consumers).
Photochemical Fuel cells Data analytics can include descriptive, predictive, informative, and
conversion
prescriptive analytics, depending on the possible usage cases. As seen in
Hydropower Hydrogen
production Fig. 14, descriptive approaches are often used to characterize grid and
Nuclear power Capacitors consumer organizational behaviors, while diagnostic models examine
Key supporting technologies for electric power and smart energy systems working conditions and grid operator decisions (see Fig. 14(d)). The
Smart grid diagnostic paradigm focuses on determining a case’s triggers, making it
Energy internet Commercial Artificial Analyzing
users intelligence
appropriate for proactive action. Forecasting is often required to antic­
Energy services Household users Cloud Optimizing ipate operational environments and potential decisions since data ana­
producers computing lytics’ primary goal are to include a preventive approach [130].
Energy traders Industrial users Big data Judging Individual functional and devices units produce thousands of terabytes
Energy production Smart buildings IoT Decision-making
of data each year, as visualized in Fig. 14(e). Due to the vast amount
Smart electric – – –
power and units of this type (i.e., sensors, customers, substation), and utility grid
energy systems operations (i.e., home energy control, management of power distribu­
tion, distributed energy resources DERs, and management), power
utilities must maintain millions of terabytes of climate and energy
consumption data, which is expected to grow in the future. As a result,
utilities must examine what increasing data entails for their conven­
tional activities and devise methods to extract value from large data
quantities. According to a global study of industry respondents and 1000
electric utility from ten countries, big data analytics is critical for a
potential smart grid and a source of potential market prospects for the
majority (80%) of electric utilities [129]. The Utility Analytics Institute
predicts that data-related costs will continue to decline, as seen in
Fig. 14(f). Every 14 months or so during the last 30 years, the expense of
storing data has also been reduced. For example, the storage of a giga­
byte of data costs around $11,200 was noted in 1995, $11 in 2000, and
three cents today [131]. Real-time data gathering and distribution have
become commercially viable due to declining data storage and mainte­
nance costs, creating significant utility opportunities to develop profit­
able ML business models.
In addition, extracting and mining useful patterns from a large
amount of data input for forecasting, long-run decision-making, and a
different key to success is inference to big data. However, there are
unique challenges that exist for ML and data analytics for illustrating
different kinds of format variations in the raw data, streaming data that
Fig. 13. Renewable energy integration.
moves at a breakneck pace, the reliability of the predictive analytics,
widely dispersed input sources, high dimensionality, poor and noisy

21
T. Ahmad et al. Renewable and Sustainable Energy Reviews 160 (2022) 112128

Fig. 14. (a) Sources of data in the power grid [128], (b) electrical and non-electrical sources big datasets in the smart grid environment, (c) big data key charac­
teristics of smart grid, (d) an overview of various analytics approaches and their key applications in smart grid, (e) big data pattern volume in power utilities, and (f)
world utilities spending expenditure on data analytics [129].

quality of data, methods scalability, imbalanced energy consumption models in different energy distribution sectors, and detail of each is
and climate input data, un-categorized and unsupervised data, the listed below:
limited number of labeled/supervised data, etc. Big data allows ML al­
gorithms to discover finer-grained trends and render more timely and 3.1. Accurately forecast the energy prices and consumption requirement
precise forecasts than ever before; however, it often poses significant
challenges/obstacles to ML, like model scalability and distributed Consumers and companies gradually generate their electricity as
computation [132]. individual power production (using wind and solar power) becomes
more manageable and cheaper [15]. Users can generate, use, and store
3. Use cases of ML for energy distribution utilities their energy by individual power production. Based on where they
reside, they might also be willing to sell excess electricity back to the
The growth of renewables, the transition to a smarter grid, and local power utility [133]. ML can help determine the most advantageous
marketing tactics transform the electricity generation environment and period to generate, store, or sell this electricity back to the utility grid.
place stress on utilities’ profit margins. On a scale, there is a greater need When rates are low, electricity should be used or retained/stored, and
to make better decisions. To stay successful, these decisions must be when rates are high, it can be sold out to the utility grid. We may render
taken quickly. ML quickly becomes the most powerful method for even more precise hourly estimates using ML algorithms to look at past
making smarter energy distribution decisions and building stronger data, energy consumption patterns (e.g., solar and wind), and weather
consumer relationships. In this section, we will cover the use of ML forecasts. This helps individuals and businesses with electricity

22
T. Ahmad et al. Renewable and Sustainable Energy Reviews 160 (2022) 112128

generation systems make rational choices on how to use their resources.

The adaptive-based neural fuzzy inference approaches, for example,
have been used to forecast short-term wind and solar forecasts for en­
ergy generation [124]. This encourages energy producers to cultivate
energy supply while still selling it back to the grid at peak rates (see
Table 9).
Any power company has to forecast the electricity needs of its con­
sumers accurately [134]. To present, there is no appropriate energy
storage solution, which ensures that the energy has to be distributed and
used almost immediately after it is generated—the use of ML techniques
to improve the predictive performance [135]. Historical data on elec­
tricity usage, weather predictions, and the kinds of companies or
buildings working on a particular day all contribute to calculating the
energy consumption. For instance, if it is a warm summer day in the
middle of the week, office buildings will consume more electricity so
that the air conditioning will be running at full capacity. Weather pre­
dictions and historical data will aid in the early detection of these trends,
preventing rolling blackouts triggered by air conditioning units during
the summer. ML identifies complex variations in a variety of driving
variables (such as the day of the week, duration, forecast solar radiation,
and wind power, past energy demand, major sporting events, air tem­ Fig. 15. The impact of merit order on renewable energy production [123].
perature, mean energy demand, pressure and moisture, and wind di­
rection to understand the energy demand fluctuations. ML allows more production load that must be bought on the power markets is limited
precise decisions than humans since it can detect more energy complex accordingly. As a result, the assured feed-in of energy produced by re­
patterns [136]. This means that when we consume electricity, we can newables reduces energy consumption. The German merit-order curve,
maximize efficiency and lower prices without making costly changes or which is a step function of single plant units in the real world, is
adjustments. simplified as a linear supply curve in the diagram. Reduced demand on
the markets tends to lower costs as long as the slope of this supply curve
is positive [139]. Since energy consumption and renewable electricity
3.2. Accurately forecast the merit order of energy prices production change hourly, calculating the merit-order effect occurs real
value is much more complicated than estimating the market value.
Utility companies can choose from various energy sources, including Permanently falling electricity costs of production – particularly in
alternatives such as wind and solar, as well as fossil fuels and nuclear renewable energy – have shifted the merit order chain, with traditional
power [137]. These various resources are sorted into a price-based merit power plants falling further behind (see Fig. 16).14 The result is obvious
order [138]. This establishes the direction in which the different sources with the increased feed-in of renewable energy sources (i.e., photovol­
of power are sold. ML can be used to evaluate both historical and taics, wind energy, or biomass).
real-time data due to access to information from a wide range of sources. During peak load times, varying wind, and PV power generation
ML algorithms are easier to account for all of the other various pricing forms with different marginal costs near zero (0) value are moving
factors – weather conditions, energy supply, demand, available re­ values into the energy market, bringing traditional power production
sources from the particular sites, historical energy consumption, etc. – to plants to the end of the merit order [140]. The merit order impact of
foresee an optimized order of merit. This is particularly useful in markets clean energy is how the electricity sector relates to this trend [15].
with a lot of penetration of renewable energy sources, like solar and Traditional power plants only meet the domestic load, which is the
wind, since these sources’ energy supply is difficult to predict. remaining energy requirement that renewable energy cannot meet.
Fig. 15 illustrates a stylized summary of sustainable energy pro­ The merit order rate is a steadfast representation algorithm that
duction for 24 h. In the short term, of a one-day-ahead energy market, it works very well for describing the short-term power industry develop­
is believed that energy consumption is inelastic. Since supply providers ment. Determining the long-term evolution of energy rates, on the other
must acquire renewable energy supplies in advance, the residual hand, necessitates the use of a pricing model that accounts for long-term
impacts. Operator decisions on implementation, extension, and
Table 9 decommissioning, and fixed overhead costs will be included in such an
ML mathematical techniques key benefits for the estimated cost of energy energy market model. The latter argument is especially important: no
distribution. power generation plant owner will construct more plants if energy
Sr. Advantages area Sub-class profits only meet the marginal costs. The better-compared dismantling
# and investment costs of nuclear power production stations, for instance,
1 Energy economics Transmission and distribution capital savings are not adequately expressed in the merit order of ML design.
Transmission and distribution operating and
maintenance savings
3.3. Forecast the consumer lifetime value
Energy efficiency
Improved utilization of assets and energy
infrastructure What is the consumer life value? The estimated amount of all
Energy theft reduction and detection at the possible sales (or profit margins) that a single consumer will produce for
distribution level a company is known as consumer lifetime value (CLV or consumer LTV).
Energy cost savings
The company’s earnings will be maximized if reliable CLV forecasts are
2 Energy safety and Energy safety
security Energy security used as the basis for marketing decisions (or profits). Notably,
3 Reliability of power Control of power quality
systems Control of power interruptions at different nodes of
feeders, and disturbing networks 14
Merit order effect (MOE), https://www.next-kraftwerke.com/knowled
4 Climate concerns Greenhouse gas emissions
ge/what-does-merit-order-mean. Accessed: 03/12/2021.

23
T. Ahmad et al. Renewable and Sustainable Energy Reviews 160 (2022) 112128

Fig. 16. Effect of the merit order on consumer energy demand.

estimating the correct consumer LTV forecasts is extremely difficult. compile the data the user carries with them – such as where they reside,
Utility operators and suppliers must give greater attention to indicators what type of device they are using, search history, their browser history,
like consumer lifetime value in an open services market [141,142]. This and how often they have used the utility website – and create an accu­
allows them to estimate how much each customer will pay throughout rate description of that individual as a customer. ML will then assess the
their agreement. ML does more than improve the customer lifetime probability of that individual being a consumer (recognized as scoring)
value and prediction accuracy [143]. We can use ML algorithms like and the best method for converting them into a consumer. This entails
deep artificial neural networks to determine an individual consumer’s utilizing highly targeted advertisements and providing a highly
total value by entering information like purchasing patterns, customer customized interface – for example, displaying a screenshot of a family
information, buying history, location, and payment behavior [144]. We of four on the homepage when a family of four enters or opens the
may also go a step further with ML and recommend maximizing con­ website.
sumer satisfaction. This may include making extremely selective deals
with related customers or using natural language processing (NLP) to
3.5. Make high-targeted offers to consumers
boost support for unhappy customers on the verge of leaving.
Customers have a variety of service suppliers in competitive elec­
3.4. Forecast the probability of winning consumers tricity markets [147]. Personalized or power utilities offerings are sig­
nificant for attracting new consumers and retaining established ones,
To remain competitive in open markets, the power utility companies mainly because brand loyalty is not as high as it once was. ML helps
must have a complete record of future consumers [145]. On the other determine the correct type of deal to make to a specific client at any
hand, ML will provide a more precise overview of consumer energy particular moment by examining buying patterns and customer details.
consumption. It also provides the data we use to enable data-driven Suppose the data shows that a consumer is about to relocate. In that
marketing strategies [62]. This ensures power utilities able to tell case, it may allow them to waive the connectivity charge at their new
when visitors or users visit the website [146]. The ML models used to address. This kind of customized product keeps us ahead of the

Fig. 17. Intelligent automation in utilities and energy distribution: (a) distribution of technical core-function cases by implementation complexity and advantages (b)
distribution of support-function cases by implementation complexity and advantages.

