Energy Conversion and Management: X 26 (2025) 100922

Contents lists available at ScienceDirect

Energy Conversion and Management: X

journal homepage: www.sciencedirect.com/journal/energy-conversion-and-management-x

A state-of-the-art comparative review of load forecasting methods:

Characteristics, perspectives, and applications
Mahmudul Hasan a , Zannatul Mifta b, Sumaiya Janefar Papiya a , Paromita Roy a, Pronay Dey c ,
Nafisa Atia Salsabil a , Nahid-Ur-Rahman Chowdhury a,* , Omar Farrok a
a
Department of Electrical and Electronic Engineering, Ahsanullah University of Science and Technology (AUST), Dhaka 1208, Bangladesh
b
Department of Electrical and Electronic Engineering, Rajshahi University of Engineering and Technology (RUET), Rajshahi 6204, Bangladesh
c
Department of Electrical and Electronic Engineering, Bangladesh University of Engineering and Technology (BUET), Dhaka 1000, Bangladesh

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

Keywords: The rapid growth in electricity demand, driven by its expanding applications across diverse sectors, has
Load forecasting emphasized the criticality of maintaining a balanced and reliable power supply. Accurate load forecasting has
Artificial intelligence in power systems become a cornerstone of modern power system management, enabling the efficient planning, operation, and
Renewable energy integration
design of electrical grids. With the increasing penetration of renewable energy sources and the rise of smart grid
Smart grid stability
technologies, the need for precise forecasting methodologies has intensified to ensure enhanced grid stability,
Machine learning for energy demand
prediction efficiency, and seamless renewable integration. This article systematically reviews contemporary state-of-the-art
Operation and planning forecasting techniques, critically analyzing their performance, applications, and outcomes. Emphasis is placed on
methodologies for predicting renewable energy availability, electricity pricing, and load demand, with an in-
depth evaluation of their modeling frameworks and predictive accuracies. The review highlights significant
advancements in artificial intelligence-based approaches, particularly machine learning and neural network
models, which consistently outperform traditional forecasting methods in terms of precision and robustness. For
enhanced clarity, key insights and comparative analyses are summarized in comprehensive tables, facilitating
efficient reference. This review aims to provide researchers with a thorough understanding of advanced fore-
casting models, their capabilities, and limitations, thereby guiding future research endeavors in the domain of
load forecasting.

1. Introduction been trying to create more efficient and environmentally friendly

methods [4]. Programming and optimal productivity from power sys-
Major countries are transforming their energy systems into clean, tems depend on the rate at which electrical load is consumed. However,
low-carbon, intelligent ones; in response to environmental pollution, electrical load consumption is dependent on a number of variables
climate change, and the depletion of fossil fuels. In this context, the including weather, wind, humidity, and other atmospheric conditions as
structural framework of the power grids’ generation and consumption well as days, months, and holidays. Precise load forecasting may lessen
sectors is experiencing considerable transformation. Electricity is a the impact of renewable energy sources being integrated into the sys-
fundamental energy source for both life and production. The steady and tem, ease unit maintenance arrangements for power plants, and incen-
consistent expansion of the economy of any nation, particularly the tivize power broker businesses to create competitive electricity pricing
balanced growth of electricity supply and power demand, is correlated plans [5].
with the favorable development of the power sector. Among these, the As demand response is gradually implemented on the load side and
precise forecast of power usage is crucial [1,2]. Energy is the foundation solar and wind power penetration rates rise on the generation side, the
of the contemporary economy, and regulations relating to climate power system is becoming extremely unpredictable on both sides. This
change, sustainability, and supply security are driving up the need for increases the difficulties in maintaining the system’s more affordable
electrification in many sectors, including transportation [3]. In order to and safe operation [6]. The ability to anticipate demand is essential for
meet their electrical needs, consumers, investors, and legislators have many power system operations, including scheduling generator

* Corresponding author.
E-mail addresses: [email protected] (N.-U.-R. Chowdhury), [email protected] (O. Farrok).

https://doi.org/10.1016/j.ecmx.2025.100922
Received 16 September 2024; Received in revised form 25 January 2025; Accepted 10 February 2025
Available online 12 February 2025
2590-1745/© 2025 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-
nc-nd/4.0/).
M. Hasan et al. Energy Conversion and Management: X 26 (2025) 100922

maintenance, integrating renewable energy sources, and even investing electricity generation and distribution, accurate electrical load modeling
in power plants and power networks. Based on load and power generator can lower the cost of operating an energy system.
forecasts, regulators keep an eye on market activity in the electrical Due to its critical role in the creation and utilization of systems, load
sector. Their action plans are determined by customers and power forecasting has received particular attention since the beginning of the
brokers. Here, Fig. 1 illustrates load factors associated with the electric electricity business [5]. Preceding research documented in the literature
load that enormously impact the load pattern and are thus very impor- indicates that the programming and optimal efficiency of power systems
tant in load forecasting. are contingent upon the magnitude of electrical load consumption. The
The uncertainty of production presents new difficulties for fore- substantial amount of renewable energy that has been added to the grid
casters and raises the bar for prediction accuracy. This study identified in recent years has drastically altered the structure of the power supply
that accurate load forecasting may reduce investment costs and improve of the power system. The overall power system is more uncertain due to
scheduling for the construction of power plants as well as distribution the unpredictable output of renewable energy sources.
and transmission networks [7]. Because of the expansion in the kinds of It influences the reliability of the load forecast as well as the study
electrical equipment that users utilize, the degree of complexity of and formulation of the power-side generation strategy, which influences
consumers’ energy systems has increased significantly in the current the load-side response to demand [9]. The next stage of the power sys-
modern power system. As a result, the load curve generated by the user’s tem Grid will be one of high perception and intelligence stage, and this
electrical load exhibits large variations. This work’s primary goal is to will introduce new demands for load forecasting of accuracy and real-
evaluate the most recent advancements of data-driven electrical load time. As a result, it plays an essential role for the smart power sys-
forecasting using various models that are appropriate and made to tem’s operating on a schedule and the full use of new energy [10].
maintain a safe, sufficient, and effective supply of electricity while also As an alternative to fuel-powered cars, electric vehicles (EVs) have
providing power system stability. become a worldwide trend. With EVs’ rapid development, the conse-
The global energy sector is currently facing two significant chal- quences of the charging needs on a power grid’s load characteristics
lenges: (a) the depletion of fossil fuel reserves, including coal and oil, have risen. Peak loads, daily load curves, and annual loads on the power
and (b) the adverse environmental impacts caused by the release of grid all change significantly as the number of EVs rises. The overall
greenhouse gases, which contribute to climate change. In this context, charging load has progressively increased as EVs have grown [12]. This
there is an urgent imperative to transition electricity generation from will have a substantial impact on a power grid’s reliability and load
fossil fuel-based sources to renewable energy systems, while simulta- characteristics, which will be discussed later in the study. So, the rapid
neously advancing the technological tools and methodologies used in expansion of EVs must be taken into account while building future
energy production. Achieving environmental, social, and economic ob- power systems [13].
jectives during this transitional phase necessitates accurate and reliable Many new unpredictable factors that have an impact on load fore-
forecasting of energy demand growth [8]. casting have emerged as the power industry develops quickly. It is
Load forecasting is a process that uses prior or historical data to challenging to predict the medium-long term load exactly because of the
forecast future load needs in electrical systems. In the quickly evolving numerous unpredictable factors that influence the electrical demand.
digital era, electricity forecasting is crucial for enabling the safe tran- The correlation between the upstream and downstream industries has an
sition of power system. By enabling more efficient management of impact on the industrial load in addition to these environmental and

Fig. 1. Electric load and corresponding load factors [11].

2
M. Hasan et al. Energy Conversion and Management: X 26 (2025) 100922

social elements, which cannot be disregarded. As a result, taking a va- Table 1

riety of parameters into consideration and increasing the accuracy of Contributions and research gaps in load forecasting domain found in the recent
load forecasting have received considerable attention [14,15]. In the literature.
development of the microgrid, nano grid, and main grid infrastructure, References Contributions Research gaps
accurate load forecasting is indispensable. The integration and optimi- [5] This study presents a thorough This study did not encompass
zation of renewable energy resources can be significantly enhanced review of mid-term load short term and long-term load
through precise load prediction [1617]. The significance of electrical forecasting. After studying the forecasting techniques or conduct
load forecasting (ELF) is growing, especially in light of the negative fundamental loads and effective a comprehensive comparison of
elements, a number of all
impacts of climate change on electricity usage and the swift advance-
forecasting technique forecasting methods.
ment of smart grid technology in recent years. The increasing use of classifications are given. The
renewable energy systems RESs in power networks and the short-term techniques’ primary benefits and
availability nature of these sources have increased the demand for limitations are examined. The
higher resolution and reliability in electrical load forecasting while also drawbacks of the neuro-fuzzy-
based methods are examined,
posing new issues [18]. The lack of adequate local energy storage has and further research is done on
heightened the significance of ELF, making it crucial at this stage for them.
optimizing energy management [19]. [20] To anticipate energy Not every LF approaches had
A considerable number of research papers and review articles have consumption, the authors been examined or compared in
attempted to conduct a this work. Only some machine
been published on load forecasting, with most reviews focusing on a
methodical and critical analysis learning (ML) based model
single forecasting aspect. Although a few studies address two or three of around seventy-seven studied.
forecasting elements, they typically examine each one independently. pertinent prior publications
This present work acted as a bridge to connect the research gap in LF. published in academic journals
Table 1 presents a summary of key contributions from previous works, during a nine-year period
(2010–2020). In particular, the
outlining the identified research gaps and the issues that have been
following themes received
addressed in the subsequent section of this current study. Here, Fig. 2 attention: (i) the forecasting
represents the basic outline of this review paper. algorithms and their fit for this
The key contributions of this study are outlined as follows: field; (ii) the theories and factors
influencing electricity
consumption and the beginnings
▪ Comprehensive classification of load forecasting techniques: of research; (iii) the pertinent
This study provides an in-depth overview of various types of error and accuracy metrics used
load forecasting methods, systematically categorizing them in electricity load forecasting;
based on their underlying methodologies. Their distinctive and (iv) the forecasting period.

features, practical applications, and predictive accuracies are

[21] In order to anticipate medium- This paper exclusively examined
critically analyzed to offer a holistic understanding of their and long-term load forecasting, short-term and mid-term load
capabilities. this paper offers an effective load forecasting methods and did not
▪ Identification of challenges and proposed solutions: The study prediction method predicted for include an analysis of long-term
various local feeds. The load
identifies key challenges in the domain of load forecasting and
suggested techniques provide forecasting techniques.
presents actionable solutions to address these limitations, more consistent and improved
thereby contributing to the advancement of forecasting prac- load forecasting performance.
tices in real-world applications. The result suggests that this
▪ Evaluation of modern forecasting techniques: Contemporary approach may be more
effectively implemented for
forecasting approaches incorporating advanced sensor tech-
requirements related to future
nologies and machine learning algorithms are thoroughly expansion.
examined. A comparative analysis of these methods is con- [22] After analyzing the results of This study compared only deep
ducted, highlighting their strengths, limitations, and potential using the feature selection learning methods for load
strategy with the operational forecasting. Mid-term and long-
for further improvement.
forecasting methods on two term methods did not discuss.
different datasets, a novel deep
These contributions serve to bridge gaps in the literature while of- learning (DL)-based ensemble
fering valuable insights for researchers and practitioners working to methodology was proposed. In
enhance the accuracy, efficiency, and applicability of load forecasting addition to providing research
findings on hour-ahead ELF, a
techniques.
subset of short-term load
The rest of this article is organized as follows. After the introduction forecasting (STLF), this article
in Section 1, the challenges and an overview of load forecasting methods reviews the most current, state-
are comprehensively described in Section 2. Then, Section 3 discusses of-the-art articles in this area.
the classification of load forecasting techniques. Section 4 describes the [23] This article presents a three-step This paper solely discussed the
assessment of the recent potential of data-driven urban
categorization of various models used in load forecasting. Measures for literature on data-driven power energy
performance assessment are illustrated in Section 5. After that a sum- load estimates. Aspects such as modeling.
mary table of very short-term, short-term, medium-term, and long-term performance, temporal
load forecasting techniques are organized in Section 6. Section 7 dis- resolution, benefits, and
downsides of load forecasting
cusses sustainable development goal-based points which are related to
systems have all been studied.
this article. Section 8 presents the discussion and future research di- The review discusses the
rections. Finally, the pertinent concluding remarks are outlined in Sec- difficulties that load forecasting
tion 9. is now facing and explains the
causes of decreased accuracy
and performance deterioration.
2. Electrical load forecasting: factors and techniques

One essential aspect of the smart grid is the prediction of electrical

3
M. Hasan et al. Energy Conversion and Management: X 26 (2025) 100922

Fig. 2. Topics covered under general procedure and framework of this study.

load. The accuracy and efficacy of load forecasting models are influ-
enced by forecasting factors, which are important variables and pa-
rameters. When forecasting energy demand, these elements need to be
carefully taken into account in order to make sure that the predictions
nearly match actual conditions.
Various criteria can classify different techniques used for electrical
load prediction: these include single-point and probabilistic forecasting,
depending on how results are interpreted; mathematical equations and
artificial intelligence methods, depending on the transparency of the
prediction model; and ultra-short-term, short-term, and mid-long-term
forecasting, based on the time scale of the prediction. All these
methods will be further presented in the next segments accordingly.