24
T. Ahmad et al. Renewable and Sustainable Energy Reviews 160 (2022) 112128

competition and decreases churn probability with the customers. structures. It is a program used to solve power engineering problems by
Fig. 17 visualized the intelligent automation in utilities and energy converting them into mathematical expressions. Table 10 visualized the
distribution using ML models [51]. Fig. 17(a) reveals that more than power system software used in different energy distribution systems.
one-third (38%) of energy and utility organizations make much effort to Fig. 19 visualizes the power system simulator used to simulate the
adopt low-benefit but straightforward cases. On the other hand, the data and power flow by supplying energy to the miniaturized power
emphasis is less than one out of five (18%) of quick wins. Just 11% of grid.15 Power system infrastructure includes substations, power plants,
organizations rely on accelerated winning in support functions. Just transmission loads, and lines, which are stimulated by ML models of
over one out of four (26%) concentrate on use cases that are complicated miniature power equipment. The simulator is used to analyze the power
to have and less convincing. It is noted that 46% of utility companies grid fault conditions, real-time impact on the utility power grid, and re-
emphasize highly complex and beneficial applications which are used in develop a real-time electric power grid by flowing a small amount of
real-time applications of power system infrastructure (see Fig. 17(b)) electric current in the miniaturized power grid. These kinds of real-time
[51]. The percentage values indicate the implementation of use cases by simulations are useful for grid fault detection, the effect on the utility
utilities and energy organizations in each quadrant. grid, and countermeasures with a high-level penetration of renewable
energy. The data network that confirms international standards
4. ML software for energy distribution IEC61588 and IEC61850 using merging units and smart ML devices
incorporated in each model makes it possible to calculate, control, and
Bottom-up physics-based tools are used for urban building energy monitor the entire miniature grid just like an actual grid.
modeling and planning [148]. The urban modeling includes three basic
classes, including 1) transport analysis and land use; 2) energy modeling 5. Challenges of machine learning for energy distribution
and urban building; and 3) urban system energy modeling. Energy
modeling and urban building are divided into major types, including i) The preferences of consumers are moving to more automated
top-down approaches and ii) bottom-up approaches. The famous tools customized services and applications. The flexibility of grid systems is
for urban building energy planning includes: a) CitySim; b) SimStadt; c) rising behind-the-meter technology such as energy storage devices,
umi; d) CityBES; e) OpenIDEAS; f) CEA; g) URBANopt; and h) TEASER. hybrid cars, solar rooftops, and advanced metering technologies that
CitySim software was developed in C++ and Java language [149]. It enhances grid data availability and granularity. Simultaneously,
aims to facilitate the design of sustainable urbanized areas. It can model competition from electricity distribution threatens firm models of util­
the energy usage of a few to tens of thousands of buildings. SimStadt is ities and accelerates a decrease in consumer loads. The most prominent
an urban energy modeling tool intended to facilitate the energy transi­ challenges of ML models are listed below:
tion preparation at the city level [150]. Energy ADE [151] and CityGML
[152] are proposed in JavaScript and incorporate the 3D city authen­ 1) Challenge of educating consumers: Many businesses face the task of
tication method. These tools are further used to explain the building’s training their consumers in their advanced technology imple­
technical systems and fabrics. The urban modeling interface (UMI) was mentations. The same is true of ML engineers. The current ca­
created to evaluate building energy usage at the community and city pacity of ML is overestimated by entrepreneurs, programmers,
scales and sustainable public transport options, daylighting, outdoor and managers. They hope that the algorithms will easily under­
convenience, and food security [153]. City Building Energy Saver (Cit­ stand and predict complex problems accurately.
yBES) tool was used to optimize and simulate the building energy effi­ 2) New technologies: In reality, it is a relatively recent commercial
ciency/performance at a large-scale level [154]. OpenIDEAS is application of ML, particularly in the field of deep learning. Users
developed to use for a Modelica-based framework. need a wide assortment of well-ordered and planned data to
City energy analyst software is used to help manage energy con­ address questions. To answer questions, consumers need a broad
sumption data [155]. Python developed energy Analysis and Simulation range of well-ordered as well as scheduled data.
for Efficient Retrofit (TEASER) to integrate the USEM and UBEM [156]. 3) Overfitting mechanism: There are millions of variables in a stan­
A detailed analysis of this software used in different building energy dard ML network, perhaps hundreds of millions. Tens of thou­
modeling and analysis is given in Ref. [148]. Roberto et al. conducted a sands of records typically comprise a training data set. Although a
comprehensive study on energy assessment tools and methods [157]. network can recognize the training set and provide 100% precise
Henrik et al. describe the EnergyPLAN that has been widely used in and accurate responses, it can prove completely ineffective when
district cooling and heating systems and gas and electricity grids [158]. new data are provided. The mechanism is referred to as overfits
Fig. 18 (a & b) demonstrates the tools used for urban energy planning and is only one of the constraints of modern deep learning
and management [159]. Energy planners are facilitated in selecting methods in ML.
these tools based on skills, goals, and big data handling. As a result, 4) Talent deficit: Although many individuals and the energy industry
energy planners determine the most appropriate method for a given attract the ML market, few experts still have the opportunity to
application based on their expertise, goals, and data availability. improve this technology. A good data scientist who has not
The power system software can be classified into three major classes: enough experience in information engineering but who knows
1) power plants analysis software; 2) power engineering protection ML.
software; 3) and renewable energy controller software. The power plant 5) ML development: There are more layers in ML development. The
analysis software is based on computations and mathematical algo­ engineers create software that learns to carry out the tasks ex­
rithms. It includes ETAP, DIgSILENT, PSS/E, CYME, etc. Simulating pected to reach the company objectives. It is much more chal­
protection mechanisms for power plants and power systems is another lenging to incorporate just one or two levels.
form of power engineering protection software. This software simulates 6) Advanced energy technologies: Advanced energy technology usu­
the activation of different power system protections, which secure ally needs increasing advanced and complex materials for pro­
power lines, transformers, and other components. The renewable energy cesses and systems. Complexity is a concern for conventional
controller used different software, including DAC, ADC, 4-bit, 8-bit, 16- methods because of the large areas of the variable to investigate.
bit, and much more. The controller language is programmed with Newly developed materials are generally needed for two decades,
computer languages such as C++, C, Java, and others. Overall, software
in the power system for energy distribution is used to model, simulate,
and quantify power plants’ configuration, transmission, electrical grids, 15
Power System Simulator, https://www.fujielectric.com/products/energy_ct
lighting and grounding systems, and many other energy distribution rl_mng/b02.html, Accessed:03/12/2021.

25
T. Ahmad et al. Renewable and Sustainable Energy Reviews 160 (2022) 112128

Fig. 18. Tools used for urban energy planning and management [159].

Table 10
Software used in power system protection.
Software Development Use of license Latest Developer Purpose in energy distribution
name history (C=Commercial) available
version

NAP 1990 C 4.0.1 Innovation Energie Constrained and initial load flow, short circuit, stability, and contingency
Développement analysis calculation.
NEPLAN 1988 C 10.8.1.2 NEPLAN AG Power system analysis, cloud computing, power system management,
real-time integration, grid code, SCADA/GIS integrations, transmission
and distribution networks, and assessment management.
SKM 1972 C 8.0.2.5 SKM Systems HI-WAVE, TMS, CAPTOR, IEE Wiring, and IEC 60909 Fault.
Analysis, Inc
CYME 1986 C 16.01 CYME International Voltage stability analysis and COM module.
PSSE 1976 C – Siemens Over timescales of a few seconds to tens of seconds, as well as in steady-
state situations.
EMTP 1982 C 4.1 EDF & RTE & Hydro- Electromagnetic transients’ programs, insulations issues.
Québec
ERACS 1990 C 3.9.10 RINA Consulting Ltd Modeling tool balanced 3-phase power systems stability analysis which
includes load flow of power systems, short/fault-circuit, protection
coordination, harmonics and G5/4, arc flash calculation modules, and
transient stability.
PSCAD 1986 C 4.003 Manitoba HVDC Short circuit analysis, power system optimization, cable pulling, power
Research Centre flow study, DC and AC arc flash measurement, contingency, and voltage
stability analysis, estimation of transmission line parameters, power
system control, and optimization, proactive device coordination.
XGSLab 2004 C 7.01 SINT Ingegneria GSA, XGSA FD, GSA FD, XGSA TD.
DIgSILENT 1985 C 2018 Dr. Martin Schmieg Grid code, power factor 2018, and station ware 2018.
ETAP 1986 C 19.0.1 Operation Power plant controller, power system management, microgrid controller,
Technology, Inc. SCADA, geospatial modeling, EMS, ADMS, transmission, and distribution
planning.

from discovery to implementation. Further, timing error pro­ storage technologies and creating experimental setups and
hibits new materials from being incorporated into products effi­ detailed evaluations for industrializing and marketing energy
ciently, compromising efficiency and competition. storage. Complete and robust cohesion, appropriate definition,
7) Technology developments: Achievable performance goals, estab­ accountability, openness, and energy storage requirements
lished ambitiously, and a comprehensive research and develop­ should also be established, offering good support to research and
ment portfolio to achieve these goals is also challenging. growth, storage applications, and energy generation, and
8) Energy storage devices and materials: The growing complexities of fostering the energy storage development technologies and
energy storage devices and energy storage systems and the vast related industries that are yet to be challenging.
volume of background data present considerable challenges for 10) Economic Challenges: Today, some countries’ energy storage
standard techniques and algorithms. New cutting-edge technol­ sector faces challenges such as lack of support for the legislation,
ogies overcome the challenges of the conventional methods for high costs, uncertain value, unhealthy business mechanism, and
increased precision, reliability, and optimization. other issues [160]. In the future, it would be important to take
9) Technology challenges: Firstly, energy storage development in­ into consideration two aspects: firstly, the proposal of alterna­
cludes innovation and breakthrough, long-term storage, a high tives to the energy storage scheme involving power generators,
level of security for electrochemical storage, and low cost. electrical companies, researchers, economic and social organi­
Furthermore, high efficiency and physical storage technology zations and, secondly, the promotion of an appropriate business
need low cost. The second point is that research concentrates on competition structure and subsidy strategy for new ML innova­
simulating energy storage and optimizing the process in various tive technologies. Based on Woori’s prediction, the world energy
energy systems, theoretically supporting the utilization of energy storage cost will rise 26% per year in the future, and the total

26
T. Ahmad et al. Renewable and Sustainable Energy Reviews 160 (2022) 112128

Fig. 19. Real-time power system simulator.