2.1. Challenges of load forecasting and proposed solutions

1. Forecasting relies on anticipated factors like weather conditions.

Regrettably, weather can be unpredictable, leading to disparities
between forecasted and actual conditions. Moreover, diverse regions
may encounter varying weather patterns, significantly influencing
electricity demand. To address the dynamic nature of factors such as
weather, economic circumstances, and consumer behavior,
employing time-varying model approaches [24] becomes essential,
allowing for the effective handling of non-stationary data. As re-
flected in Fig. 3, a number of criteria must be considered in order to
provide exact data for the development of an effective load fore-
casting model. These criteria may be divided into the time factor, Fig. 3. The factors affecting the LF models [25].
atmospheric conditions, economics, population, and facilities.
2. Consumer usage patterns differ based on the type of meter used, with During the summer, exceptionally high temperatures lead to
notable variations between smart and traditional meters and across increased demand for cooling in residential areas [27]. In response to
different tariff structures. The utility must acknowledge these dif- the growing impact of temperature on power load forecasting, M.
ferences and construct distinct forecast models for each metering Wang et al. [28] presented a short-term load forecasting model that
system. Aggregating these individual forecasts will yield a more ac- incorporates comprehensive meteorological effects and utilizes
curate final prediction. Evidence suggests that load forecasting based ElmanNN. In order to improve the precision of load forecasting, this
on smart meter data yields enhanced accuracy compared to forecasts work incorporates a temperature correction model to consider the
based on traditional meters [26]. collective influence of hourly air temperature, utilizing a correction
3. Power systems exhibit differences in size, complexity, and genera- coefficient for temperature changes.
tion mix, necessitating STLF models that can effectively handle these 5. Although power consumption data can be easily gathered, their ac-
variations. One strategy for achieving scalability and adaptability is curacy and validity may be questioned. Current electric load fore-
through transfer learning [26]. This entails utilizing a pre-trained casting techniques are dependent on both data integrity and quality,
model from one power system and adjusting it to suit the charac- which are compromised by malicious cyberattacks and noisy real-
teristics of another, allowing for more efficient and effective world environments, respectively [29]. A. Moradzadeh et al. [20]
forecasting. suggested a cybersecurity-oriented deep learning model for STLF. In
4. Temperature exerts the most significant influence on electricity addition to processing massive amounts of data and performing
consumption. Typically, the surge in energy demand within the acceptable performance against cyberattacks, the developed AE-
residential sector is attributed more to cooling needs than heating.

4
M. Hasan et al. Energy Conversion and Management: X 26 (2025) 100922

CLSTM model was also able to eliminate noise from raw information driving electricity consumption [33]. Both developed and industri-
of power consumption and model electric load data in time series. alized countries, as well as developing ones, have had difficulty with
6. When compared to traditional power systems, microgrid load load forecasting. While these constraints are somewhat less severe
modeling and forecasting can be more complicated. This is due to the for developed nations, the main issues facing developing nations are
fact that microgrid load time series are typically more erratic than the absence of essential data, appropriate load forecasting models,
power system load time series [30]. A microgrid’s load is more and necessary institutions [34].
variable because of its smaller size as opposed to a conventional
power system. Stated differently, load time series in small-scale
2.2. Overview of load forecasting methods
systems become less smooth because of low inertia. In conse-
quence, conventional clustering and load forecasting techniques
Fig. 4 provides a comprehensive depiction of the load forecasting
cannot be directly applied to microgrids. This necessitates the crea-
process. A set of data samples that serve as the model’s input features are
tion of novel methods to address the increased uncertainty in load
the first step in any load forecasting algorithm. Load datasets typically
behavior [31]. Machine learning techniques that have been shown to
contain past energy usage data, meteorological data, and time-related
be highly effective for microgrid load forecasting include Artificial
variables like different weekday and holiday hours that can affect the
Neural Networks (ANN), Support Machine Vectors (SVM), Random
consumption of electricity. Models using meteorological explanatory
Forest, and k-Nearest Neighbors (kNN). In addition, compared to
variables often rely on weather data. Ideally, predicted load should use
combining other single predictive techniques for load forecasting,
ex-ante forecasts, which rely on predicted weather variables from mul-
SVM or ANN-based double predictive models have demonstrated
tiple stations to reflect real-world conditions. Nevertheless, a lot of
better outcomes [32].
research employs ex-post forecasts with real weather data, which results
7. Electricity demand is notably influenced by the economic status of a
in forecast accuracy that is probably exaggerated [36]. The subsequent
country. In developed nations, the demand growth rate averages
step involves preparing the load data through data preprocessing,
around 0.7 % annually, while in developing countries, it has surged
employing various processing techniques. Following data preprocessing,
to approximately 3 % per year. Furthermore, the primary de-
the load data is analyzed, and the performance of the chosen load
terminants of demand differ between developed and developing
forecasting technique is assessed. Evaluation is conducted using error
nations. In developed countries, factors such as digitalization and
criteria, and the LF technique with the lowest error potential is chosen
electrification predominantly drive demand growth. Conversely, in
for load estimation.
developing countries, income levels play a more significant role in
Traditional load forecasting models commonly employ econometric

Fig. 4. Overview of load forecasting method [35].

5
M. Hasan et al. Energy Conversion and Management: X 26 (2025) 100922

models and time series models. Time series methodologies are utilized to In summary, the survey of literature suggests that an accurate and
analyze the evolution of the load, revealing its dynamic characteristics, precise forecasting techniques can enhance the following operation and
and enabling the deduction of future trends using mathematical tools planning aspects.
[23]. In the time series model, all the historical data are analyzed to
select a forecasting model that is the best fit for the given data as ▪ Supply-demand matching
illustrated in Fig. 4. ▪ Increasing the reliability of the electric grid’s attributes by
Fig. 5 depicts the evolution of the MTLF models for the forecasting of making it simple for administrators to intend and requiring
load over the next three months, six months, nine months, one year, and suppliers to come to intended conclusions
six years. Forecasting is the projection of future circumstances and ▪ Optimizing power load, which is needed during peak times, by
events based on historical data. Making decisions requires careful ensuring that the manufacturer’s energy supply is reduced
forecasting [37]. Despite common misconceptions, forecasting is ▪ Reaching both financial and environmental goals is facilitated
different from prediction in that it primarily draws on historical data by the harmonious integration of sustainable resources
sources. [38]. In the comprehensive method of load forecasting, as seen ▪ Maintaining a systematic framework can reduce total system
in Fig. 6, raw data is pre-processed before being given to prediction costs and save on operating and preservation charges.
models, whose performance is assessed using MAPE, MAE, and RMSE
analysis. 3. Categorization of load prediction methods
Basically, there are three different types of load forecasting: short,
medium, and long-term. An hour to no less than a week beforehand is The following categories for load forecasting are determined by time
covered by the short-term forecasting. A week to a year is the time frame zone:
for medium-term forecasting, and a year and beyond is the time frame
for long-term forecasting [40]. The different load forecasting types are • Very short-term load forecasting (VSTLF)
also classified according to their intended purposes. To maintain system • Short-term load forecasting (STLF)
stability and save operating costs for load balancing and system opera- • Medium-term load forecasting (MTLF)
tion, short-term forecasts are employed. For effective operational • Long-term load forecasting (LTLF)
scheduling and generation optimization, medium-term forecasts are
employed. For investment planning for the infrastructure growth of the The time horizon for VSTLF, STLF, MTLF, and LTLF and the time
power system, long-term forecasts are used. Each load forecasting ac- periods required for various planning, maintenance, and operational
tivity is distinct [41]. activities are shown in Fig. 9.
Policies play a major role in the medium and long-term load pre- As can be seen in Table 2, the time range for short-term load fore-
diction in particular. Fig. 7 depicts the impact of three different policy casting is 24 h to 1 week. The unit commitment procedure requires this
perspectives—economic, industrial, and environmental protection—on kind of forecasting. Mid-term load forecasting, which is important in
the amount of electricity consumed. There are eight subcategories establishing a plant’s reserve capacity, considers periods of time from
within each policy factor category. This comprises the per capita one week to one year. The time horizon for long-term load forecasting is
disposable income of urban residents, the population of permanent often more than one year. Forecasting of this kind is carried out in order
residents, and the regional GDP. Also includes the proportion of elec- to expand a plant’s capacity in the future.
tricity consumed by secondary and tertiary industry, the proportion of Load forecasting methods depend on data size and time scale.
electricity consumed by residents, energy consumption per 10,000 yuan Mathematical methods work well for small data, while AI methods are
GDP, and volume of exports as referred to Fig. 8 [29]. suited for large datasets. Regression analysis is typically used for mid- to

Fig. 5. Electric load forecasting system schematic diagram [21].

6
M. Hasan et al. Energy Conversion and Management: X 26 (2025) 100922

Fig. 6. General flow of load forecasting [39].

Fig. 7. The influence of policy factors on load [42].

long-term forecasts with periodicity, and other methods are preferred 3.2. Survey on short-term load forecasting method
for short-term forecasts with more volatility [44].
It is a method for projecting energy use for the next few minutes to
several days. In several grid activities requiring dispatch analysis and
3.1. Survey on very short-term load forecasting method
reliability analysis, it plays a crucial role. Additionally, it aids in
avoiding both overestimation and underestimate of the energy demand,
The time scale used in this kind of forecasting ranges from minutes to
thus enhancing grid reliability. Basically, the STLF is essential in
hours (0–3 h). It can assist in managing unpredictable fluctuations in the
developing financially sound, dependable, and secure operational plans
output of renewable energy that can be anticipated in a very short
for the power system. STLF used to balance the supply of power and load
amount of time. It can be applied to a wide range of RESs, such as
demand as an essential part of scheduling electricity generation,
forecasting wind and solar power generation. A few minutes in advance,
providing a foundation for energy dispatch and administration. Addi-
forecasts are created, and the forecast values are sent to the control unit
tionally, precise STLF results can reduce energy waste and avert energy
for transmitting in real-time. In order to react promptly to intraday
emergencies [52]. The deterministic point forecasts and the probability
variations in electricity demand, this kind of forecasting is employed
forecasts are two categories into which conventional STLF techniques
[45]. The key findings from the reviewed articles on VSTLF are sum-
can be broadly divided [53]. STLF techniques are generally divided into
marized in Table 3.
four types. Those are as follows:
The Gradient Boosting Regression Trees (GBRT) algorithm improves
forecasting accuracy while requiring less computation time. Integrating
• Statistical Technique
production data through the integration of a MES in the ETA-Factory or
• Artificial Intelligent (AI) Technique
by repeating the machine learning process with time steps shortened
• Knowledge-Based Expert Systems
from minutes to seconds, which helps to achieve a better accuracy [51].
• Hybrid Techniques [54]
In [47], the authors present TPE algorithm which improves XGBoost’s
hyperparameters which increases, the accuracy.