energy storage market value in 2020 is noted up to $16 billion potential outages of electricity, and financial loss. The error in the
[160]. While energy storage has several market dynamics, the ML forecasting model lacks the overall system efficacy and per­
key hurdles remain high costs, inadequate subsidy policies, an formance. Increased accuracy and an understanding of these
indeterminate pricing structure, and the business model. improvements to grid loads prevent cost-effectiveness errors and
11) Energy-efficiency perspective: ML strategies’ promise is great in make smarter grid management decisions.
addressing future green 5G and 6G energy management chal­ 15) Solar Energy: Solar power systems and associated technologies
lenges. ML’s methodologies such as federated learning, deep have become renewable energy sources. Their deployment at the
learning, and optimization may be discussed for the design and international level is extensively recognized. Due to the relative
optimization of the network orchestration and architecture cost- cost of installation,
effectively. ML learned network complexity for designing 5G and
6G air interfaces by acquiring dynamic network topology and poor conversion efficiency and battery capacity problems, solar
varying traffic patterns. ML would be more far-reaching and power is still not a widespread source of electricity compared with
critical as saving electricity employing diversified 5G and 6G conventional power sources. Despite the difficulties in deploying ML
technologies, such as smart grid, smart transmission and distri­ techniques, there is much groundbreaking research on emerging prac­
bution networks, smart cities, and factory automation. On the tices such as AI, deep learning, big data, IoT, and new approaches to
other end, coordination and computing are generally needed in enhance the performance of solar power transformation efficiency in the
ML techniques. This could present a major obstacle for the design field of energy are needed. Many companies are working to overcome
and deployment of energy-efficient ML algorithms as well as of the above-said challenges with the use of ML in energy distribution.
future 5G and 6G networks in smart energy systems [161]. Table 11 demonstrates the consortiums, companies, and foundations
12) Demand response: First, demand response agents need to operate working on ML in energy distribution. According to our best knowledge,
in a partially observable atmosphere, i.e., agents are unable to just a few companies are listed in Table 11.
fully understand the units that were used for demand response,
other environments, and agents [162]. 6. ML opportunities towards a smart and sustainable future
13) Intermittent nature of renewable energy: Power utilities have chal­
lenges and prospects for innovation and new technology in the The combination of resilient, low latency connectivity, high-speed,
energy sector. The intermittent nature of renewables requires the and technologies such as the ML will transform towards sustainable
pace of decision-making to leverage renewables that contradict smart energy industry distribution. With the environmental effects of
the conventional prediction models. In the meantime, the power generation and consumption growing, utilities aggressively pur­
deployment of intelligent meters provides a thin and large energy sue novel alternatives to reduce their environmental effects. ML is
usage. Efficient processing and use of this information may pro­ currently being integrated by leaders of both the traditional and
vide new load forecasting prospects. renewable energy markets in many ways to increase alternative energy
14) COVID-19 Impact on energy demand and requirement: The coro­ use while improving the accuracy of renewable energy forecasts, energy
navirus pandemic provides an understanding of the energy use conservation, and usability. Among them, key points of ML towards a
and thus highlights vulnerabilities within utilities’ environments smart and sustainable future are listed below:
as one of the new challenges that utilities face. Poor predictions
or a general lack of forecasts negatively impact public infra­ 1) Energy investors: In the context of an ever more volatile economy,
structure, leading to waste of energy, higher operating costs, energy investors looking towards resilience are re-developing ML

27
T. Ahmad et al. Renewable and Sustainable Energy Reviews 160 (2022) 112128

Table 11 Table 11 (continued )

Consortiums, companies, and foundations working on ML in energy distribution Sr. Name of the Company providing Objective of services
[163]. # company services
Sr. Name of the Company providing Objective of services 9 SmartNet Power grid reliability SmartNet offers resources to
# company services (European Union) and stability enhance cooperation
1 BeeBryte Demand-side BeeBryte is intended to between distribution system
(Singapore, and management and minimize the electricity operators and transmission
France) energy demand using ML algorithms for system operators by sharing
forecasting energy charges and information on the control
automatic controls on heat- and the collection of
cooling devices (for auxiliary services from
example, HVAC), district actors.
generators, charging points 10 McKinsey, Utilityx, Predictive Utilityx supports
for electric vehicles, or (The United States maintenance control infrastructure management
batteries. of America) in predictive engineering to
2 IBM Watson (The Power grid IBM uses analytics to help optimize efficiency.
United States of reliability, and make power-decision Advanced analytics was
America) stability decisions. IBM balances used to turn network data
supply and demand for into a health-driven and
sustainable, stable, and essential asset-based
efficient electricity services approach.
from traditional and 11 Kunumi and PSR Market operations Kunumi and PSR combine
renewable sources. It (Brazil) and advanced energy ML and modern theoretical
monitors and manages grids demand forecasting approaches, including
globally. energy planning, operation,
3 DeepMind, Google Demand-side DeepMind designs programs and trade, to maximize and
(The United States management and to solve the complex nature anticipate uncertainty in
of America) energy demand of power system problems. energy systems.
forecasting In an attempt to reduce 12 Infosys (India) Demand-side By applying ML to data
power consumption, management and generated by the smart
DeepMind tested its ML energy demand meters, advanced sensors,
techniques on Google’s data forecasting and smart devices behind
centers. meters, Infosys supports
4 Tomorrow Demand-side The ML models were stakeholders from the
(Denmark) management and developed by Tomorrow energy sector.
energy demand that systematically derives 13 Grid Edge (United Grid reliability and The company grid edge
forecasting insights from different types Kingdom) stability renders the cloud-based ML
of data on CO2 emissions. models which allow to
Such observations are then optimize, forecast, and
used by other resources, control the supply energy
including the demand.
ElectricityMap, which 14 EWeLiNE Renewable energy EWeLiNE developed the ML
shows the CO2 emissions (Germany) production and models, which are used for
from the development, forecasting renewable energy demand
export, and import of energy forecasting. EWeLiNE
in various regions collects real-time wind
worldwide. turbine and solar data from
5 Verv (United Demand-side To learn about home energy across Germany and
Kingdom) management and products and their incorporates it into a
energy demand behaviors, Verv energy calculating method for
forecasting assistant aims to reduce the renewable energy
household’s energy costs by generation for the
using ML. corresponding 48 h.
6 DCbrain (France) Grid reliability and DCbrain allows flow and
stability consumption optimization,
power system network and power to direct data-driven decisions. It also reduces the electricity
anomaly detection and consumption in a building or manufacturing facility with the same
prevention, and network technology. ML is expected to be applied in several diverse fields of
evolution simulation.
7 EUPHEMIA, N- Optimized energy EUPHEMIA is a method of
energy efficiency and the development of clean energy as well as
SIDE (Europe) market operation coupled energy integrated sustainable future [37]. The key studies or investigations should aim
into European energy to upgrade new energy technologies for energy conservation
markets for determining accompanied by renewables’ increasing use. Recent technical ad­
spot volumes and prices in
vancements in ML technology have allowed applying ML methods in
the European energy
markets the energy and environmental sectors to guarantee sustainable
8 Fraunhofer Power grid reliability The Fraunhofer Institute has growth [164].
(Germany) and stability developed an ML algorithm 2) Smart grid and distribution generation [165]: As an important part of
capable of logging and energy distribution, global warming, and climate change impose a
containing up to 4.3 million
data sets every day,
significant shift in the use of resources and more sustainable use of
analyzing these data to renewable resources: development, supply, and consumption. In the
reliably forecast network energy transformation era, smart grids should improve power net­
operators, identifying and works and systems. The energy market has changed inexorably with
acting on network
the application of integrated grids, microgrids, and technological
abnormalities within 20–50
ms. innovations. The most interesting factor is that smart grids maintain
safety in supply and allow customers to engage as prosumers in the

28
T. Ahmad et al. Renewable and Sustainable Energy Reviews 160 (2022) 112128

energy market. Smart grids use automated and advanced technolo­ high-dimensional input data are now being used effectively. There­
gies to control and track power transmission from all generating fore, recent advances in ML have raised the probability that
sources to meet end-user demand rapidly, quickly, and efficiently. data-orientated materials research would transform breakthroughs
Intelligent smart grids improve flexibility, device durability, and to provide innovative energy material development strategies. In
stability and minimize the climate’s disturbances, expense, and ef­ addition, recent developments in data-driven materials science also
fects. Some emerging technologies, including distributed generation show that ML technologies’ use would make it much simpler and
and microgrids, deliver electricity locally, create broader and more accelerate the discovery and delivery of state-of-the-art energy ma­
stable networks, and lower line overload. Electricity storage com­ terials [168,169]. ML applications is used in different energy mate­
plements renew energy while local energy supplies contribute to the rials, for illustrate, alkaline ion battery materials (e.g., electrolytes,
micro-grids reduction of all blackouts. The microgrid is fitted with electrodes, etc.) [170], catalytic materials [171,172], photovoltaic
local storage and power supply infrastructure based on independent materials (e.g., property screening and forecasting, solar conversion
distributor power supplies. To allow owners to produce their elec­ efficiency, organic PV, etc. (See Fig. 20)) [38,173,174], carbon di­
tricity, microgrids often reduce their energy dependence by oxide capture materials [175,176], and catalytic materials [177].
contributing to cost reduction and avoiding peak consumption costs. 2) Energy storage technologies: In new energy applications such as inte­
The microgrid will generate income if it produces an excess of grating renewable energies and peak load shaving technologies,
electricity allocated or supplied to the energy utilities. Whatever is energy storage technologies are well recognized. Fig. 21 visualized
said above is possible only due to advanced ML technologies. the operational power range versus the time of many new energy
3) ML reshaping the energy industry [92]: Renewable energy will be storage technologies. It demonstrates the suitability of many utility
supplied by 2050 up to 70% of the world’s total energy demand applications. For renewable energy integration and utility, life cycle,
[166]. The method we consume electricity is simultaneously output power, and energy storage capacity are the key criteria of
evolving. Large consumption of electric cars is projected in the material capacity and storage performance. In particular, the use of
coming years. Digitalization enables these key patterns of decar­ ML and the proliferation of storage would allow renewable energy to
bonization and decentralization. This reorganization of the energy be incorporated and dispatched to promote smarter grids that
distribution is made possible by modern ML technologies. depend less as well as inefficient power generation plants.
4) Data-driven strategies [167]: Data-driven strategies with the deploy­ 3) The United States Energy Department: The United States National
ment of ML in the utilities and power industry renders the platform Laboratories and Energy Department established an environment of
for management discussion around the implementation of research to achieve their scientific, technology, and safety missions
data-driven management and data-science to deliver energy strategy using ML techniques. This environment expands and enhances ML
in long-term perceptive, and consequently, competitive advantage programs, software, algorithms, processes, and applications. DOE
for a sustainable future. and its labs work together to promote technology advancement that
5) Decision making [136]: ML can enhance AI system efficiency for en­ ensures stability in the United States and competition with other
ergy distribution decision-making. Reinforcement learning can be federal departments, the industry, and academia [179].
used in any situation requiring decision-making in an inconsistently
changing environment. Machine control, robotics, and tuning (i.e., The United States Energy Department proposed different strategies
heating, ventilation, and air conditioning), optimization of the sup­ towards the use of ML in energy distribution. Strategy-1: initiate long-
ply chain (i.e., ML demonstrates how much of a given object to buy term investments in ML area for energy distribution; Strategy-2: the
from when and where), and crop yield optimization are all examples. development of effective techniques for human-ML collaboration;
Reinforcement learning (RL), a sub-class of ML, nearly often gener­ Strategy-3: address as well as understand the ethical, legal, and societal
ally requires instruction in a simulated environment. How will RL be repercussions of ML; Strategy-4: ensure the security and safety of ML
used to meet the Sustainable Development Goals? By training robots, systems; Strategy-5: proposed shared environments and shared public
drones, and smart devices how to make better choices in diverse use datasets for ML testing and training; Strategy-6: evaluate and measure
cases, RL will help us meet any or all of the Sustainable Development the ML techniques through benchmarks and standards; Strategy-7: better
Goals (SDGs), for illustrate, (Sustainable Development Goal 3 for understanding for research and development of ML techniques; and
well-being and health), (Sustainable Development Goal 9 for infra­ Strategy-8: expansion of the private and public partnership to accelerate
structure and industry innovation), (Sustainable Development Goal the advances in ML techniques.
12 for production and responsible consumption), etc.
4) The use of ML in Energy Science: The science office has allocated 13
7. Recent progress on properties and discovery of machine million US dollars to the development of ML as a platform for
learning in the energy distribution forecasting and scientific investigation. Developing applications
and algorithms for particular scientific programs would cost
It is widely expected that ML can change almost every part of our about $11 million.16 The remaining funds will be used to boost
society. The most commonly used subset of AI today is ML techniques. the predictability of ML models. The electricity office reported
Deep learning models or a subset of ML utilizes multi-level neural net­ federal support for many ML-related initiatives totaling $7
works in ML to learn from the massive amount of data stored from million. These initiatives are intended to help in the application
different energy sources, devices, and power system infrastructure. The and development of faster grid simulation and analytics in the
key point on the recent progress of ML in energy distribution is listed future.12 The “Advanced Research Projects Agency-Energy” has
point-by-point below: launched a new initiative called design intelligence for energy

1) New energy materials [7]: Recently, the screening of new materials

and their simulation of quantitative relations of structural behavior 16
Department of Energy, Department of Energy Announces $20 Million for
have become a hot and trendy subject in energy materials because of
Artificial Intelligence Research, https://www.energy.gov/articles/department
the many opportunities and challenges, including a low probability -energy-announces-20-million-artificial-intelligence-research, Accessed: 03/
of performance, high computational costs, and time consumption 12/2021.
connected with classical techniques of energy material development
[7]. Currency algorithms for regression, classification, clustering or
dimensional reduction of large collections of particularly

29
T. Ahmad et al. Renewable and Sustainable Energy Reviews 160 (2022) 112128

Fig. 20. The processes of ML emerging technologies for advanced organic solar cell stages of development [7].