7
M. Hasan et al. Energy Conversion and Management: X 26 (2025) 100922

Fig. 8. Power load hierarchy ladder and different policy impacting variables [42].

Fig. 9. Types of load forecasting based on forecasting horizon [25].

Table 2
Main categories of load forecasting based on time-horizon [43].
Type Time Forecasted Accuracy Operation Planning
Intervals Outcomes

VSTLF Few minutes/hour Load-frequency Moderate Energy Management System (EMS) Load-frequency control
STLF 24 h-1 week Load curves Fixed load Economic Unit commitment
curves load dispatch
MTLF 1 week-1 year Load curves Capacity≫ Unit Reserve planning
Error commitment
LTLF >1 year Energy Fixed Planning of the power system Future capacity expansion
needed energy

The two main techniques used in the majority of studies on short- as presented in Fig. 10, which consists of four modules: data acquisition
term load forecasting are as follows: statistical analysis and artificial models, data preprocessing, model training, and load forecasting model.
intelligence based techniques known as machine learning [10]. The significance of STLF lies in its ability to assist facilities in plan-
Hong et al. [35] provided a basic framework for the STLF approach ning for peak electrical load generation and load shedding, and other

8
M. Hasan et al. Energy Conversion and Management: X 26 (2025) 100922

Table 3
Summary of reviewed articles on VSTLF.
Sl. References Year Methods Real-time applications Features Error
No.

1. [46] 2019 Machine learning approach (MLA) Germany (ETA research factory- • It is used to predict the electric load of the ETA research 0.73
TU Darmstadt) plant.
• It can choose the optimal hyperparameters and
appropriate characteristics for a forecasting model.
• Most peaks cannot be predicted by the model.
2. [41] 2021 Extreme Gradient Boosting (XGBoost) Real wind farm, Spain • It is used for wind power forecasting approach. 1.4271
Optimized by TPE Algorithm • It is tested with the real wind farm data.
• There are Excessive manual interventions for
hyperparameter tuning.
3. [42] 2021 Markov-chain mixture (MCM) Residential areas electricity, • Its implementation aims to predict home power usage one N/A
Australia step ahead with a half-hour precision.
• It is computationally inexpensive. For each forecast,
MCM employs a single data point.
4. [49] 2022 Deep-Auto former model Local houses from Los Angeles, • It is used to deal with the VSTLF in a microgrid system. 0.51
New York, and Texas
5. [50] 2023 Multiplicative Neuron Model Electric load of Uttarakhand, • When compared to MLP, it lowers both the likelihood of 0.0020
India overfitting and the computational complexity.
• It may be used for extremely short-term load forecasting
instead of the many neutrino MLP model.

Fig. 10. The short-term load forecasting framework’s schematic overview [35].

operational decisions. Machines learning methods such as ANN, CNN, Fast and precise converging STLF (AFC-STLF model) improves ac-
SVR, LSTM etc. are some popular methods for STLF. Table 4 presents a curacy by reducing average execution time [77]. With the actual values,
comparative analysis of these techniques [43]. the ECNN outperforms the conventional CNN in terms of accuracy [78].
For demand response and energy administration in the power sys- In time-series data, the Residual Connection Bidirectional Long
tem, forecasting short-term residential load is becoming increasingly Short-Term Memory with Attention Mechanism (RBiLSTM-AM) model
important [56]. Fig. 11 presents the general outline of the framework for performs exceptionally well at identifying intricate patterns and
one hour to 24 h ahead of short-term load forecasting. The framework is nonlinear relationships. In [79], the RBiLSTM-AM model significantly
comprised of six phases that involve data collection, feature selection, increases prediction accuracy by incorporating residual connections and
data processing, model training and testing and finally selecting the best attention mechanisms.
model for the given data.
Scholars specializing in power systems regard Short-Term Load
Forecasting (STLF) as exceptionally compelling and intellectually stim- 3.3. Survey on medium term load forecasting method
ulating. M. Li et al. proposed a prediction model for STLF on the basis of
Temporal Convolutional Network (TCN). In the same year, another team It is employed for time spans that extend from a few days to several
of researchers [43] adopted the Iterative Resblocks Based Deep Neural months into the future. It supports fuel supply, adequacy evaluation, and
Network (IRBDNN) method for STLF. Some statistical analysis methods maintenance for smart grid systems. Moreover, it contributes to risk
have also been employed in a large number of researches. Among sta- management and is crucial for assessing the financial qualities of the
tistical models, Autoregressive Moving Average (ARMA), Autoregressive energy system. Due to its crucial function in the operation and planning
Integrated Moving Average (ARIMA), Seasonal Autoregressive Inte- of the power system, which includes its best planning of generation
grated Moving Average (SARIMA) models showed better performance units, comprehensive planning program for customer service, and
for STLF. Table 5 presents a comprehensive summary of prior studies on economical supply, MTLF needs more consideration [37].
STLF. MTLF can be made for a period of one to twelve months, and a
In order to improve the predictive accuracy, XGBoost constructs an medium-term time horizon of three months to three years is recom-
ensemble of decision trees, adding each tree one after the other to mended. However, from another angle, MTLF is implemented for a
minimize the loss function [76]. duration of one to many years [80]. Operators can use MTLF to inform
their decisions regarding unit commitment, system safety analysis,

9
M. Hasan et al. Energy Conversion and Management: X 26 (2025) 100922

Table 4
A comparative analysis of different ML methods for short-term load prediction [48].
Machine Advantages Disadvantages Mitigations
Learning
techniques

Decision Trees • Capability of obtaining • Delicate with data • Monitoring for noise in a production environment, or put in place
(DT) nonlinearity • Prone to mistakes in sampling steps
• Resilient to anomalies • It lacks a configurable solution • feature reduction technique
• Adaptable • Greater time needed for instruction • Linear Regression model for linearity
ANN • Increased predictive precision • Low repeatability • Appropriate network structure is achieved through experience and
• Uncomplicated architecture • Prone to being too big trial and error
• Able to extract nonlinearity and • Having trouble managing several
complicated features from the variables
data
Random Forests • Effective management of missing • To make a difference, input data • Feature selection and hyperparameter tuning are crucial steps in
(RF) data attributes need to be strongly associated optimizing the performance
• Less subtle in their shapes with a goal output
• Decreased variance and error • Tuning a hypermeter is challenging
• Extraction of feature importance
XGBoost • Emphasis on feature attraction • Tuning hyperparameters is challenging • By introducing penalty terms to the objective function during
• Less susceptible to anomalies • Having trouble understanding training, it prevents overfitting by providing L1 and L2
• Decrease the data’s bias • It performs poorly with sparse data regularization terms using the ’alpha’ and ’lambda’ parameters.
• Excellent output of work
Multi ¡Layer • Strong performance in modeling • Attaining local minima can provide • Using more parameters, data, and time to converge helps to make it
Perceptron and pattern recognition challenges easier to implement, train, and understand
(MLP) • Improved results in nonlinear • It interferes with training time
problems • Slower convergence is possible
LSTM • Improved results with sequential • High training time • Use of hyperparameter optimization can control the behavior and
data (time-series) • Need specialized skills to adjust performance of the model, such as the number of layers, units,
• Able to handle big datasets hyperparameters epochs, learning rate
• Capable of processing non-linear • Generalization is hard to do
data and picking up complex
information
CNN • Significantly more effective in • Not universal • Adding More Images, Regularization, and Reducing architecture
terms of training time than LSTM • Need specialized skills to adjust complexity will help to reduce overfitting
• Capable of handling both 1-D and hyperparameters
2-D data
• Capable of handling non-linear
data and learning intricate fears
from data
SVR • Functions well with data that is • The size of the data affects performance • Scaling can increase the model’s stability and accuracy.
highly dimensional since training time grows as data size
• Able to manage structured and does
semi-structured data • The size of the data affects performance
• Reduced likelihood of over fitting since training time grows as data size
does
• Finding an appropriate kernel function is
difficult
• It takes knowledge of data distribution to
identify an appropriate kernel function

Fig. 11. Experiment for short term electricity demand forecasting [57].

10
M. Hasan et al. Energy Conversion and Management: X 26 (2025) 100922

Table 5
An overview of the STLF articles that have been evaluated.
Sl. References Year Methods Real-time applications Features Error
No.

1. [58] 2020 Hybrid method based on TCN and National Electric market (NEM) in • It’s used to reduce the discrepancy N/A
attention mechanism Australia and the National Centers between the actual and anticipated
for Environmental Information numbers.
(NCEI) in the United States
2. [55] 2020 IRBDNN Smart grid, Boston Area • By utilizing the spatio-temporal corre- 0.6159
lation between the load data from ap-
pliances, it enhances the performance of
STLF.
3. [59] 2022 Novel NODE (neural ordinary Hourly load data, University of • It is applied to improve the forecasting 0.165
differential equation) with LSTM Texas at Dallas accuracy.
• Complex and dynamic future trends can
be more accurately predicted using
NODE-based models.
• It is not adopted to use multivariate time
series datasets; when future steps are
short, NODE continuity is not
remarkable.
4. [60] 2020 Deep belief neural network (DBN) Eight different buildings, hourly • It reflects better adaptability. 0.276
load data taken from University of • For low-voltage consumers, it increases
Texas at Dallas the precision of short-term load
predictions.
5. [61] 2020 Extreme learning machine (ELM) Smart metering terminal (SMT) • Its applications include the widespread 0.393
enables huge electric big data, great collection of electric large data.
opportunity for electric shop • The reduction of data transmission and
market cloud computing tasks, and its tendency
to achieve the minimum training error.
• Requires a lot of electrical data.
6. [62] 2020 A CNN-BiGRU hybrid neural Load datasets-New Zealand & load • Unlike other models, it can thoroughly 0.05
network-based multi-model forecast example-Zhejiang, China find the numerous aspects of the
integrated prediction algorithm and provided load data and produce a
a random forest algorithm-based superior prediction result.
feature selection technique
7. [63] 2020 ARMA, ARIMA & ARIMAX Karnataka State Demand Data-2019 • The daily hourly loads and time of day ARMA has forecasting
are taken into consideration by the error of 17.7 %, while
ARMA and ARIMA models. While the that of ARIMA is 4 %
loads behave quite differently on and ARIMAX is 3.6 %
weekdays, weekends, and public
holidays, the ARIMAX Model accounts
for these events.
8. [64] 2022 Regression, SARMA and Ensemble Korean National Holiday • It functions by considering the four 0.026
learning main variables that influence the net
load profiles: weather, trend, BTM PV,
and calendar.
9. [65] 2020 Cross Electrical load data-Maine, the • The secret to raising the forecasting 0.5979
multi-model and second decision region of New England, Singapore accuracy of multi-model approaches is
mechanism (CMSDM) & New South Wales of Australia the model aggregation strategy.
10. [66] 2020 CNN-LSTM model Bangladesh Power System • It is put into practice to forecast load 0.0985
across various time spans.
• The long sequence time series electric
load data is handled by it.
• The accuracy of load forecasting is
impacted by the convolutional layer of
CNN’s scale choice.
11. [67] 2021 A bottom-up probabilistic load High voltage transmission • An HV substation is the intended use. 0.018
forecasting method based on Substations, China Better HV transmission substation
MV load data load forecasts might be obtained with
the help of the suggested approach.
12. [68] 2022 Neural Four seasons data given as input for • Better management of the energy 2.35
Networks determining totality of factors network is made possible by the
affecting energy consumption, the development of more precise electrical
ambient temperature and amount energy consumption projections based
of precipitation on artificial neural networks.
• It is also important to note that
electrical load consumption forecasting
may be done using low-power computer
equipment by utilizing a small number
of hidden layer neurons.
13. [69] 2022 Time Augmented Transformer New South Wales load data that was • The model uses less data to capture load 2.62
(TAT) made available to the public by the characteristics and produce more
Australian National Electricity accurate predictions, occupying a more
Market desirable prediction performance and
practicability.
(continued on next page)

11
M. Hasan et al. Energy Conversion and Management: X 26 (2025) 100922

Table 5 (continued )
Sl. References Year Methods Real-time applications Features Error
No.