5) Renewable Energy and Grid Modernization [180]: The United States

DOE and its national laboratories are taking advantage of recent
developments in ML to further use their vast computing capa­
bilities and technological workforce in favor of grid automation
and energy conservation, and sustainable energy utilization.
6) National Security and Cybersecurity [37]: ML technologies built by
the United States DOE and its National Laboratories prevent
malicious cyber-attacks on critical energy infrastructure, and
reduce the grid disruptions and shutdowns. Preventing power
failures has major social and economic implications around the
world. The ML technologies facilitate the modernization of the
power grid using autonomous energy systems. The long-term
objective is fully achieved through key objectives: 1) secure; 2)
flexible; 3) scalable; 4) robust; 5) resilient; 6) affordable; 7)
reliable; and 8) real-time interoperability.
7) Geothermal: ML techniques that are successfully implemented
could aid in the exploration of geothermal wells, improve drilling
accuracy, and lower costs.18
8) Biomass: The Idaho National Laboratory scientists utilize the ML
Fig. 21. Applications of energy storage technologies for energy utilities [178]. models for data analysis on bio-refinery processing to direct
operational design changes that optimize performance while
mitigating device damage.19
consumption reduction and advanced technology upgrades,
which would grant up to $20 million to projects.17 ML can be
implemented in energy systems by project teams and instigating
energy technology development.
18
Office of Energy Efficiency & Renewable Energy, Energy Department
Awards $5.5 Million to Apply Machine Learning to Geothermal Exploration,
https://www.energy.gov/eere/articles/energy-department-awards-55-mill
17
Design Intelligence Fostering Formidable Energy Reduction and Enabling ion-apply-machine-learning-geothermal-exploration. Accessed: 03/12/2021.
19
Novel Totally Impactful Advanced Technology Enhancements, https://arpa-e. Artificial intelligence helps turn biomass into energy, https://inl.gov
energy.gov/technologies/programs/differentiate, Accessed: 03/12/2021. /article/systems-engineering/. Accessed: 03/12/2021.

30
T. Ahmad et al. Renewable and Sustainable Energy Reviews 160 (2022) 112128

9) ML Supercomputing Capabilities: The US DOE has success in the markets (e.g., consumption patterns, price fluctuations, produc­
national laboratories’ world-class supercomputing capabilities. tion metrics, and historical prices, etc.) [186].
The labs are also the home of four of the world’s ten fastest su­
percomputers – establishing the US as the worldwide leader in the 8. Conclusion
area of high-tech ML computing. The most powerful supercom­
puters include: 1) Cascade (e.g., 3 Petaflops); 2) Cori (e.g., 28 The supply of continuous energy to consumers has many challenges
Petaflops); 3) Mira (e.g., 10 Petaflops); 4) Joule 2.0 (e.g., 6 to address for power grids. Energy usage, weather patterns, even wild­
Petaflops); 5) Sequoia (e.g., 20 Petaflops); 6) Lassen (e.g., 23 card incidents, internal failure, interferences, and lightning strokes from
Petaflops); 7) Theta (e.g., 7 Petaflops); 8) BeBop (e.g., 2 Peta­ wild animals can impact power delivery in energy distribution networks.
flops); 9) Summit (e.g., 200 petaflops); and 10) Trinity (e.g., 41 The ML is increasingly being applied to help forecast potential brownout
petaflops). Exascale computing is the next generation of conditions. This study covers the detailed review analysis for data-
computing systems. The Exascale computer can perform quintil­ driven probabilistic ML and its application for energy distribution. ML
lion computations per second, demonstrating the unprecedented models have been applied in a variety of applications for energy distri­
breakthroughs in ML. More detailed analysis and recent progress bution. Power utilities and grid operators use ML to classify grid failure
concerning ML techniques’ supercomputing capabilities are conditions at generation, transmission, and the distribution level. Smart
given in Ref. [179]. meters and sensors such as phase measurement units render energy
10) Technology Advancement: To simulate the electrical system and consumption information in real-time. When we combine ML tools with
the complexity of different issues much faster than possible with simulated and historical data, ML is used to help and mitigate grid
current commonly used techniques, Argonne scientists are failure conditions, using different models such as demand-response
working on optimization methods that use ML and AI. The key optimization and grid balancing. Furthermore, recent advances in
emphasis is to speed up the electric system planning and daily data-driven systems engineering show that ML techniques can be used to
load flow analysis and calculations. facilitate and improve the deployment and development of advanced
11) Demand Response [181]: The demand response management and energy materials and electric infrastructure at the energy distribution
infrastructure are intended to track and use the real-time energy level.
consumption information to deliver the energy prices to thou­ ML models are used to figure out the fault conditions more precisely,
sands of consumers in real-time via the utility power grid. Cus­ supporting consumers to mitigate service interruption. For example,
tomers may respond to these rates and enable them to change discrete wavelet transforms with a SVM are used to locate the fault
their energy consumption in terms of gird conditions. Continued condition at energy distribution networks and systems. The grid balance
development can boost stability, cost savings, and sustainability - ensuring that electricity supplies meet energy demand - is one of the
by helping end-user applications consider how they need power main tasks for energy utilities and grid operators. However, weather-
grid changes. Information and such resilience would encourage dependent renewable sources make energy forecasting even harder.
the potential penetration of renewable energy into the power The most accurate tool for forecasting the demand for renewable energy
grid. generation is ML in energy distribution. ML leads to many weather
12) High Fidelity: Argonne National Laboratory increases high-fidelity variables (e.g., wind speed, wind direction, solar radiation, etc.) and
simulation with ML, such that the optimization of the design makes perfect energy forecasts by applying specialized methods such as
process is accelerated significantly while sustaining the data short-term neural networks. This saves operators money and conserves
durability. In this scenario, a task or job that takes hours using power plant energy.
high-fidelity alone takes a milliseconds optimization process More and more cyberattacks are being used to target important
augmented by ML. power system infrastructure, such as hijacking the power system control
13) Smart Energy Materials Manufacturing [38]: Advanced processing and demanding money. Utilities and grid operators use ML to distin­
techniques require multi-component structures and intricate guish between a malfunction (for illustrating, short circuit faults) or a
thermo-chemical processes resulting in a high data volume, often disruption in the power grid and a smart cyber-attack (such as false data
at a high pace. Optimally, operators can get immediate data/­ injection to the power system). Symbolic dynamic filtering theory is
feedback on the characteristics of manufacturing of energy ma­ used for feature extraction to explore causal interactions between the
terials and process variables in real-time to find new processes subsystems without overburdening computer systems. When tested at
and phenomena faster and adapt to increase effectiveness and power distribution systems, 99% of cyber-attacks were correctly iden­
efficiency. However, existing methods offer “postmortem” tified with a true positive rate of 98% and a demonstrably false rate of
data/feedback long after the manufacturing process is completed. less than 2%.
ML methods are used to optimize the analyze the manufacturing The use of ML models has been well recognized in energy market
process. information management (e.g., financial assurance, market mitigation
14) Solar Energy breakthroughs with ML [58]: There are five directions and monitoring), market operations (e.g., economic dispatch, unit
where ML takes advantage of advances in the field of solar commitment, security constrained, energy reserve calculation and re­
technology. These areas of research include 1) smart infrastruc­ quirements, system black start capability, voltage control and reactive
ture design for solar energy forms and systems (e.g., optimal supply), power system modeling (e.g., data acquisition, supervisory
sizing, weather conditions, accurate geographical placement, control, and power system modeling), outage and forecast (e.g., climate
etc.); 2) smart maintenance of solar power plants (e.g., advanced forecast, energy demand forecast, and outage scheduling), transmission
sensors, failure prediction, anomaly detection, and real-time scheduling (e.g., tagging, external scheduling, transmission capacity
power system data) [182]; 3) solar energy generation fore­ measurement, and energy distribution network reservation), grid oper­
casting (e.g., historical satellite data, environmental and ations (e.g., dispatch management, state estimation, dynamic limit
real-time forecast information can be integrated into and evalu­ estimation, automatic energy production control, resource performance
ated using optimization analytics to support hardware mainte­ monitoring, voltage stability analysis, resource capacity analysis, flow
nance decisions) [183]; 4) optimization process of power gate calculation and power flow, and contingency analysis), grid set­
transmission and distribution networks (e.g., integration of en­ tlements (e.g., internal scheduling, billing, energy tariffs, metering,
ergy consumption data [184], proactive monitoring of distrib­ reliability, and ancillary services), utility organization (e.g., customer
uted grid, etc.) [185]; and 5) well understanding of solar energy data management), energy distribution control (e.g., load monitoring,
alarm processing, and logging), energy security (e.g., policy