• It is more capable of figuring out the

complicated input sequence’s dynamic
dependencies.
• Compared to other models, it can
produce superior prediction outcomes
with less historical data.
• It has proven to be highly successful in
short-term electrical load forecasting;
when the prediction step is increased,
the accuracy of the forecast rapidly
declines.
14. [70] 2023 Nonlinear Spiking Neural P Systems Task conducted on datasets of • It is capable of accurately capturing the 0.004
(LF-NSNP model) different time and variable types relationship between the electrical load
and compared with various sequence’s temporal properties.
Vaseline models • Its MAPE metric value is really low.
15. [71] 2023 RF-CNN-SVM N/A • It is employed to meet the increasing 0.6030
demand for short-term load forecasting
accuracy.
• It has better load forecasting effect
compared to CNN-SVM, CNN-LSTM and
SVM models.
• It does not consider some factors such as
electricity price and holidays that may
affect the load.
16. [72] 2023 AM-CIF-LSTM Medical, ecological and other fields • In emergency situations, it is utilized to 2.820
precisely forecast the load pattern.
• There is a decrease in total prediction
error and an increase in prediction
accuracy.
• Obtained prediction results contain less
information.
17. [73] 2023 SARIMA Load forecasting in Hanoi City • It is implemented to Forecast economic 0.5
objects; domestic and foreign research
on load forecasting.
• It satisfies the conditions of random
residuals and forecasting conditions.
• The approach and model have certain
restrictions on the predicting range.
18. [59] 2023 LSTM-CNN-based SAM Applied in residential, building, and • It is applicable to load forecasting across 0.02
industrial levels various time horizons and aggregation
levels.
• It may decrease the amount of input
data without affecting forecast
accuracy.
19. [74] 2023 An ultra short-term power load Data obtained from a random days • From the power load data, it is able to MAE = 0.015,
forecasting method based on similar electric load (in every 4 h with 15 extract patterns of short-term local RMSE = 0.028
day clustering and EEMD min interval) reliance.
(Ensemble Empirical Mode • Based on the power load’s periodicity, it
Decomposition) can simplify the optimization process
and capture long-term trends.
• It achieves greater prediction accuracy
compared to other models of a similar
kind.
20. [75] 2023 Temporal Applied in different time horizons • It has ignored external validity and 2.43
Fusion Transformer (TFT) and network levels (grid and proven experimentally inconclusive.
substation level) • In the electric grid, hierarchical load
forecasting can be used to achieve it.
• The TFT appears to require more
computing work to train because to its
substantially larger parameter set.

dispatching timetable, and analysis of load flow. Consequently, frames are used by researchers for MTLF. While some researchers have a
increasing MTLF accuracy is essential for raising system efficiency and one-month time horizon, others have a one-year time horizon. Addi-
cutting expenses [81]. A reliable method for MTLF, that is high- tionally, their forecasting models differ. In a study F. Yaprakdal et al.
precision load forecasting, can effectively support power management, [17] have employed statistical learning models like Linear Regression
and this is acknowledged as the foundation of economic dispatch in (LR), Decision Tree (DT), Random Forest (RF). Deep learning alongside
power systems [14]. other machine learning models have demonstrated effectiveness and
Based on a recent study [40], in the schematic design of Fig. 12, it surpassed conventional stat5istical time series methods. Table 6 pro-
can be observed that neural networks and deep learning have been vides an overview of the most significant research that has been carried
implemented to establish forecasting models for short-term as well as out recently with respect to mid-term load forecasting, with a focus on
long-term load forecasting. intelligent techniques [83]. Deep learning alongside other machine
A survey of the literature indicates that different methods and time learning models have demonstrated effectiveness and surpassed

12
M. Hasan et al. Energy Conversion and Management: X 26 (2025) 100922

Fig. 12. Electric load forecasting system schematic design for mid-term load forecasting [82].

conventional statistical time series methods. Table 6 offers a detailed literature to address this issue. As illustrated in Fig. 13, forecasting
overview on the most prominent recent researches conducted in the models can be broadly categorized into three categories based on sta-
realm of MTLF emphasizing the application of advanced intelligent tistical, AI prediction, and combined prediction models. Statistical
methodologies. model generally includes regression method, exponential smoothing,
CNN and LSTM provide more accurate forecasting for MTLF. and Kalman filtering among others. According to recent studies, deep
Improved performance can be achieved by optimizing the parameters of learning and machine learning models are the two most prevalent AI
these algorithms, such as squared error, Mean Absolute Error (MAE), prediction models used in load forecasting.
Mean Absolute Percentage Error (MAPE), and Root Mean Square Error Exponential smoothing, linear regression, the similar day approach,
(RMSE) [82].In [88] using a real-world dataset, the proposed hybrid and autoregressive moving average (ARMA) are the statistical tech-
LSTM–BPNN-to-BPNN model is compared against several benchmark niques most frequently utilized in the literature. The complexity of the
models, including ARIMA, MR, SVR, BPNN, and LSTM. Deriving load nonlinear interactions between the input and output variables cannot be
characteristic indicators from analyzing peak load characteristics helps fully learned using these methods, though. As a result, they don’t offer
to reduce predictive error significantly. adequate solutions for these issues. ML techniques, however, have the
capacity to represent intricate nonlinear mapping between inputs and
outputs, uncover hidden patterns in massive volumes of data, and pro-
3.4. Survey on long term load forecasting method
vide scalability, which allows them to address the limitations of statis-
tically based models [57]. Neural networks are an appropriate load
The time frame used in this kind of forecasting can be anything from
forecasting technique because of the nonlinearity of the load [90]. The
months to even years. For all production and load growth planning
two primary types of load forecasting models are those that employ
procedures involving long-term planning, LTF is crucial. The main
parametric and non-parametric techniques. Nonparametric approaches
benefit of LTF is its ability to forecast long-term trends while mitigating
are built on artificial intelligence, whereas parametric methods are
the consequences of short-term random changes. In general, long-term
based on analytical frameworks [91]. Table 7 provides an overview of
load forecasting is the prediction for the next five to twenty years.
the findings from the reviewed articles.
There is a non-linear association between this kind of forecasting and
A CNN-BiLSTM model, introduced in [100], combines a convolu-
other elements as well [81]. The goal of LTLF is to coordinate the
tional neural network with a bi-directional long short-term memory
planning of several divisions in a utility including system planning,
(BiLSTM) to efficiently process sequence data. By using Bayesian opti-
finance, management of demand side, as well as power supply. Future
mization, the model’s hyperparameters are automatically adjusted to
generation and transmission plans are directly impacted by the LTLF’s
improve accuracy, while reducing the number of parameters to enhance
accuracy. As a result, LTLF is crucial for planning and forecasting the
performance and reduce computational load. In [92], an LSTM-RNN
upcoming state of a power grid. The development of distributed gen-
model is proposed, where the high resolution of the data enhances ac-
eration (DG), electric vehicles (EVs), and financial factors are only a few
curacy, as the Recurrent Neural Network (RNN) model is particularly
of the many unpredictable elements that have an impact on LTLF [12].
effective for time-series data.
Three basic approaches—traditional statistical models, machine
learning models, and deep learning models—have been put out in the

13
M. Hasan et al. Energy Conversion and Management: X 26 (2025) 100922

Table 6
Overview of MTLF publications that have been reviewed.
Sl. References Year Methods Real-time applications Features Error
No.

1. [71] 2018 Ant-Lion Optimizer Least-square Data taken from PJM website-a • MLTF will calculate the amount of power used 4.3560
Support Vector Machine (ALO-LSSVM) regional transmission organization within a certain period of time. According to the
and Least Square Support Vector (RTO) and selected from Duke prediction, since there won’t be any
Machine (LS-SVM) overgeneration, operational costs will be
eliminated.
• In addition, there is a chance to prevent the
power outage or trip.
2. [84] 2018 Quantile Regression (QR) Actual data set from United States • It efficiently obtains the probabilistic load 1.753
forecasting result in addition to achieving a more
accurate point prediction result than individual
models.
• The future load probability distribution is
described in greater depth, and dispatchers and
planners can benefit from the additional
information it offers.
• Compared to the separate models, it produces
prediction results that are more accurate and
thorough.
3. [85] 2019 A hybrid method, composed of SVR Method tested on EUNITE • With optimization techniques, it is possible to 1.39
and Symbiotic Organism Search competition dataset maximize the amount of most important
Optimization (SOSO) method characteristics while minimizing redundancy,
and this provides a foundation for further study.
It is simplifying the primary issue, which
facilitates its resolution.
4. [86] 2020 Generalized regression neural network Data collected from Kano • This method aids in lowering significant <10
(GRNN), feed-forward neural network electricity distribution company technical losses and enhancing power
(FFNN) and radial basis function (KEDCO) distribution quality.
neural network
5. [87] 2020 Hybrid system based on neural Non-metal mineral industry • It makes system limitations and the structural 2.30
network and technological theories Black metal smelting and rolling range of power users possible to take into
processing industry consideration.
Chemical raw material and • It is possible to make an accurate prediction by
chemical product manufacturing investigating the cause of load fluctuations when
industry there is a “turning point” in the load.
Coal mining and washing industry
Textile industry Accommodation
and dining industry
6. [21] 2022 Deep learning models such as LSTM, Method used for the benefit of the • The electric utility company’s future needs for 3 Months
MLP, CNN + LSTM, and 1D-CNN electric utility company personnel recruitment, equipment sales, and Prediction:
expansions are better met by this model. MLP: 2.72 %
• It could be better suitable for a possible problem 6 Months
with power growth. Prediction:
MLP: 2.36 %
9 Months
Prediction:
MLP: 2.38 %
1 Year
Prediction:
MLP: 2.30 %
6 Years
Prediction:
MLP: 2.49 %

7. [88] 2022 Hybrid LSTM–BPNN-to-BPNN Method applied to predict the • It shows a noticeable gain over the LSTM, 1.47
monthly peak load of a year highlighting the need to take future data into
compared to the previous year account for medium- and long-term peak load
annual peak load forecasting.
• When predicting peak demand over the medium
and long terms, this model performs well.
8. [19] 2022 Linear regression (LR), decision tree Smart Grid • Deep learning and other machine learning 1.61
(DT), random forest (RF), gradient models have proven effective and outperform
boosting, adaBoost, conventional time series and regression methods.
RNN and LSTM
9. [89] 2022 Linear regression, Artificial Neural Indonesia is a tropical nation with • The used technique falls under the category of 2.63
Networks, and Splitted Linear unique features compared to medium-term load forecasting, which is charac-
Regression others with four distinct seasons. terized by results in the form of daily energy
consumption or peak loads alone. Each model’s
findings are presented as half-hour data, which
may be utilized for short-term load forecasting.
10. [45] 2022 Linear regression, support vector Forecast real load power system • It enables the design, administration, and 1.30
regression (SVR), decision tree operation of isolated EPSs to employ information
regression, random forest (RF),
(continued on next page)

14
M. Hasan et al. Energy Conversion and Management: X 26 (2025) 100922

Table 6 (continued )
Sl. References Year Methods Real-time applications Features Error
No.

XGBoost, adaptive boosting over technology more extensively and predict

decision dependability to rise.
trees (AdaBoost), AdaBoost over linear • This study makes it possible to plan the loading
regression mode of generators in an isolated power system,
which in turn allows for the optimization of the
generators’ working mode while accounting for
the water resources that are accessible.

Fig. 13. Classification chart of load forecasting models [39].

3.5. Summary on load forecasting methods earlier loads by using STLF as a base. The VSTLF can be obtained by
adding future residuals to the STLF.
To enhance the short-term model, precise prediction information of
the influencing components is crucial. Since instantaneous demand 4. Classification of load forecasting models
might vary, the fundamental goal is to find the relationship between
load demand and factors. MTLF is typically used for load forecasting There are two primary categories of forecasting models: artificial
across time periods ranging from days to months. During the height of intelligence (AI) based models and classical statistic models. Traditional
summer or winter, it gains popularity. LTLF considers load durations statistical models include things like moving averages, exponential
ranging from a few weeks to many years. It considers a number of var- smoothing, regression analysis, stochastic time series models, etc. ANNs,
iables, including weather information, install device characteristics at fuzzy logic, data mining, machine learning, genetic algorithms, and
interest locations, load history, and customer count. Long-term load knowledge-based expert systems are examples of AI-based models. The
forecasting techniques take the economy into consideration. In the case several load forecasting models are described here and comparative
of VSTLF, instead of using the modelling relationships between time, analysis described in Table 8. [43,101].
weather, and load, the estimate of recently discovered load to the near
future is employed. Artificial neural networks, genetic algorithms, and
autoregressive moving average models are among the very few tech- 4.1. Similar day approach
niques available for VSTLF. Rarely is STLF used for durations of minutes
to hours. For both daily operations and system operations, STLF is a Similar day approaches are searched using historical data. Typically,
crucial source of information. Because the STLF may be used to estimate it looks through the data for days that are comparable to the projected
the long-time load, researchers are becoming more interested in creating day within the last one, two, or three years. Certain indices, such the
prediction models. date, the specific week’s day, and the weather, are regarded as
Based on this extensive literature survey, it can be said that STLF is comparative individualities. This method can involve a regression
the most widely used among the four approaches. Because of its inherent analysis or a linear combination technique with several comparable
relationship to other classes of forecasts, it is essential to the develop- days, as opposed to considering a load of similar single days. The con-
ment of operational plans for power systems. The STLF was transformed dition typically clarifies this is as below.
into the MTLF and LTLF by the addition of the economic element. D = √(W1 (ΔL1 )2 + W2 (ΔHh )2 + W3 (ΔTt )2 ) (1)
Additionally, the STLF model can be transformed into VSTLF by
including prior hour loads as an input factor. The autocorrelation be- where, ΔL1 is the variation in load between the predicted day and his-
tween the loads of the previous and current hours can be captured by torical days, ΔTt is the deviation in temperature between the forecast
STLF. It is possible to create new series and acquire the residuals of day and historical days, and ΔHh is deviation in humidity between the

15
M. Hasan et al. Energy Conversion and Management: X 26 (2025) 100922

Table 7
Summary of reviewed articles on LTLF.
Sl. References Year Methods Real-time applications Findings Error
No.

1. [11] 2018 Regression Model The ISO New England Market • Forecasts peak load of a month 26.09
• Population, weather, and number of households were %
ANN model ​ considered as fuzzy variables 8.30 %
Fuzzy ANN model ​ • The regression model employs linear variable 5.73 %
combinations and is unsuitable for simulating highly
nonlinear complicated data
• Fuzzy ANN provides better forecast result
2. [92] 2018 LSTM-RNN model Real time data of ISO New England Market • Forecasts electricity load for a period of five years 6.54 %
• Takes around 30 min to compute
• Monthly maximum load demand can also be calculated
• For longer-term forecasting of energy consumption,
LSTM-RNNs are the best option because of their faster
convergence and higher accuracy
3. [93] 2018 Feedforward ANN ​ • The effectiveness of the most popular ML techniques of 1.5 %
SVM ​ LTLF has been investigated 1.7 %
RNN New England Network • Applied in LTLF for monthly load forecasting 1.9 %
GRNN ​ • Weather indicators and historical load data were taken as 2.3 %
KNN ​ input variables 2.3 %
Gaussian Process ​ • When compared to other methods, feedforward ANN 2.0 %
Regression (GPR) yields superior results
4. [94] 2019 Box-Jenkins (BJ) technique ​ • Presents LTLF study based on demographic, economic, 4.76 %
(ARIMA model) technologic, and environmental factors
Nonlinear autoregressive Tunisian electric consumption • For 5–10 years period BJ approach gives satisfactory 0.81 %
neuronal network, results, its effectiveness is restricted beyond a 10-year data
NAR period
Multiple linear regressions, ​ • NAR works well with data period beyond 10 years 3.26 %
MLR • MLR is suitable for linear variables
Nonlinear autoregressive ​ • NARX gives better performance for 5-year period as well 0.58 %
neuronal network as beyond 15-year period
with exogenous variables, • BJ shows lowest performance and NARX shows best
NARX performance of all
5. [95] 2019 Multivariate adaptive Turkish electricity distribution company • Applied for modelling electrical load in the distribution 4.0 %
regression splines (MARS) network
ANN ​ • Hourly load, wind, humidity, and temperature data are 4.5 %
Linear regression (LR) ​ included in the input dataset 4.7 %
• MARS, ANN and LR yields precise load forecasting
• MARS technique gives more precise forecasting result as
its MAPE and RMSE vales are lower. Processing data set
with missing value and handling mixed data is easier with
MARS than ANN and LR
6. [96] 2020 Deep Neural Network Data collected from electric utilities from • LTLF was computed for two commercial buildings with 0.21 %
(DNN) different customers and two industrial distinct energy consumption patterns
buildings at Naju-si, Jeonnam, South • The input dataset includes the hourly power usage data
Korea spanned three years, from 2017 to 2019
LSTM ​ • The proposed LSTM model consists of two layers, while 0.48 %
the proposed DNN model has four layers
• In terms of MAPE, the performance of the proposed DNN
outperformed that of the proposed LSTM
7. [97] 2021 Improved CNN + RFR ​ • Forecasts the maximum load of a distribution generator 13.68
• The improved CNN + RFR employs fewer data feature %
Improved CNN + BP ​ parameters 14.41
• CNN + RFR involves less recording time and MAPE was %
Improved CNN + DTR ​ smaller 15.03
%
Improved CNN + SVR Smart Grid 15.20
%
8. [98] 2021 Functional Nonparametric ​ • Uses trend line checking to forecast the trend of annual 1.17 %
BP-ARIMA ​ load 1.93 %
Combined Functional Regional power load • Combines annual LF with monthly periodic LF 1.59 %
Nonparametric and BP- • Combined BP-ARIMA increases forecasting accuracy and
ARIMA data stability
9. [21] 2022 LSTM ​ • 3, 6, 9 months and 1 and 6 years ahead load forecasting 2.49 %
CNN ​ was done 3.03 %
Multilayer Perceptron ​ • LSTM layer contained 64 neurons and ‘RELU’ as an 5.42 %
(MLP) activation function. 3 layers were present in MLP
CNN + LSTM Future requirements for electric utility • For long-term load forecasting (ahead 6-year forecasting), 3.51 %
company MLP was found to be the best method with lowest pre-
diction error
10. [99] 2022 ANN with Backpropagation Data on actual monthly power use from • Domestic, industrial, agricultural, governmental energy 26 %
Algorithm (BPA) 2011 to 2020, as supplied by the Ministry consumption have been taken as the input to the model
of power in Iraq and monthly peak load was the forecasted output
• In contrast to ANN + traditional BPA approach, the ANN
+ APBA yielded the lowest MAPE of 0.045
(continued on next page)

16
M. Hasan et al. Energy Conversion and Management: X 26 (2025) 100922

Table 7 (continued )
Sl. References Year Methods Real-time applications Findings Error
No.

ANN with Adaptive ​ • Whereas BPA was unable to comprehend the significant 4.5 %
Backpropagation variation in the load dataset, ABPA is capable of
Algorithm (ABPA) successfully capturing the unexpected pattern of changing
load data

forecast day and historical days. The weighted factor (Wi) is defined by seasonal trend and auto correlation, are essentially assumption tech-
temperature and load data (i = 1, 2, 3,…). The least squares approach, niques. These tactics recognize these underlying characteristics and
which is built through regression analysis utilizing temperature and establish a relationship between them. The time series approach has
historical load data, is used to generate this factor. Therefore, load, been widely employed in the field of prediction, including the fore-
humidity, and temperature patterns are considered while choosing casting of electric load, throughout the previous few decades. The time
similar days for load forecasting. series strategies that are commonly employed in STLF are the classical
ones, such as ARMA ARIMA, ARMAX (autoregressive moving average
4.2. Statistical technique with exogenous variables), and ARIMAX (autoregressive integrated
moving average with exogenous variables). An ARIMA is an extension of
Mathematical formulas that indicate the relationship between load an ARMA model, which represents non-stationary forms, while an
and several input parameters are the foundation of statistical ap- ARMA model represents stationary forms.
proaches. These techniques can be categorized into the following This approach predicts electrical load by using the rolling window-
groups. based strategy, which is based on the moving average notion. The
following equation serves as a mathematical representation of the S.M.
a) The approach of multiple regression A. strategy.
b) Simple moving average
1
c) Exponential smoothing Yt+1 = Σti=0 = t − Nt+1 Xi (4)
N
d) Time series methodologies
where N is the number of terms included in the average, Xi is the
observation at time t, and Yt+1 is the prediction value at time t + 1. One,
4.2.1. Multiple regression method two, or three time periods can be forecasted using this method. But
In the subject of STLF, this particular type of statistical technique is because forecasting error rises with increasing time period, it is often
employed most commonly. Building a relationship between load con- applied to one-time period forecasting.
sumption and other variables, such as the day type index, the weather,
and finally the client classes, is necessary to forecast electrical load. 4.2.4. Time series approach
Different regression techniques are typically employed to create such Time series approaches, which include internal elements like sea-
relationships. Additionally, the correlation of twelve coefficients be- sonal trend and auto correlation, are essentially assumption techniques.
tween the load and potential independent factors are used to identify These tactics recognize these underlying characteristics and establish a
independent variables. The following multiple regression equation is relationship between them. The time series approach has been widely
created to represent the load pattern. employed in the field of prediction, including the forecasting of electric
→ load, throughout the previous few decades. The time series strategies
y = c0 x0 + c1 x1 + ⋯ + cn xn = C ⋅→
x +e (2)
that are commonly employed in STLF are the classical ones, such as
ARMA, ARIMA, ARMAX, and ARIMAX. An ARIMA is an extension of an
wherein the definition of the electrical load is y, xi indicates the variable
ARMA model, which represents non-stationary forms, while an ARMA
of interest, and the regression coefficient is cj for ∈ [0, n], and e is the
model represents stationary forms. Furthermore, the ARMA and ARIMA
modeling error. The row vector → x = [x0 ⋅x1 ⋯xn ] contains independent
models also use time and electrical load as inputs. Because ARIMAX
variables which are affecting the load, and the column vector → c = depends on weather and time of day, it is the most effective classical
[cc ⋅c1 ⋯cn ]t comprises of unknown regression coefficients. model for STLF among all of them. Here, in Table 8, a comprehensive
summary of the various LFT is presented.
4.2.2. Exponential smoothing
It is a time-honored process that is frequently applied to STLF issues.
4.3. Summary
The following is a mathematical representation of this model.
Yt+1 = aXt + (1 − a)Yt (3) In conclusion, statistical procedures are easier and need less
computational resources. These are used to determine the relationship
When the anticipated value at that moment t and t +1 is Yt and Yt+1 between predicted power and meteorological variables. The depend-
respectively. Real worth at the moment t is xt the most recent obser- ability of energy datasets with large and nonlinear data declines. It be-
vation is defined by the weight of a, and the most recent forecast a = N1 is comes challenging for those techniques to regulate complicated weather
indicated by the weight of (1- a) is the smoothing factor, and N is the patterns. On the other hand, time series analysis approaches can take
total number of observations used to calculate the average. Equation (3) into account seasonal impacts. The downside is that they are numeri-
makes clear that the value derived from the forecast at (t + 1)th any cally unstable.
given instant is dependent upon both the actual value that was seen at t
that instant and the anticipated value for that time period. The expo- 5. Measures for performance assessment of methods used
nential smoothing method uses prior observations in this way, requiring
less memory on the computer. To ensure that real load values are predicted accurately, load fore-
casting systems are evaluated based on several criteria. To maximize the
4.2.3. Simple moving average accuracy of their models, many scholars have employed recently
Time series approaches, which include internal elements like developed statistical metrics, such as probabilistic load forecasting