31
T. Ahmad et al. Renewable and Sustainable Energy Reviews 160 (2022) 112128

management, energy authentication and authorization), energy demand References

and consumption data management (e.g., data visualization, business
intelligence, data archive and historian). [1] McCollum DL. Machine learning for energy projections. Nat Energy 2021:1–2.
https://doi.org/10.1038/s41560-021-00779-9. News & Vie.
Recent progress in science and engineering data-driven materials has [2] Lu T, Sherman P, Chen X, Chen S, Lu X, McElroy M. India’s potential for
shown that ML technologies’ implementation will make the design, integrating solar and on- and offshore wind power into its energy system. Nat
research, deployment, and development of advanced energy materials Commun 2020;11:1–10. https://doi.org/10.1038/s41467-020-18318-7.
[3] Zhang Weiping, Maleki * Akbar, Rosen Marc A. A heuristic-based approach for
much easier. This study explores that is the most widely used ML tech­ optimizing a small independent solar and wind hybrid power scheme
niques in energy distribution development and research include super­ incorporating load forecasting. J Clean Prod 2019;241:18.
vised learning (e.g., support vector regression, SVM, RVM, decision [4] Schiermeier Q. And now for the energy forecast. Nature 2016;535:212–3.
[5] Omer AM. Focus on low carbon technologies: the positive solution. Renew
tree/RF, and regression models), unsupervised learning, and deep Sustain Energy Rev 2008;12:2331–57. https://doi.org/10.1016/j.
learning (for illustrating, EML, artificial neural networks, RNN, CNN- rser.2007.04.015.
based methods). Several modern ML techniques combine multiple [6] David TW, Anizelli H, Jacobsson TJ, Gray C, Teahan W, Kettle J. Enhancing the
stability of organic photovoltaics through machine learning. Nano Energy 2020;
methods and algorithms (i.e., integrating RNN and CNN, combining
78:105342. https://doi.org/10.1016/j.nanoen.2020.105342.
DNN and RNN) to achieve the highest level of training, test, and vali­ [7] Liu Y, Esan OC, Pan Z, An L. Machine learning for advanced energy materials.
dation learning performance. Energy AI 2021;3:100049. https://doi.org/10.1016/j.egyai.2021.100049.
The key breakthrough of successful ML application in energy dis­ [8] IRENA. Renewable energy highlights. 2020. p. 1–2. https://doi.org/https://
www.irena.org/-/media/Files/IRENA/Agency/Publication/2020/Jul/
tribution and remaining challenges is pointed out like standardization Renewable_energy_highlights_July_2020.pdf?
and improvement of energy infrastructure, model visualization and la=en&hash=75B114DB7A55F4260F41F64C4DFF793DB2044306.
automatic closed-loop optimization, experimental exploration, sup­ [9] Mai T, Hand MM, Baldwin SF, Wiser RH, Brinkman GL, Denholm P, et al.
Renewable electricity futures for the United States. IEEE Trans Sustain Energy
porting policies, and interdisciplinary communications. The energy in­ 2014;5:372–8. https://doi.org/10.1109/TSTE.2013.2290472.
dustry moves from a reliable, low-risk environment, regulation-driven [10] DSIRE. Database of state incentives for renewables & efficiency. NC Clean Energy
to a modern technology-driven, more sophisticated, as well as uncertain Technol Cent NC State Univ; 2021. . [Accessed 28 February 2021].
[11] (SEPA) SEPA. Utility carbon reduction tracker. Sepa 2020 [https://doi.org/
marketplace with the use of ML models. Changes, as well as emerging https://sepapower.org/utility-carbon-reduction-tracker/].
technology and business models used in energy distribution, would be [12] Kuhn T. Delivering America’s energy future. Elec Perspect 2018;43:6–7 [https://
essential. Unrecognizing this could threaten the viability of power doi.org/https://www.eei.org/issuesandpolicy/finance/wsb/Documents/EEI_
WSB_Remarks.pdf].
companies. Our study findings show that ML can play an important role [13] IRENA. REmap 2030 - a renewable energy roadmap./Publications/2014/Jun/
in navigating and managing this emerging energy distribution system. It REmap-2030-Full-Report 2014:188. https://doi.org/https://www.irena.org/-/
will probably be one of the key enablers. However, the utilities seem media/Files/IRENA/Agency/Publication/2014/IRENA_REmap_summary_
findings_2014.pdf.
reluctant to adopt this available resource (e.g., ML techniques) due to
[14] Mutezo G, Mulopo J. A review of Africa’s transition from fossil fuels to renewable
unknown uncertainties. We believe this is a mistake. As such, this energy using circular economy principles. Renew Sustain Energy Rev 2021;137:
study’s recommendations illustrate how power companies can take 110609. https://doi.org/10.1016/j.rser.2020.110609.
advantage of the benefits of ML and keep ahead of the ML curve rela­ [15] Antonopoulos I, Robu V, Couraud B, Kirli D, Norbu S, Kiprakis A, et al. Artificial
intelligence and machine learning approaches to energy demand-side response: a
tively quickly. But to do this, they have to become smart and act quickly. systematic review. Renew Sustain Energy Rev 2020;130:109899. https://doi.org/
10.1016/j.rser.2020.109899.
Credit author statement [16] IRENA. Innovation landscape for a renewable-powered future: solutions to
integrate variable renewables. Abu Dhabi: Int Renew Energy Agency; 2019.
p. 1–164.
Tanveer Ahmad: Conceptualization, Methodology, Data curation, [17] IRENA. Global energy transformation. Int Renew Energy Agency 2018:1–76.
Writing – original draft, Visualization, Validation, Writing- Reviewing [18] Murphy KP. Probabilistic machine learning: an introduction (adaptive
computation and machine learning). MIT Press; 2020. p. 1–932.
and Editing. Rafal Madonski: Methodology; Writing- Reviewing and [19] Murphy KP. Probabilistic machine learning: an introduction (adaptive
Editing. Dongdong Zhang: Investigation; Methodology; Writing – orig­ computation and machine learning). MIT Press; 2020. p. 1–932. Massachuse.
inal draft. Chao Huang: Conceptualization; Funding acquisition. Asad [20] Cheng L, Yu T. A new generation of AI: a review and perspective on machine
learning technologies applied to smart energy and electric power systems. Int J
Mujeeb: Data curation; Methodology. Energy Res 2019;43:1928–73. https://doi.org/10.1002/er.4333.
[21] Donti PL, Kolter JZ. Machine learning for sustainable energy systems. Annu Rev
Declaration of competing interest Environ Resour Oct. 2021;46(1):719–47. https://doi.org/10.1146/annurev-
environ-020220-061831.
[22] Sharmila P, Baskaran J, Nayanatara C, Maheswari R. A hybrid technique of
The authors declare that they have no known competing financial machine learning and data analytics for optimized distribution of renewable
interests or personal relationships that could have appeared to influence energy resources targeting smart energy management. Procedia Comput Sci 2019;
the work reported in this paper. 165:278–84. https://doi.org/10.1016/j.procs.2020.01.076.
[23] Perera ATD, Kamalaruban P. Applications of reinforcement learning in energy
systems. Renew Sustain Energy Rev Mar. 2021;137:110618. https://doi.org/
Acknowledgments 10.1016/j.rser.2020.110618.
[24] Westermann P, Evins R. Using Bayesian deep learning approaches for
uncertainty-aware building energy surrogate models. arXiv:2010.03029 [cs, stat].
This work was supported in part by the National Key Research and Oct. 2020 [Online]. Available: http://arxiv.org/abs/2010.03029. [Accessed 21
Development Program of China (Grant No. 2019YFE0118000), the Key November 2021].
Laboratory of Special Machine and High Voltage Apparatus (Shenyang [25] Sun M, Zhang T, Wang Y, Strbac G, Kang C. Using bayesian deep learning to
capture uncertainty for residential net load forecasting. IEEE Trans Power Syst
University of Technology), Ministry of Education, Grant KFKT202006, Jan. 2020;35(1):188–201. https://doi.org/10.1109/TPWRS.2019.2924294.
and Guangxi Young and Middle-aged Scientific Research Basic Ability [26] Pfingsten T. Fully bayesian inference and model selection for Gaussian process
Promotion Project, 2020KY01009, National Natural Science Foundation regression. 2019. https://doi.org/10.13140/RG.2.2.21150.38727.
[27] Munkhammar J, Widen J, Grahn P, Ryden J. A Bernoulli distribution model for
of China under Grant 62002016, in part by Beijing Natural Science plug-in electric vehicle charging based on time-use data for driving patterns. In:
Foundation under Grant 9204028, in part by Guangdong Basic and 2014 IEEE international electric vehicle conference (IEVC), florence; Dec. 2014.
Applied Research Foundation under Grant 2019A1515111165. The p. 1–7. https://doi.org/10.1109/IEVC.2014.7056224.
[28] Coulier A. Multiscale modeling in systems biology: methods and perspectives.
authors are grateful for the sponsorship of the UM-Macau Postdoc Talent
Acta Univ Upsal 2021;2051. https://doi.org/10.33063/diva-442412.
Program under the cover of the Government of Macau, SAR. [29] Pu B, Nan F, Zhu N, Yuan Y, Xie W. Ufngbm (1,1): a novel unbiased fractional
grey Bernoulli model with Whale Optimization Algorithm and its application to
electricity consumption forecasting in China. Energy Rep Nov. 2021;7:7405–23.
https://doi.org/10.1016/j.egyr.2021.09.105.