17
M. Hasan et al. Energy Conversion and Management: X 26 (2025) 100922

Table 8 Table 8 (continued )

An overview of the benefits and drawbacks of different forecasting models. Technique Benefits Drawbacks Mitigations
Technique Benefits Drawbacks Mitigations
data has well- • Might need and updating
Similar day • A simple • The approach • Combine the defined further models as new
approach technique that heavily depends similar day structures. preprocessing or data becomes
does not require on historical approach with • Able to adjust model available.
intricate data, which other techniques. over time to modifications to • Accurate
modeling. might not • Boosting the changes in the deal with non- forecasting can
• This approach entirely weighting system data. stationary time be improved by
utilizes accommodate by highlighting series. using ensemble
historical data novel events or the most similar methods to
from days with unforeseen days. include external
similar circumstances. variables.
characteristics to • Similar days are Traditional • The methods are • Not be as precise • Hybrid models
predict load. assumed to show LFT well-respected as more integrating real-
consistent load and frequently sophisticated time data with
behavior. applied in the ones. machine learning
Statistical • Models are often • Might be • Regularization field. • Not be able to algorithms like
technique simple to difficult to techniques can be • Reasonably easy manage the ARIMA, and neu-
understand and represent used to eliminate to comprehend data’s ral networks can
easy to interpret. intricate overfitting. and put into complicated or significantly
• It usually comes patterns and • Forecasting practice. non-linear enhance adapt-
with built-in non-linear cor- accuracy can be • They don’t need relationships. ability and
systems for relations in the increased over a lot of • Extreme accuracy.
handling data data. time by information. occurrences or
noise and • It is challenging constantly missing data can
outliers. to make more reviewing the be too much for
advancements model’s them to manage.
without performance. Clustering- • The strategies • For these • Robust clustering
completely Based can find links methods to algorithms such
transforming the LFT and patterns in reach great as DBSCAN or
model. the data that accuracy, a lot of OPTICS can
Multiple • Able to manage • The accuracy of • Using different standard data may be improve data
regression multiple types of the load forecast ensemble approaches can needed. analysis
method predictor and the methods or miss. • Require a lot of outcomes by
variables. regression regression • Can be applied computing managing noise
• A well-known coefficients can models to to recognize and power to run. and outliers.
method with a be greatly identify non- predict trends in • They could be
large number of impacted by linear particular data challenging to
resources for outliers or relationships. sets or decipher and
application and important data • Reliability can be groupings. comprehend.
interpretation. points. increased by • Can deal with
• Gives numerical • It might have using robust extreme events
insights into trouble regression or missing data.
how several capturing techniques that AI-based LFT • Can deal with • To achieve great • Utilizing
elements affect intricate, non- are less intricate or accuracy, a lot of dimensionality
load forecasts. linear relation- vulnerable to nonlinear data data may be reduction
ships between outliers. relationships. needed. methods like PCA
load and • Can be applied • Need a lot of or t-SNE prior to
predictors. to recognize and computing clustering to
Exponential • Suitable for real • Might find it • Combining with predict trends in power to run effectively
smoothing time or near real difficult to other forecasting particular data them. manage high-
time forecasting identify methods such as sets or • Could be dimensional data.
applications. complicated machine learning groupings. challenging to • Appling Mini-
• Less vulnerable patterns or algorithms or • Can deal with decipher and Batch K-Means or
to noise or abnormalities in regression extreme events comprehend. BIRCH algo-
abrupt changes the data. models. or missing data. • These methods rithms for
in the data. • Sometimes it might not apply improved mem-
becomes more to different ory utilization
difficult to kinds of data. and computa-
identify longer- tional efficiency.
term patterns or Time Series • These methods • Included a novel • Using seasonal
data changes. LFT are well-known feature selection decomposition
Simple • Shows the • Might not keep • Evaluating and often technique into a methods to
moving average of up with real load several S.M.A. applied in the hybrid distinguish
average (S. historical load changes. variations. industry. forecasting between trend,
M.A.) values over a • Less appropriate • Removing • Relatively approach. seasonal, and
given time for time series outliers or simple to • Not be as precise residual elements
period which forecasts that applying data implement and as more within the data.
makes it easy to exhibit transformation understand. sophisticated • Consistently
interpret. significant techniques • Can deal with ones. verifying the
• Calculations are seasonal before utilizing S. extreme events • Not be able to uniformity of
numerically patterns. M.A. or missing data. manage the data collection
efficient. data’s intervals and
Time series • Can generate • Can have trouble • Regularly complicated or manage any
approach precise forecasts processing monitoring non-linear missing or
when applied erratic data. forecast accuracy relationships. irregular data
(continued on next page)

18
M. Hasan et al. Energy Conversion and Management: X 26 (2025) 100922

Table 8 (continued ) ∑n
(xi − yi )2
Technique Benefits Drawbacks Mitigations R2 = ∑i=1
n 2
(6)
˙ (xi − yi )
i=1
• For them to points
reach high accordingly.
The average values of all the samples are indicated here by yi .
accuracy, a lot of
data might be
needed. 5.3. Mean square error
Meta- • Can deal with • For these • Integrating with
Heuristic intricate data methods to machine learning
Based LFT relationships. reach great techniques, such
Mathematically, MSE can be calculated as follows:
• Can be applied accuracy, a lot of as neural n
1∑
to recognize and data may be networks, can MSE = (xi − yi )2 (7)
predict trends in needed. enhance n i=1
particular data • Need a lot of forecasting
sets or computing performance. The mean square error (MSE) of an estimator calculates the average of
groupings. power to run • Utilizing
the deviations or errors. The second moment of error, or MSE, considers
• Can deal with them. parallelization or
extreme events • It’s possible that distributed an estimator’s bias as well as its variance.
or missing data. these methods computing
• This techniques cannot be used methods helps in 5.4. Mean Absolute error
for global to different alleviating the
optimization can kinds of data. computational
identify the best • In order to complexity. The MAE is the measure of variation between two successive vari-
answers. identify the ables. If, for instance, variables y and x represent the actual and expected
global optimal values, the MAE can be computed as follows:
or near-optimal.
n
∑ (yi − xi )
MAE = (8)
n
metrics. Owing to widespread adoption and exceptional academic ben- i=1

efits in business, the research on probabilistic forecasting is still in its

early stages of development. The following section displays the most 5.5. Mean Absolute Percentage error
significant static measures that researchers employ. Fig. 14 illustrates a
listing of accuracy metrics. The accuracy of the predictor was quanti- The following formula is used to calculate the MAPE:
tatively measured using the coefficient of determination (R2), mean n
100 ∑ (yi − xi )
square error (MSE), mean absolute error (MAE), and root mean squared MAPE = (9)
n i=1 n
error (RMSE) to assess its quality. The well-known error prediction
measures found in the literature are presented in the next section
[102,103]. 5.6. Summary on assessment methods

From the exhaustive literature survey, it is found that every metric

5.1. Root Mean squared error has both benefits and drawbacks. The RMSE provides a two-degree loss
function, although it assigns greater weight to large errors than to minor
Metrics are needed to test a predictor’s accuracy quantitatively in ones. Naturally, MAE can be used to calculate the average error. MAPE is
order to assess its quality. For this reason, the current study established a independent of scale and can be used for both high and low volume
quantity known as RMSE, which is defined as follows: items. Variable penalties could result in skewed forecasts. The weak
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ points of MAPE are small and zero denominators that are difficult to
n
∑ control. Due to zero or extremely low levels of load at the aggregate
RMSE = (xi − yi )2 (5)
i=1 level, these weak areas are unrelated to issues with standard load fore-
casting. Furthermore, these metrics’ values change depending on the
Where, n represents the total number of samples in the training dataset parameters and datasets. Therefore, comparing the results of several
and xi and yi the measured and projected values of the i-th sample. The methodologies is very challenging considering different operational
superior collection of chosen descriptors is indicated by a reduced RMSE constraints.
value.
6. Characterization of the state-of-the-art methods of applied
load forecasting
5.2. Coefficient of determination
This section outlined the many approaches taken by survey articles
The following function is used to calculate R2: to the topic of LF models using various sensors and ML based methods.

Fig. 14. Accuracy metrics for evaluating the performances of the forecasting methods.

19
M. Hasan et al. Energy Conversion and Management: X 26 (2025) 100922

The primary focus of energy forecasting techniques is on future load of forecast performance. In another work, R. K. Agrawal et al. [83]
demands that will eventually be used in infrastructure development and utilized LSTM-RNN model for long term load forecasting. When applied
planning. Various types of sensors, such as temperature, humidity, and to real-time data, this model proved to be incredibly accurate and
consumption sensors, as well as energy-related data primarily measured suitable for long-term load forecasting (LTLF). Table 10 summarizes a
by smart meters, are used in establishing energy forecasting. Sensors thorough analysis of numerous machine learning-based load forecasting
could provide useful data to enhance energy management of any techniques.
building. In order to achieve improved forecasting results, appropriate A wide array of techniques has been proposed to accurately forecast
selection of sensor characteristics for the load forecasting input should electrical load, including linear and non-linear models, statistical
be made. With the increased use of sensors, it becomes a significant task models, deterministic and stochastic models, machine learning and deep
to optimize the power system’s dispatch using clever algorithms in order learning methods, as well as hybrid models that combine classical and
to achieve efficient functioning. In a study, D. Zhang et al [91] proposed modern approaches. Here, in Table 11, the advantages and disadvan-
a model for predicting real time LF based on CNN-BiLSTM. The power tages of a few state-of-the-art forecasting techniques are briefly sum-
system load in real time was accurately predicted by the model. In some marized and compared.
additional research, real-time sensor data was used, and other mod-
els—such as CNN, LSTM, and RNN—were used to forecast load. Table 9
presents a summary of the articles that have been reviewed in this 6.1. Role of improved intelligent optimization algorithm and combined
particular area of study. prediction model
In previous years, researches on load forecasting have often relied on
conventional techniques like regression-based analysis and time-series For intelligent and optimized load forecasting, a single optimization
models of analysis. However, these methods often exhibit limitations algorithm frequently faces a few drawbacks, including computational
such as poor accuracy, inadequate incorporation of factors like meteo- complexity, increased error rates and decreased computing efficiency. In
rological data, and low sensitivity with regard to input data. The order to achieve greater accuracy with a lower error rate, academics
tremendous advancements in computer science and artificial intelli- have been developing hybrid optimization models for years. In general,
gence have made machine learning (ML) models a potential tool for load hybrid models combine two or more separate methods, each of which
forecasting in recent years. Load prediction of buildings and commu- contributes to more precise and effective forecasting. The role of
nities using machine learning has been the subject of numerous research improved intelligent optimization algorithm such as ANN, CNN, DNN
papers. In a study [98], H. Habbak et al. concluded that AI-based LF etc. and combined prediction model (i.e., CNN-LSTM, CNN-GRU and
techniques utilizing ML, especially neural network (NN) models like LSTM-Transformer) in enhancing the accuracy of load forecasting model
Artificial Neural Network (ANN) outperformed other methods in terms presented here. Some papers such as [11,93,96,99] presented intelligent
optimization algorithm such as ANN, CNN, DNN etc. showed better

Table 9
Summary of sensor based reviewed articles.
Sl. References Year Methods Real time applications Findings Error
No.