32
T. Ahmad et al. Renewable and Sustainable Energy Reviews 160 (2022) 112128

[30] Saidi AYN, Ramli NA, Muhammad N, Awalin LJ. Power outage prediction by [55] Zhu K, Liu X, Pong PWT. On-site real-time current monitoring of three-phase
using logistic regression and decision tree. J. Phys.: Conf. Ser. Jul. 2021;1988(1): three-core power distribution cables with magnetic sensing. Proc IEEE Sensors
12039. https://doi.org/10.1088/1742-6596/1988/1/012039. 2018;2018-Octob:1–4. https://doi.org/10.1109/ICSENS.2018.8589880.
[31] Terrén-Serrano G, Martínez-Ramón M. Multi-layer wind velocity field [56] Power I, Society E. IEEE guide for application of optical instrument transformers
visualization in infrared images of clouds for solar irradiance forecasting. Appl for protective relaying IEEE guide for application of optical instrument
Energy Apr. 2021;288:116656. https://doi.org/10.1016/j. transformers for protective relaying. EEE std C37241-2017. 2018. p. 1–50.
apenergy.2021.116656. https://doi.org/10.1109/IEEESTD.2018.8307291.
[32] Xu S, Wu W, Yang Y, Wang B, Wang X. Analytical solution of stochastic real-time [57] Ghoddusi H, Creamer GG, Rafizadeh N. Machine learning in energy economics
dispatch incorporating wind power uncertainty characterized by Cauchy and finance: a review. Energy Econ 2019;81:709–27. https://doi.org/10.1016/j.
distribution. IET Renew Power Gener Jul. 2021;15(10):2286–301. https://doi. eneco.2019.05.006.
org/10.1049/rpg2.12163. [58] Wei J, De Luna P, Bengio Y, Aspuru-Guzik A, Sargent E. Use machine learning to
[33] Li C, Conejo AJ, Liu P, Omell BP, Siirola JD, Grossmann IE. Mixed-integer linear find energy materials. Nature 2017;552:23–5. https://doi.org/10.1038/d41586-
programming models and algorithms for generation and transmission expansion 017-07820-6.
planning of power systems. Eur J Oper Res Mar. 2022;297(3):1071–82. https:// [60] Jain A, Ong SP, Hautier G, Chen W, Richards WD, Dacek S, et al. Commentary:
doi.org/10.1016/j.ejor.2021.06.024. the materials project: a materials genome approach to accelerating materials
[34] Pallonetto F, De Rosa M, D’Ettorre F, Finn DP. On the assessment and control innovation. Apl Mater 2013;1:1–12. https://doi.org/10.1063/1.4812323.
optimisation of demand response programs in residential buildings. Renew [61] Chen C, Zuo Y, Ye W, Li X, Deng Z, Ong SP. A critical review of machine learning
Sustain Energy Rev Jul. 2020;127:109861. https://doi.org/10.1016/j. of energy materials. Adv Energy Mater 2020;10:1–36. https://doi.org/10.1002/
rser.2020.109861. aenm.201903242.
[35] Borunda M, Jaramillo OA, Reyes A, Ibargüengoytia PH. Bayesian networks in [62] Zhang L, Wen J, Li Y, Chen J, Ye Y, Fu Y, et al. A review of machine learning in
renewable energy systems: a bibliographical survey. Renew Sustain Energy Rev building load prediction. Appl Energy 2021;285:116452. https://doi.org/
Sep. 2016;62:32–45. https://doi.org/10.1016/j.rser.2016.04.030. 10.1016/j.apenergy.2021.116452.
[36] Amarasinghe K, Marino DL, Manic M. Deep neural networks for energy load [63] Ahmad R, Alsmadi I. Machine learning approaches to IoT security: a systematic
forecasting. In: 2017 IEEE 26th international symposium on industrial electronics literature review. Internet of Things 2021;14:100365. https://doi.org/10.1016/j.
(ISIE), edinburgh, United Kingdom, Jun; 2017. p. 1483–8. https://doi.org/ iot.2021.100365.
10.1109/ISIE.2017.8001465. [64] Kang Y, Li L, Li B. Recent progress on discovery and properties prediction of
[37] Ahmad T, Zhang D, Huang C, Zhang H, Dai N, Song Y, et al. Artificial intelligence energy materials: simple machine learning meets complex quantum chemistry.
in sustainable energy industry: status quo, challenges and opportunities. J Clean J Energy Chem 2021;54:72–88. https://doi.org/10.1016/j.jechem.2020.05.044.
Prod 2021;289:125834. [65] Turetskyy A, Wessel J, Herrmann C, Thiede S. Battery production design using
[38] Yılmaz B, Yıldırım R. Critical review of machine learning applications in multi-output machine learning models. Energy Storage Mater 2021:1–24.
perovskite solar research. Nano Energy 2021;80:105546. https://doi.org/ [66] Rahnama A, Zepon G, Sridhar S. Machine learning based prediction of metal
10.1016/j.nanoen.2020.105546. hydrides for hydrogen storage, part I: prediction of hydrogen weight percent. Int
[39] Amasyali K, El-Gohary N. Machine learning for occupant-behavior-sensitive J Hydrogen Energy 2019;44:7337–44. https://doi.org/10.1016/j.
cooling energy consumption prediction in office buildings. Renew Sustain Energy ijhydene.2019.01.261.
Rev 2021;142:110714. https://doi.org/10.1016/j.rser.2021.110714. [67] Zhou Y, Zheng S, Zhang G. A state-of-the-art-review on phase change materials
[40] Xu X, Xu X, Jia Y, Xu Y, Xu Z, Xu Z, et al. A multi-agent reinforcement learning- integrated cooling systems for deterministic parametrical analysis, stochastic
based data-driven method for home energy management. IEEE Trans Smart Grid uncertainty-based design, single and multi-objective optimisations with machine
2020;11:3201–11. https://doi.org/10.1109/TSG.2020.2971427. learning applications. Energy Build 2020;220:110013. https://doi.org/10.1016/
[41] Du P. Ensemble machine learning-based wind forecasting to combine NWP output j.enbuild.2020.110013.
with data from weather station. IEEE Trans Sustain Energy 2019;10:2133–41. [68] Gao T, Lu W. Machine learning toward advanced energy storage devices and
https://doi.org/10.1109/TSTE.2018.2880615. systems. iScience 2021;24:101936. https://doi.org/10.1016/j.isci.2020.101936.
[42] Hossain Lipu MS, Hannan MA, Hussain A, Saad MH, Ayob A, Uddin MN. Extreme [69] Wang Y, Liu J, Han Y. Production capacity prediction of hydropower industries
learning machine model for state-of-charge estimation of lithium-ion battery for energy optimization: evidence based on novel extreme learning machine
using gravitational search algorithm. IEEE Trans Ind Appl 2019;55:4225–34. integrating Monte Carlo. J Clean Prod 2020;272:122824. https://doi.org/
https://doi.org/10.1109/TIA.2019.2902532. 10.1016/j.jclepro.2020.122824.
[43] Tayal A, Solanki A, Singh SP. Integrated frame work for identifying sustainable [70] Seferlis P, Varbanov PS, Papadopoulos AI, Chin HH, Klemeš JJ. Sustainable
manufacturing layouts based on big data, machine learning, meta-heuristic and design, integration, and operation for energy high-performance process systems.
data envelopment analysis. Sustain Cities Soc 2020;62:102383. https://doi.org/ Energy 2021;224:120158. https://doi.org/10.1016/j.energy.2021.120158.
10.1016/j.scs.2020.102383. [71] Narciso DAC, Martins FG. Application of machine learning tools for energy
[44] King JE, Jupe SCE, Taylor PC. Network state-based algorithm selection for power efficiency in industry: a review. Energy Rep 2020;6:1181–99. https://doi.org/
flow management using machine learning. IEEE Trans Power Syst 2015;30: 10.1016/j.egyr.2020.04.035.
2657–64. https://doi.org/10.1109/TPWRS.2014.2361792. [72] Ikeda S, Nagai T. A novel optimization method combining metaheuristics and
[45] Zhang D, Li S, Sun M, O’Neill Z. An optimal and learning-based demand response machine learning for daily optimal operations in building energy and storage
and home energy management system. IEEE Trans Smart Grid 2016;7:1790–801. systems. Appl Energy 2021;289:116716. https://doi.org/10.1016/j.
https://doi.org/10.1109/TSG.2016.2552169. apenergy.2021.116716.
[46] Huang C, Zhang H, Song Y, Wang L, Ahmad T, Luo X. Demand response for [73] Ahmad T, Chen H. A review on machine learning forecasting growth trends and
industrial micro-grid considering photovoltaic power uncertainty and battery their real-time applications in different energy systems. Sustain Cities Soc 2020;
operational cost. IEEE Trans Smart Grid 2021:In press. https://doi.org/10.11 54:102010. https://doi.org/10.1016/j.scs.2019.102010.
09/TSG.2021.3052515. [74] Ahmad T, Zhang D. Using the internet of things in smart energy systems and
[47] Susto GA, Schirru A, Pampuri S, McLoone S, Beghi A. Machine learning for networks. Sustain Cities Soc 2021;68:102783. https://doi.org/10.1016/j.
predictive maintenance: a multiple classifier approach. IEEE Trans Ind Inf 2015; scs.2021.102783.
11:812–20. https://doi.org/10.1109/TII.2014.2349359. [75] Ali SS, Choi BJ. State-of-the-art artificial intelligence techniques for distributed
[48] Ibrahim MS, Dong W, Yang Q. Machine learning driven smart electric power smart grids: a review. Electron 2020;9:1–28. https://doi.org/10.3390/
systems: current trends and new perspectives. Appl Energy 2020;272:115237. electronics9061030.
https://doi.org/10.1016/j.apenergy.2020.115237. [76] Tayal D. Achieving high renewable energy penetration in Western Australia using
[49] Zhou Y, Zheng S, Liu Z, Wen T, Ding Z, Yan J, et al. Passive and active phase data digitisation and machine learning. Renew Sustain Energy Rev 2017;80:
change materials integrated building energy systems with advanced machine- 1537–43. https://doi.org/10.1016/j.rser.2017.07.040.
learning based climate-adaptive designs, intelligent operations, uncertainty-based [77] Advances S. Using AI to improve energy and resource efficiency in various
analysis and optimisations: a state-of-the-art review. Renew Sustain Energy Rev industries. Eur Community 2021. https://doi.org/https://cordis.europa.eu/
2020;130:109889. https://doi.org/10.1016/j.rser.2020.109889. article/id/415798-using-ai-to-improve-energy-and-resource-efficiency-in-
[50] Kellner T. Wind in the cloud? How the digital wind farm will make wind power various-industries, . [Accessed 3 March 2021].
20 percent more efficient. GE Reports 2015:1–40. https://doi.org/http://www. [78] Tomazzoli C, Scannapieco S. Machine learning for energy efficiency: automatic
gereports.com/post/119300678660/wind-in-the-cloud-how-the-digital-wind- detection of electric loads from power consumption. In: 2017 AEIT int annu conf
farm-will/. infrastructures energy ICT oppor foster innov AEIT 2017 2017; 2017-January.
[51] Capgemini. Intelligent automation in energy and utilities. Capgemini Res Inst; p. 1–6. https://doi.org/10.23919/AEIT.2017.8240544.
2019. p. 1–40. https://doi.org/https://www.capgemini.com/wp-content/ [79] Paris EH, Washington SR, Chui M, Francisco S, London TA. Artificial intelligence
uploads/2019/05/Digital-Report-%E2%80%93-Automation-in-Utilities-1.pdf. the next digital frontier ? McKinsey Co; 2017. p. 1–80. Sree Ramas.
[52] Pong PWT, Annaswamy AM, Kroposki B, Zhang Y, Rajagopal R, Zussman G, et al. [80] Burlig F, Knittel C, Rapson D, Reguant M, Wolfram C. Machine learning from
Cyber-enabled grids: shaping future energy systems. Adv Appl Energy 2021;1: schools about energy efficiency. J Assoc Environ Resour Econ 2020;7:1181–217.
100003. https://doi.org/10.1016/j.adapen.2020.100003. https://doi.org/10.1086/710606.
[53] Donnal JS, Leeb SB. Noncontact power meter. IEEE Sensor J 2015;15:1161–9. [81] Owoyele O, Pal P. A novel machine learning-based optimization algorithm
https://doi.org/10.1109/JSEN.2014.2359619. (ActivO) for accelerating simulation-driven engine design. Appl Energy 2021;
[54] Tobin NP, McCormack M, O’Loughlin E, Tobin K. Remote current sensing on 285:116455. https://doi.org/10.1016/j.apenergy.2021.116455.
overhead power lines. In: IEE conf publ; 2001. p. 1. https://doi.org/10.1049/cp: [82] Zhou Y, Zheng S. Climate adaptive optimal design of an aerogel glazing system
20010727. with the integration of a heuristic teaching-learning-based algorithm in machine