1. [100] 2023 CNN combined with bi-directional Smart grid real-time load The smart grid real-time load can be accurately 9.45 %
long and short-term memory BiLSTM prediction model predicted by the model.
(CNN-BiLSTM)
2. [104] 2021 DNN Day-ahead hourly load In terms of accuracy, the clustered models 7.27 %
LSTM, LR forecasting at the level of outperformed the individual models.
distribution transformers Proposed scheme used real-world data and maintained 7.18 %
the forecasting accuracies.
20.32 %
3. [78] 2020 Two stage electricity load forecasting Energy generation data and Because the ECNN’s curve closely matches the actual ​
framework and enhanced CNN hourly electricity load data of data, it obtains a better accuracy.
classifiers the ISO GSA was employed to tune CNN super parameters to
achieve better efficiency and higher accuracy.
4. [105] 2020 • Distributed VMD-DBN STLF Smart grid Using the hybrid STLF on the Hadoop MapReduce 2.057 %
architecture results in increased accuracy and (Yichun
efficiency. City)

1.605 %
(Fuhzou
City)
5. [92] 2018 Long-Short-Term-Memory based Real time data of the ISO The LSTM-RNN model has been applied to real-time 0.0654
Recurrent Neural Network (LSTM- electricity market data and is ideally suited for time series electrical data.
RNN) The increased precision is facilitated by the great
resolution of the data.
6. [106] 2017 • IoT-Enabled Smart Meter Systems Smart grids with vast areas, The prediction precision for daily total consumption is 1.26 %
variable climates, and social greatly increased by using a two-step forecasting
conventions technique.
7. [107] 2017 Combined CNN and clustering A large real-world power This approach has demonstrated that forecast errors 3.055 %
technique industry data set can be reduced by effectively grouping different (Summer)
customers based on consumption behavior.
It has also been tested with two different real smart 7.4102 %
meter datasets. (Winter)
8. [108] 2015 k-means clustering algorithm Household load data from smart A substantial real-world power industry data set is 0.5 %
NN-Based Load Forecasting meters utilized to assess the model. (Dataset 1)
The data set is divided into subsets using the K-Means
approach to facilitate quicker and more effective model 1.07 %
creation. (Dataset 2)

20
M. Hasan et al. Energy Conversion and Management: X 26 (2025) 100922

Table 10
Summary of ML based reviewed articles.
Sl. References Year Methods Real time applications Findings Error
No.

1 [109] 2023 Multiple Linear Smart grid MLR models are extremely important in LF since they are clear, N/A
Regression Models interpretable, and widely used.
(MLR) Clustering-Based Load Forecasting Techniques can find
K-Means Clustering patterns and linkages in the data that older approaches might
Neural networks (NN) miss. NN helps to get an accurate model of some real-life
systems.
2. [104] 2021 DNN Day-ahead hourly load forecasting at The accuracy of the clustered models was significantly higher 0.0727
LSTM the level of distribution transformers than that of the individual models.
LR The forecasting results showed that the suggested technique 0.0718
produced precise clusters and saved 44 % of the training time.
The suggested technique may be applied at any level to 0.2032
massive electrical networks and enormous data in smart grids.
3. [110] 2021 ANNs Data (from Greek connected power The availability and quality of the training data determine how <0.02
system) used to evaluate proposed well the ANNs forecast.
model
4. [111] 2019 Regression Simple NN Industrial load Algorithms for machine learning can greatly reduce forecast 0.2661
Approach mistakes.
Classical regression 0.2814
algorithm
LSTM 0.1710
Random Forest
0.1847
5. [112] 2019 Feed-forward DNN Ho Chi Minh City electricity load The quantity, scope, and nature of the chosen datasets 0.350
Recurrent DNN forecasting determine the efficacy of the FF-DNN and R-DNN deep
learning algorithms.
Combined model of FF-DNN and R-DNN gives faster and more
accurate forecasting results
6. [92] 2018 LSTM-RNN Real time data of the ISO New England The LSTM-RNN model has been applied to real-time data and is 0.0654
electricity market proven to be very accurate.
It is particularly well-suited for time series power data.
7. [107] 2017 Combined CNN and A large real-world power industry data The model is tested using a sizable real-world power industry 3.0554
clustering technique set data set. (summer)
However, it is unclear if the settings represent the best possible
combinations of these variables. 7.4102
(winter)
8. [113] 2017 Multidirectional LSTM The smart grid dataset from UCI MLSTM approach outperforms the other ML approaches. 0.02
Machine Learning Repository MSLTM model is used to analyze the power utilization and
predict the stability of the SG dataset
9. [114] 2017 RNNs Short bench-mark international airline Complex univariate electric load time series with substantial 0.0535
passenger data set and a long non-stationarity and non-seasonality can be forecasted by
electricity consumption data set LSTM.
For more precise forecasting, the LSTM-based RNN N can take
advantage of the long-term dependencies in the electric load
time series.
10. [115] 2017 DNN Demand side empirical load database The DNN model performed better than ARIMA, SNN, and 0.0329
DSHW forecasting models.
The DNN with ReLU is simple to train and produces reliable
and accurate predictions.

performance and better accuracy than other models. From Table 11, ML the Bangladeshi power system, the created model’s performance is
based technique used for produces more precise, reliable and accurate assessed. In comparison to the LSTM, RBFN, and XGboost models, the
predictions in load forecasting. Combined prediction model or hybrid suggested CNN-LSTM model yields the lowest MAE, RMSE, and MAPE
model presented better output or results. values. In each table, in comparison to the LSTM network, the CNN-
This paper suggested an LSTM-CNN model architecture based on LSTM often yields lower MAE, RMSE, and MAPE, respectively. There-
SAM. CNN + LSTM based SAM outperformed LSTM, CNN GRU based fore, in short-term electrical load forecasting, CNN-LSTM is found to be
SAM, and LSTM based SAM. So, the hybrid LSTM-CNN model was better more effective than the suggested strategy [66]. This research suggests a
than only LSTM model because LSTM-CNN model has a 0.164 lower VMD-Pyraformer-Adan model for short-term power load forecasting
MAPE, a 70.154 lower MAE, and a 94.030 lower RMSE where LSTM has because of the power loads’ volatility, periodicity, and nonlinearity. The
a 0.196 MAPE, a 83.866 MAE, and a 111.850 RMSE [59]. Initially, the VMDPyraformer-Adan model has the lowest MAPE (0.65) and RMSE
power system load dataset’s input attributes are chosen using the (70732) rather than single LSTM model which showed MAPE (1.46) and
random forest technique to produce more precise predicting outcomes. RMSE (159364). Since the projected values more closely resemble the
A CNN-BiGRU hybrid neural network with multi-model fusion is then actual values, the VMDPyraformer-Adan model put out in this research
presented to forecast short-term load which showed higher accuracy has a higher accuracy in short-term power load forecasting than the
using both MAPE and RMSE. The integrated CNN-BiGRU model is su- other models [71]. This article suggested a hybrid LSTM–BPNN-to-
perior than the single CNN-BiGRU model, as evidenced by the fact that BPNN model that combines a back propagation neural network (BPNN)
the prediction accuracy of the stack model is higher for both datasets and a long short-term memory network (LSTM) to extract the features of
than that of any single model [62]. The suggested approach is used to future information and previous data independently. Using a real-world
gain the benefits of both modules by utilizing the hidden characteristics dataset, the suggested model was compared to a number of benchmark
of CNN and LSTM networks. By examining the electrical load forecasts of models, including the ARIMA, MR, SVR, BPNN, and LSTM, in order to

21
M. Hasan et al. Energy Conversion and Management: X 26 (2025) 100922

Table 11 Table 11 (continued )

A short comparison of various modern LF techniques summarized by the Forecasting Merits Demerits Mitigations
authors. methods
Forecasting Merits Demerits Mitigations prediction goals
methods with various
Statistical • Simple and • Less dependability • By including data sets.
methods requiring little for big and non- meteorological ANN • No further • Black box type • By using the
computing linear data in the factors in the knowledge is nature. right kernel
power. energy dataset. training dataset, required for • Computationally function, one
• Use to ascertain • Difficult to control performance statistical costly. may attain a
how expected complicated may be further training. • Possess a high level of
power and meteorological enhanced. • Capacity to propensity to accuracy in
meteorological circumstances • It is necessary to watch how overfit. monthly load
characteristics that are present. generate independent and • Building an forecasting.
are related. extremely dependent empirically based • For long-term
precise weather variables model. power load fore-
parameter interact. • Large numbers of casting, a com-
forecasts on a • Capable of datasets are bined fore-
continual basis. identifying required to train casting model
Machine • Easy to use and • Reduced • Using every the model, and it can offer a more
learning capable of dependability in interpretability association is sophisticated. objective load
handling big situations tools or between the • Neural network forecasting
datasets. involving big and techniques, like predictor and data technique.
diverse data sets, model variables. preprocessing are
leading to point visualization. • A large number required.
prediction. • Overfitting can of training
be avoided by algorithms are
using available.
regularization • Able to regulate
approaches non-linear inter-
Deep • Effective with • Overfitting, point • Methods like action in load
learning big, non-linear prediction, and knowledge consumption
datasets. appropriate distillation and through weight
• Display hyper-parameter quantization can changes during
improved tweaking are lower training.
multilayer required. computational Hybrid • Incredibly • Specific • Capable of
neural network • The quality of the complexity methods precise and applications, creating more
results. input data and the expandable. • Incredibly costly complex
• No need for architecture • By preserving to compute. forecasting
additional design affect the benefits of • They struggle with models for
neural networks performance. the strategy, consistent conventional
or pre- integrate prediction ANNs with
processing of the various model because of their several
data. weights to intricate learning forecasting
Probabilistic • Efficient, • Highly expensive • Accuracy and enhance model system. outputs
deep reliable, and due to reliability of the performance. • May show low
learning provide computational training data can efficiency, an
prediction arrangement. be accomplished underfitting issue,
intervals. by performing and a lengthy
extensive data training period.
cleansing,
outlier
identification, confirm its efficacy. According to the comparative results, the suggested
and bias model fared better in terms of MAPE. Here, MAPE values for LSTM,
reduction BPNN, and LSTM-BPNN models showed 3.55 %, 4.09 %, and 1.47 %
techniques.
Time series • Reliable in • Numerical • Assessing
respectively [88]. This work proposed a hybrid CNN-BiLSTM Bayesian
analysis adopting various instability. medium- to long- optimization-based smart grid real-time load prediction model that
seasonal effects. range electric improves the model’s accuracy and usefulness while successfully
power forecasts resolving the gradient explosion and disappearance issues rather than
by including
LSTM or CNN techniques [100].
meteorological
data to account It can be concluded that ML based technique used for load fore-
for seasonal casting produces more precise, reliable and accurate predictions than
variations. statistical approaches. Studies showed that combined prediction model
ARIMA • Appropriate for • Linearization of • Utilizing distinct or hybrid model presented better results. Although some single predic-
non-linear data is required. linkages based
models. • Suitable for data on various
tion models such as ANN, CNN, DNN etc. showed better performance
• Calculate the fit that is stationary. forecasting and better accuracy than other models.
of a regression • Preprocessing of goals.
model via data is necessary. • Swapping out 7. Usefulness of this study regarding the fulfillment of
forecasting. • Loss of the current
sustainable development goals
information can techniques with
happen. alternative
• Can not provide approaches that The Sustainable Development Goals (SDGs), sometimes known as the
precise result for can accomplish Global Goals, were adopted by the United Nations in 2015 with the
time series data. the same intention of eradicating poverty, protecting the environment, and
ensuring that everyone lives in a prosperous and peaceful world by