33
T. Ahmad et al. Renewable and Sustainable Energy Reviews 160 (2022) 112128

learning-based optimization. Renew Energy 2020;153:375–91. https://doi.org/ [108] Lee J, Davari H, Singh J, Pandhare V. Industrial Artificial Intelligence for industry
10.1016/j.renene.2020.01.133. 4.0-based manufacturing systems. Manuf Lett 2018;18:20–3. https://doi.org/
[83] Al-Shahri OA, Ismail FB, Hannan MA, Lipu MSH, Al-Shetwi AQ, Begum RA, et al. 10.1016/j.mfglet.2018.09.002.
Solar photovoltaic energy optimization methods, challenges and issues: a [109] Wang H, Lei Z, Zhang X, Zhou B, Peng J. A review of deep learning for renewable
comprehensive review. J Clean Prod 2021;284:125465. https://doi.org/10.1016/ energy forecasting. Energy Convers Manag 2019;198. https://doi.org/10.1016/j.
j.jclepro.2020.125465. enconman.2019.111799.
[84] Kong X, Li C, Zheng F, Wang C. Improved deep belief network for short-term load [110] Ahmad T, Zhang H, Yan B. A review on renewable energy and electricity
forecasting considering demand-side management. IEEE Trans Power Syst 2020; requirement forecasting models for smart grid and buildings. Sustain Cities Soc
35:1531–8. https://doi.org/10.1109/TPWRS.2019.2943972. 2020;55. https://doi.org/10.1016/j.scs.2020.102052.
[85] Lopez KL, Gagne C, Gardner MA. Demand-side management using deep learning [111] Tavana M, Shaabani A, Santos-Arteaga FJ, Vanani IR. A review of uncertain
for smart charging of electric vehicles. IEEE Trans Smart Grid 2019;10:2683–91. decision-making methods in energy management using text mining and data
https://doi.org/10.1109/TSG.2018.2808247. analytics. Energies 2020;13:1–13. https://doi.org/10.3390/en13153947.
[86] Mahmud K, Khan B, Ravishankar J, Ahmadi A, Siano P. An internet of energy [112] Nisar A, Ruiz F, Palacios M. Organisational learning, strategic rigidity and
framework with distributed energy resources, prosumers and small-scale virtual technology adoption: implications for electric utilities and renewable energy
power plants: an overview. Renew Sustain Energy Rev 2020;127:109840. https:// firms. Renew Sustain Energy Rev 2013;22:438–45. https://doi.org/10.1016/j.
doi.org/10.1016/j.rser.2020.109840. rser.2013.01.039.
[87] IRENA. Innovation landscape for a renewable-powered futureSOLUTION V; 2019. [113] Gupta V, Bali R. Artificial intelligence in the energy world—getting the act
p. 1–6. /Publications/2019/Feb/Innovation-Landscape-for-a-Renewable- together. 2021. https://doi.org/10.1007/978-981-15-8297-4_63. 793–8.
Powered-Future https://doi.org/https://www.irena.org/publications/2019/Feb/ [114] Nikkei. AI to propel wind farm efficiency in Japan. AsiaNikkeiCom/Business/
Innovation-landscape-for-a-renewable-powered-future. Companies/AI-to-Propel-Wind-Farm-Efficiency-in-Japan 2017. https://doi.org/
[88] Vázquez-Canteli JR, Nagy Z. Reinforcement learning for demand response: a https://asia. nikkei.com/Business/AI-to-propel-wind-farmefficiency- in-Japan.
review of algorithms and modeling techniques. Appl Energy 2019;235:1072–89. [115] Company M. Performance management and optimization for the electricpower
https://doi.org/10.1016/j.apenergy.2018.11.002. industry. McKinsey Co; 2019. p. 1–24 [https://doi.org/www.mckinsey.com/
[89] Haq EU, Lyu X, Jia Y, Hua M, Ahmad F. Forecasting household electric appliances solutions/utilityx].
consumption and peak demand based on hybrid machine learning approach. [116] Mohanta BK, Jena D, Satapathy U, Patnaik S. Survey on IoT security: challenges
Energy Rep 2020;6:1099–105. https://doi.org/10.1016/j.egyr.2020.11.071. and solution using machine learning, artificial intelligence and blockchain
[90] Wen L, Zhou K, Li J, Wang S. Modified deep learning and reinforcement learning technology. Internet of Things 2020;11:100227. https://doi.org/10.1016/j.
for an incentive-based demand response model. Energy 2020;205:118019. iot.2020.100227.
https://doi.org/10.1016/j.energy.2020.118019. [117] Henzelmann Torsten, Hammermeister F. Artificial intelligence: a smart move for
[91] Ponnaganti P, Pillai JR, Bak-Jensen B. Opportunities and challenges of demand utilities. Rol Berger ThinkAct 2018.
response in active distribution networks, vol. 7. Wiley Interdiscip Rev Energy [118] Heffron R, Körner MF, Wagner J, Weibelzahl M, Fridgen G. Industrial demand-
Environ; 2018. p. 1–16. https://doi.org/10.1002/wene.271. side flexibility: a key element of a just energy transition and industrial
[92] Ghobakhloo M. Industry 4.0, digitization, and opportunities for sustainability. development. Appl Energy 2020;269:115026. https://doi.org/10.1016/j.
J Clean Prod 2020;252:119869. https://doi.org/10.1016/j.jclepro.2019.119869. apenergy.2020.115026.
[93] Hu M, Chen H, Shen L, Li G, Guo Y, Li H, et al. A machine learning bayesian [119] Verhees B, Raven R, Kern F, Smith A. The role of policy in shielding, nurturing
network for refrigerant charge faults of variable refrigerant flow air conditioning and enabling offshore wind in The Netherlands (1973-2013). Renew Sustain
system. Energy Build 2018;158:668–76. https://doi.org/10.1016/j. Energy Rev 2015;47:816–29. https://doi.org/10.1016/j.rser.2015.02.036.
enbuild.2017.10.012. [120] Magazzino C, Mele M, Schneider N. A machine learning approach on the
[94] Kück M, Freitag M. Forecasting of customer demands for production planning by relationship among solar and wind energy production, coal consumption, GDP,
local k-nearest neighbor models. Int J Prod Econ 2021;231:107837. https://doi. and CO2 emissions. Renew Energy 2021;167:99–115. https://doi.org/10.1016/j.
org/10.1016/j.ijpe.2020.107837. renene.2020.11.050.
[95] Yang L, He M, Zhang J, Vittal V. Support-vector-machine-enhanced markov [121] Gorjian S, Sharon H, Ebadi H, Kant K, Scavo FB, Tina GM. Recent technical
model for short-term wind power forecast. IEEE Trans Sustain Energy 2015;6: advancements, economics and environmental impacts of floating photovoltaic
791–9. https://doi.org/10.1109/TSTE.2015.2406814. solar energy conversion systems. J Clean Prod 2021;278:124285. https://doi.org/
[96] Kawase H, Mori Y, Hasegawa H, Sato KC. Dynamic router performance control 10.1016/j.jclepro.2020.124285.
utilizing support vector machines for energy consumption reduction. IEEE Trans [122] Elshawi R, Sakr S. Big data systems meet machine learning challenges: towards
Netw Serv Manag 2016;13:860–70. https://doi.org/10.1109/ big data science as a service. Big Data Res 2018;14:1–11.
TNSM.2016.2605640. [123] Onile AE, Machlev R, Petlenkov E, Levron Y, Belikov J. Uses of the digital twins
[97] Perera ATD, Kamalaruban P. Applications of reinforcement learning in energy concept for energy services, intelligent recommendation systems, and demand
systems. Renew Sustain Energy Rev 2021;137:110618. https://doi.org/10.1016/ side management: a review. Energy Rep 2021;7:997–1015. https://doi.org/
j.rser.2020.110618. 10.1016/j.egyr.2021.01.090.
[98] Zhang X, Li Y, Lu S, Hamann HF, Hodge BM, Lehman B. A solar time based analog [124] Zhou Y, Liu Y, Wang D, Liu X, Wang Y. A review on global solar radiation
ensemble method for regional solar power forecasting. IEEE Trans Sustain Energy prediction with machine learning models in a comprehensive perspective. Energy
2019;10:268–79. https://doi.org/10.1109/TSTE.2018.2832634. Convers Manag 2021;235:113960. https://doi.org/10.1016/j.
[99] Han Y, Xiu L, Wang Z, Chen Q, Tan S. Artificial neural networks controlled fast enconman.2021.113960.
valving in a power generation plant. IEEE Trans Neural Network 1997;8:373–89. [125] Dogan A, Birant D. Machine learning and data mining in manufacturing. Expert
https://doi.org/10.1109/72.557689. Syst Appl 2021;166:114060. https://doi.org/10.1016/j.eswa.2020.114060.
[100] Dai X, Liu J, Zhang X. A review of studies applying machine learning models to [126] Liu Y, Guo B, Zou X, Li Y, Shi S. Machine learning assisted materials design and
predict occupancy and window-opening behaviours in smart buildings. Energy discovery for rechargeable batteries. Energy Storage Mater 2020;31:434–50.
Build 2020;223:110159. https://doi.org/10.1016/j.enbuild.2020.110159. https://doi.org/10.1016/j.ensm.2020.06.033.
[101] Ray PK, Mohanty SR, Kishor N, Catalao JPS. Optimal feature and decision tree- [127] Oracle Oracle. Big data for the enterprise. Oracle White Pap 2012:16.
based classification of power quality disturbances in distributed generation [128] Zhang Y, Huang T, Bompard EF. Big data analytics in smart grids: a review.
systems. IEEE Trans Sustain Energy 2014;5:200–8. https://doi.org/10.1109/ Energy Informatics 2018;1:1–8. https://doi.org/10.1186/s42162-018-0007-5.
TSTE.2013.2278865. [129] Bhattarai BP, Paudyal S, Luo Y, Mohanpurkar M, Cheung K, Tonkoski R, et al. Big
[102] Ramos V, Giannini G, Calheiros-Cabral T, Rosa-Santos P, Taveira-Pinto F. Legal data analytics in smart grids: state-of-theart, challenges, opportunities, and future
framework of marine renewable energy: a review for the Atlantic region of directions. IET Smart Grid 2019;2:141–54. https://doi.org/10.1049/iet-
Europe. Renew Sustain Energy Rev 2021;137:110608. https://doi.org/10.1016/j. stg.2018.0261.
rser.2020.110608. [130] Gu B, Sheng VS. A robust regularization path algorithm for ν-support vector
[103] Liukkonen M, Hiltunen T. Adaptive monitoring of emissions in energy boilers classification. IEEE Transact Neural Networks Learn Syst 2017;28:1241–8.
using self-organizing maps: an application to a biomass-fired CFB (circulating https://doi.org/10.1109/TNNLS.2016.2527796.
fluidized bed). Energy 2014;73:443–52. https://doi.org/10.1016/j. [131] GTMResearch. Sas. The soft grid 2013-2020 : big data & utility analytics for smart
energy.2014.06.034. grid. 2013. p. 1–15.
[104] Saraswat SK, Digalwar AK. Evaluation of energy alternatives for sustainable [132] Zhou L, Pan S, Wang J, Vasilakos AV. Machine learning on big data: opportunities
development of energy sector in India: an integrated Shannon’s entropy fuzzy and challenges. Neurocomputing 2017;237:350–61. https://doi.org/10.1016/j.
multi-criteria decision approach. Renew Energy 2021;171:58–74. https://doi. neucom.2017.01.026.
org/10.1016/j.renene.2021.02.068. [133] Bayram IS, Saffouri F, Koc M. Generation, analysis, and applications of high
[105] Baharoon DA, Rahman HA, Fadhl SO. Publics’ knowledge, attitudes and resolution electricity load profiles in Qatar. J Clean Prod 2018;183:527–43.
behavioral toward the use of solar energy in Yemen power sector. Renew Sustain https://doi.org/10.1016/j.jclepro.2018.02.084.
Energy Rev 2016;60:498–515. https://doi.org/10.1016/j.rser.2015.12.110. [134] Patnam BSK, Pindoriya NM. Demand response in consumer-Centric electricity
[106] Kotsiopoulos T, Sarigiannidis P, Ioannidis D, Tzovaras D. Machine learning and market: mathematical models and optimization problems. Elec Power Syst Res
deep learning in smart manufacturing: the smart grid paradigm. Comput Sci Rev 2020;193:106923. https://doi.org/10.1016/j.epsr.2020.106923.
2021;40:100341. https://doi.org/10.1016/j.cosrev.2020.100341. [135] Fathi S, Srinivasan R, Fenner A, Fathi S. Machine learning applications in urban
[107] Qu YJ, Ming XG, Liu ZW, Zhang XY, Hou ZT. Smart manufacturing systems: state building energy performance forecasting: a systematic review. Renew Sustain
of the art and future trends. Int J Adv Manuf Technol 2019;103:3751–68. https:// Energy Rev 2020;133:110287. https://doi.org/10.1016/j.rser.2020.110287.
doi.org/10.1007/s00170-019-03754-7. [136] Seyedzadeh S, Pour Rahimian F, Oliver S, Rodriguez S, Glesk I. Machine learning
modelling for predicting non-domestic buildings energy performance: a model to

34
T. Ahmad et al. Renewable and Sustainable Energy Reviews 160 (2022) 112128

support deep energy retrofit decision-making. Appl Energy 2020;279:115908. [164] Nižetić S, Šolić P, López-de-Ipiña González-de-Artaza D, Patrono L. Internet of
https://doi.org/10.1016/j.apenergy.2020.115908. Things (IoT): opportunities, issues and challenges towards a smart and sustainable
[137] Said Z, Hachicha AA, Aberoumand S, Yousef BAA, Sayed ET, Bellos E. Recent future. J Clean Prod 2020;274:122877. https://doi.org/10.1016/j.
advances on nanofluids for low to medium temperature solar collectors: energy, jclepro.2020.122877.
exergy, economic analysis and environmental impact. Prog Energy Combust Sci [165] Feng C, Wang Y, Chen Q, Ding Y, Strbac G, Kang C. Smart grid encounters edge
2021;84:100898. https://doi.org/10.1016/j.pecs.2020.100898. computing: opportunities and applications. Adv Appl Energy 2021;1:100006.
[138] Sun W, Zhang C. Analysis and forecasting of the carbon price using https://doi.org/10.1016/j.adapen.2020.100006.
multi—resolution singular value decomposition and extreme learning machine [166] Amin AZ. Global energy transformation. Int Renew Energy Agency 2018:1–76.
optimized by adaptive whale optimization algorithm. Appl Energy 2018: https://doi.org/10.1057/9780230244092.
1354–71. https://doi.org/10.1016/j.apenergy.2018.09.118. [167] You NZF. New York State’s 100% renewable electricity transition planning under
[139] Sensfuß F, Ragwitz M, Genoese M. The merit-order effect: a detailed analysis of uncertainty using a data-driven multistage adaptive robust optimization approach
the price effect of renewable electricity generation on spot market prices in with machine-learning. Adv Appl Energy n.d.;Available:100019. https://doi.org/
Germany. Energy Pol 2008;36:3086–94. https://doi.org/10.1016/j. https://doi.org/10.1016/j.adapen.2021.100019.
enpol.2008.03.035. [168] Vasudevan RK, Choudhary K, Mehta A, Smith R, Kusne G, Tavazza F, et al.
[140] Kell AJM, McGough AS, Forshaw M. The impact of online machine-learning Materials science in the artificial intelligence age: high-throughput library
methods on long-term investment decisions and generator utilization in generation, machine learning, and a pathway from correlations to the
electricity markets. Sustain Comput Informatics Syst 2020;30:100532. underpinning physics. MRS Commun 2019;9:821–38. https://doi.org/10.1557/
[141] Freitas Gomes IS, Perez Y, Suomalainen E. Coupling small batteries and PV mrc.2019.95.
generation: a review. Renew Sustain Energy Rev 2020;126:109835. https://doi. [169] Huang JS, Liew JX, Ademiloye AS, Liew KM. Artificial intelligence in materials
org/10.1016/j.rser.2020.109835. modeling and design. Arch Comput Methods Eng 2020. https://doi.org/10.1007/
[142] Strielkowski W. Consumers, prosumers, and the smart grids. Soc Impacts Smart s11831-020-09506-1.
Grids 2020:191–238. https://doi.org/10.1016/b978-0-12-817770-9.00006-7. [170] Zhou M, Gallegos A, Liu K, Dai S, Wu J. Insights from machine learning of carbon
[143] GuidoPerboli; Ehsan Arabnezhad. A Machine Learning-based DSS for mid and electrodes for electric double layer capacitors. Carbon N Y 2020;157:147–52.
long-term company crisis prediction. Expert Syst Appl n.d.;174:114758. https://doi.org/10.1016/j.carbon.2019.08.090.
[144] Strielkowski W. Traditional power markets and an evolution to smart grids. Soc [171] Orupattur NV, Mushrif SH, Prasad V. Catalytic materials and chemistry
Impacts Smart Grids 2020:9–54. https://doi.org/10.1016/b978-0-12-817770- development using a synergistic combination of machine learning and ab initio
9.00002-x. methods. Comput Mater Sci 2020;174:109474. https://doi.org/10.1016/j.
[145] Babuta A, Gupta B, Kumar A, Ganguli S. Power and energy measurement devices: commatsci.2019.109474.
a review, comparison, discussion, and the future of research. Meas J Int Meas [172] Yang W, Fidelis TT, Sun WH. Machine learning in catalysis, from proposal to
Confed 2021;172:108961. https://doi.org/10.1016/j. practicing. ACS Omega 2020;5:83–8. https://doi.org/10.1021/
measurement.2020.108961. acsomega.9b03673.
[146] du Plessis AA, Strauss JM, Rix AJ. Short-term solar power forecasting: [173] Zhang L, He M, Shao S. Machine learning for halide perovskite materials. Nano
investigating the ability of deep learning models to capture low-level utility-scale Energy 2020;78:105380. https://doi.org/10.1016/j.nanoen.2020.105380.
Photovoltaic system behaviour. Appl Energy 2021;285:116395. https://doi.org/ [174] Odabaşı Ç, Yıldırım R. Machine learning analysis on stability of perovskite solar
10.1016/j.apenergy.2020.116395. cells. Sol Energy Mater Sol Cells 2020;205:110284. https://doi.org/10.1016/j.
[147] Park C, Heo WG. Review of the changing electricity industry value chain in the solmat.2019.110284.
ICT convergence era. J Clean Prod 2020;258:120743. https://doi.org/10.1016/j. [175] Shi Z, Yang W, Deng X, Cai C, Yan Y, Liang H, et al. Machine-learning-assisted
jclepro.2020.120743. high-throughput computational screening of high performance metal-organic
[148] Ferrando M, Causone F, Hong T, Chen Y. Urban building energy modeling frameworks. Mol Syst Des Eng 2020;5:725–42. https://doi.org/10.1039/
(UBEM) tools: a state-of-the-art review of bottom-up physics-based approaches. d0me00005a.
Sustain Cities Soc 2020;62:102408. https://doi.org/10.1016/j.scs.2020.102408. [176] He Y, Cubuk ED, Allendorf MD, Reed EJ. Metallic metal-organic frameworks
[149] Robinson D, Haldi F, Kämpf J, Leroux P, Perez D, Rasheed A, et al. Citysim: predicted by the combination of machine learning methods and ab initio
comprehensive micro-simulation of resource flows for sustainable urban calculations. J Phys Chem Lett 2018;9:4562–9. https://doi.org/10.1021/acs.
planning. IBPSA 2009 - Int Build Perform Simul Assoc 2009 2009. 1083–90. jpclett.8b01707.
[150] Nouvel R, Brassel K-H, Bruse M, Duminil E, Coors V, Eicker U, et al. SIMSTADT, a [177] Toyao T, Suzuki K, Kikuchi S, Takakusagi S, Shimizu KI, Takigawa I. Toward
new workflow-driven urban energy simulation platform for CityGML city models. effective utilization of methane: machine learning prediction of adsorption
In: CISBAT int conf; 2015. p. 889–94. energies on metal alloys. J Phys Chem C 2018;122:8315–26. https://doi.org/
[151] Agugiaro G, Benner J, Cipriano P, Nouvel R. The Energy Application Domain 10.1021/acs.jpcc.7b12670.
Extension for CityGML: enhancing interoperability for urban energy simulations. [178] Vazquez S, Lukic S, Galvan E, Franquelo LG, Carrasco JM, Leon JI. Recent
Open Geospatial Data, Softw Stand 2018;3:1–2. https://doi.org/10.1186/s40965- advances on energy storage systems. In: IECON proc (industrial electron conf;
018-0042-y. 2011. p. 4636–40. https://doi.org/10.1109/IECON.2011.6120075.
[152] CityGML. Open geospatial consort. 2021 [https://doi.org/https://www.citygml. [179] Intelligence A. Spotlight: artificial intelligence. Using artificial intelligence to
org/]. advance the state of multiple industries. September 2019. 2019. https://doi.org/
[153] MIT. Sustainable design lab. Umi - urban modeling interface. 2013 [https://doi. https://www.energy.gov/sites/prod/files/2019/10/f67/2019-09-30_Spotlight-
org/http://web.mit.edu/sustainabledesignlab/projects/umi/index.html]. Artificial_Intelligence_0.pdf.
[154] Hong T, Chen Y, Piette Xl MA. City 2017 [https://doi.org/https://citybes.lbl.gov/ [180] Shaukat N, Ali SM, Mehmood CA, Khan B, Jawad M, Farid U, et al. A survey on
]. consumers empowerment, communication technologies, and renewable
[155] Fonseca JA, Nguyen TA, Schlueter A, Marechal F. City Energy Analyst (CEA): generation penetration within Smart Grid. Renew Sustain Energy Rev 2018;81:
integrated framework for analysis and optimization of building energy systems in 1453–75. https://doi.org/10.1016/j.rser.2017.05.208.
neighborhoods and city districts. Energy Build 2016;113:202–26. https://doi.org/ [181] Dehghanpour K, Hashem Nehrir M, Sheppard JW, Kelly NC. Agent-based
10.1016/j.enbuild.2015.11.055. modeling of retail electrical energy markets with demand response. IEEE Trans
[156] Remmen P, Lauster M, Mans M, Fuchs M, Osterhage T, Müller D. TEASER: an Smart Grid 2018;9:3465–75. https://doi.org/10.1109/TSG.2016.2631453.
open tool for urban energy modelling of building stocks. J Build Perform Simul [182] Zhao W, Zhang H, Zheng J, Dai Y, Huang L, Shang W, et al. A point prediction
2018;11:84–98. https://doi.org/10.1080/19401493.2017.1283539. method based automatic machine learning for day-ahead power output of multi-
[157] Menghi R, Papetti A, Germani M, Marconi M. Energy efficiency of manufacturing region photovoltaic plants. Energy 2021;223:120026. https://doi.org/10.1016/j.
systems: a review of energy assessment methods and tools. J Clean Prod 2019; energy.2021.120026.
240:118276. https://doi.org/10.1016/j.jclepro.2019.118276. [183] Khodayar M, Mohammadi S, Khodayar ME, Wang J, Liu G. Convolutional graph
[158] Lund H. Tool: the energyplan energy system analysis model. Renew Energy Syst A autoencoder: a generative deep neural network for probabilistic spatio-temporal
Smart Energy Syst Approach to Choice Model 100% Renew Solut Second 2014: solar irradiance forecasting. IEEE Trans Sustain Energy 2020;11:571–83. https://
53–78. https://doi.org/10.1016/B978-0-12-410423-5.00004-3. doi.org/10.1109/TSTE.2019.2897688.
[159] Ferrari S, Zagarella F, Caputo P, Bonomolo M. Assessment of tools for urban [184] Guevara E, Babonneau F, Homem-de-Mello T, Moret S. A machine learning and
energy planning. Energy 2019;176:544–51. https://doi.org/10.1016/j. distributionally robust optimization framework for strategic energy planning
energy.2019.04.054. under uncertainty. Appl Energy 2020;271:115005. https://doi.org/10.1016/j.
[160] Yao L, Yang B, Cui H, Zhuang J, Ye J, Xue J. Challenges and progresses of energy apenergy.2020.115005.
storage technology and its application in power systems. J Mod Power Syst Clean [185] Hu L, Che XL, Zheng SQ. Online system for grid resource monitoring and machine
Energy 2016;4:519–28. https://doi.org/10.1007/s40565-016-0248-x. learning-based prediction. IEEE Trans Parallel Distr Syst 2012;23:134–45.
[161] Lu Y, Zheng X6G. A survey on technologies, scenarios, challenges, and the related https://doi.org/10.1109/TPDS.2011.108.
issues. J Ind Inf Integr 2020;19:100158. https://doi.org/10.1016/j. [186] Wen S, Zhang C, Lan H, Xu Y, Tang Y, Huang Y. A hybrid ensemble model for
jii.2020.100158. interval prediction of solar power output in ship onboard power systems. IEEE
[162] Oconnell N, Pinson P, Madsen H, Omalley M. Benefits and challenges of electrical Trans Sustain Energy 2021;12:14–24. https://doi.org/10.1109/
demand response: a critical review. Renew Sustain Energy Rev 2014;39:686–99. TSTE.2019.2963270.
https://doi.org/10.1016/j.rser.2014.07.098. [187] Paris EH, Washington SR, Chui M, Francisco S, London TA. Artificial intelligence
[163] Artificial Intelligence and Big Data Innovation Landscape Brief. Irea 2019. IRENA; the next digital frontier. McKinsey Co. Sree Ramas 2017:1–80.
2019. p. 1–24. https://doi.org/https://www.irena.org/-/media/Files/IRENA/ [188] IRENA - International Renewable Energy Agency. REmap 2030 - a renewable
Agency/Publication/2019/Sep/IRENA_AI_Big_Data_2019.pdf? energy roadmap. IRENA; 2014. p. 1–56. https://www.irena.org/-/media/Files/
la=en&hash=9A003F48B639B810237FEEAF61D47C74F8D8F07F. IRENA/Agency/Publication/2014/IRENA_REmap_summary_findings_2014.pdf.

35