22
M. Hasan et al. Energy Conversion and Management: X 26 (2025) 100922

2030. The interdependence of social justice, the environment, and 9, and 13 are related to the demand forecasting system. Three of the
economic sustainability is acknowledged in the 17 Sustainable Devel- SDG’s main goals will be achieved if both the installation capacity and
opment Goals, and development must strike a balance in these three the maximum demand are predicted properly.
domains. The SDGs’ 17 goals, shown in Fig. 15, serve as a guide for
achieving a more promising and sustainable society for all. 8. Discussions and future research trends
SDGs are broken down into four categories to examine the progress
made toward each type of goal: environmental protection (SDG Maintaining a balance between power generation and demand is
12,13,15), social development (SDG 4,5,10, the SDG 16,17), economic essential to ensure a steady supply of energy, since its relevance in
development (SDG 8,9,11), and essential needs (SDG 1,2,3,6,7) [116]. everyday life has expanded due to its increased use across several in-
Among the 17 goals, goal 7 is related to energy. This goal aims to dustries. The power system is become increasingly unpredictable on
guarantee that everyone has access to inexpensive, clean energy, which both sides as demand response is progressively used on the load side and
is essential for the growth of education, business, agriculture, commu- renewable penetration rates increase on the generating side. This makes
nications, and healthcare. Using renewable energy sources to produce it harder to maintain the system’s safer and more economical operation.
electricity has been proven to be the best way to meet goal 7, as these An extensive examination of the load forecasting (LF) techniques and
sources are less expensive and help to reduce environmental pollution their many applications is the goal of this work. Although several studies
and global warming. Research indicated that attaining sustainable and surveys have looked into LF techniques from different angles, this
development and progress requires implementing efficient indicators study aims to provide a more thorough and current assessment, bringing
and providing equal access to electrical energy [117]. fresh insight to the subject. The features and accuracy of several load
In off-grid communities especially isolated microgrids with several forecasting methods, such as Very Short-Term Load Forecasting
Distributed Energy Resources (DERs) will help in achieving SDG 7. (VSTLF), Short-Term Load Forecasting (STLF), Medium-Term Load
However, because development aid resources are limited, extremely Forecasting (MTLF), and Long-Term Load Forecasting (LTLF), have been
precise designs that reduce mini-grid operating and investment costs are thoroughly examined in this work.
necessary. This can only be accomplished by carefully analyzing and Moreover, the performance evaluation metrics such as Root Mean
forecasting the community’s future energy needs [119]. Appropriate Squared Error (RMSE), Coefficient of determination (R2), Mean Square
energy management and effective grid operation are essential for Error (MSE), Mean Absolute Error (MAE), and Mean Absolute Percent-
achieving energy sustainability. The implementation of energy-efficient age Error (MAPE) were discussed to ensure the better accuracy of
measures lowers net costs; conversely, their non-implementation would different LF models. Traditional statistical models and models based on
lead to increased demand for hydrogen, low-carbon power, and carbon artificial intelligence (AI) are the two primary categories into which
capture and storage. Characterizing and analyzing the data from smart forecasting models are often divided. While statistical techniques
meters after it has been processed and stored can lead to a number of include Simple Moving Average (SMA), Exponential Smoothing, Multi-
advantages in terms of technology, economy, society, and the environ- ple Regression, and time series models like ARMA, ARIMA, and ARI-
ment [120]. MAX, AI-based models include neural networks, machine learning, and
Energy accessibility is crucial for economic growth (SDG 8) since it genetic algorithms.
affects industry, innovation, and infrastructure development. The lack of In short-term load forecasting, ARIMAX is especially helpful for
energy in rural areas is a major obstacle to achieving SDG 7. In order to including factors like weather and time of day. Additionally, this study
achieve the SDG 7, proper management calls for multifaceted scenario has concentrated on intelligent optimization algorithms including ANN,
planning that takes installed electricity generation capacity into ac- CNN, DNN, and others, as well as hybrid load forecasting models (CNN-
count. To plan for capacity, the maximum demand must be estimated. LSTM, CNN-GRU, and LSTM-Transformer, etc.) that have demonstrated
Socioeconomic covariates, in particular the connections between de- superior accuracy and performance over previous models. Finally, the
mand, income, and population growth, should be examined for long- significance of Sustainable Development Goals (SDGs) is also empha-
term capacity planning [121]. The creation of precise energy demand sized in the report, especially SDGs 7, 8, 9, and 13, which significantly
prediction models is necessary to achieve a number of sustainable rely on precise estimates of installation capacity and maximum demand.
development goals, including infrastructure and innovation (SDG 9), In future, to improve the accuracy of forecasting, the following ap-
clean and affordable energy, and climate action (SDG 13). proaches might be considered.
Energy demand prediction models have the potential to reduce the
negative consequences of climate change and advance a more environ- (i) Creation of novel techniques for data pre-processing: Power and
mentally friendly and sustainable energy system. Therefore, SDGs 7, 8, wind data are unclear and have not been thoroughly examined

Fig. 15. Sustainable Development Goals (SDGs) for attaining social, economic, and environmental sustainability [118].

23
M. Hasan et al. Energy Conversion and Management: X 26 (2025) 100922

and analyzed. It implies that current data-based analyses are These recommendations aim to provide a clear roadmap for
ineffective for power and wind data, which lead to skewed find- advancing the accuracy, efficiency, and applicability of load forecasting
ings. Therefore, in order to address complicated uncertainty techniques, ultimately contributing to the development of more resilient
problems and reduce uncertainty, new data pre-processing tech- and sustainable power systems.
niques must be developed. This issue can be resolved by List of acronyms.
combining different pre-processing techniques. ANN–Artificial neural network
(ii) Increase in the number of influencing features: A variety of causes CNN–Convolutional neural network
might cause fluctuations in data. Therefore, the changes in the DNN- Deep neural networks
datasets may be accurately described by using more topograph- LF–Load forecasting
ical and climatic characteristics at the study site. By installing LSTM–Long short-term memory
LTLF–Long-term load forecasting
current, appropriate sensors and collecting the data, the issue can MAE–Mean absolute error
be resolved. MAPE–Mean absolute percentage error
(iii) Enhanced capacity for non-linear fitting: Generally, the chal- ML–Machine learning
lenging topography and climate variations make it challenging to MTLF–Mid-term load forecasting
RMSE–Root mean squared error
assess and characterize data fluctuations.
RNN- Recurrent neural network
(iv) However, the non-linear fitting tendency of the various models is STLF–Short-term load forecasting
superior, and their limited capacity to learn and the oscillations SDGs- Sustainable development goals
in the training data limit their output.Thus, in order to improve SGs–Smart grids
the non-linear fitting capability, it is necessary to mix multiple SVR–Support vector regression
VSTLF–Very short-term load forecasting
models.
(v) Integration of wind, solar, and load forecasting: Future research is
required to develop new techniques for integrated wind, solar,
and electrical load forecasting. Predicting wind/solar power or
CRediT authorship contribution statement
wind speed/irradiance initially could be a good approach. Then,
using the anticipated wind/solar statistics as partial inputs, one
Mahmudul Hasan: Writing – original draft, Visualization, Valida-
could forecast the load.
tion, Methodology, Formal analysis, Resources. Zannatul Mifta:
Writing – original draft, Resources, Methodology, Investigation, Formal
All things considered, it is critical that future research stay focused
analysis, Data curation. Sumaiya Janefar Papiya: Writing – original
on creating more precise and dependable LF models for traditional as
draft, Validation, Methodology, Investigation, Formal analysis, Visual-
well as smart grids, since these are essential to the functional and effi-
ization. Paromita Roy: Writing – original draft, Visualization, Valida-
cient running of the future power grid.
tion, Resources, Methodology, Formal analysis. Pronay Dey:
Validation, Methodology, Investigation, Formal analysis, Resources,
9. Conclusions
Visualization. Nafisa Atia Salsabil: Writing – original draft, Visualiza-
tion, Resources, Methodology, Investigation, Validation. Nahid-Ur-
This study emphasizes the critical role of precise power load fore-
Rahman Chowdhury: Writing – review & editing, Project administra-
casting in the development of modern power networks and highlights
tion, Conceptualization, Resources, Visualization, Writing – original
how advancements in data processing and computational techniques
draft. Omar Farrok: Writing – review & editing, Validation, Supervi-
have significantly enhanced predictive capabilities. The review pre-
sion, Project administration, Resources.
sented in this paper provides a comprehensive classification and analysis
of the most recent forecasting models, offering a synthesized overview of
Declaration of competing interest
their applications and practical implementations. The findings demon-
strate that advancements in artificial intelligence technology, particu-
The authors declare that they have no known competing financial
larly in machine learning and deep learning algorithms, have
interests or personal relationships that could have appeared to influence
substantially improved the accuracy of demand forecasting. Among the
the work reported in this paper.
various techniques evaluated, artificial intelligence-based methods,
such as machine learning and neural network models, consistently
Data availability
achieved superior forecast performance. Based on the analysis, several
key recommendations for advancing the field and guiding future
Data will be made available on request.
research are proposed:

References
▪ The integration of load forecasting techniques with smart grid
systems represents a pivotal research direction. Machine [1] Zhang H, et al. Research on medium- and long-term electricity demand
learning models hold significant potential to enhance fore- forecasting under climate change. Energy Rep 2022;8:1585–600. https://doi.org/
casting accuracy and operational efficiency in smart grid 10.1016/j.egyr.2022.02.210.
[2] M. Hasan, N. A. Salsabil, I. J. Khan, and N. Chowdhury, “Feasibility of a 50 kW
applications.
Solar Photovoltaic System for Easy-Bike Charging in Bangladesh”.
▪ Incorporating distributed energy resources and renewable en- [3] Hasan M, et al. A critical review on control mechanisms , supporting measures ,
ergy sources into forecasting models is crucial for developing and monitoring systems of microgrids considering large scale integration of
renewable energy sources. Energy Rep 2023;10(November):4582–603. https://
more comprehensive and sustainable approaches to power load
doi.org/10.1016/j.egyr.2023.11.025.
forecasting within smart grid systems. [4] Mesa-Jiménez JJ, Tzianoumis AL, Stokes L, Yang Q, Livina VN. Long-term wind
▪ Combining machine learning techniques with statistical and solar energy generation forecasts, and optimisation of Power Purchase
modeling methods may further enhance the precision and ef- Agreements. Energy Rep 2023;9:292–302. https://doi.org/10.1016/j.
egyr.2022.11.175.
ficiency of forecasting methodologies. Utilizing real-time data [5] Wang H, Alattas KA, Mohammadzadeh A, Sabzalian MH, Aly AA, Mosavi A.
streams and implementing advanced sensor technologies could Comprehensive review of load forecasting with emphasis on intelligent
significantly improve the ability to predict and respond effec- computing approaches. Energy Rep 2022;8:13189–98. https://doi.org/10.1016/j.
egyr.2022.10.016.
tively to variations in load demand. [6] Hanifi S, Liu X, Lin Z, Lotfian S. A critical review of wind power forecasting.
Energies 2020;13(15):1–24.

24
M. Hasan et al. Energy Conversion and Management: X 26 (2025) 100922

[7] Zhu J, Dong H, Zheng W, Li S, Huang Y, Xi L. Review and prospect of data-driven [35] Li Y, Han D, Yan Z. Long-term system load forecasting based on data-driven linear
techniques for load forecasting in integrated energy systems. Appl Energy 2022; clustering method. J Mod Power Syst Clean Energy 2018;6(2):306–16. https://doi.
vol. 321, no. June:119269. https://doi.org/10.1016/j.apenergy.2022.119269. org/10.1007/s40565-017-0288-x.
[8] Nafil A, Bouzi M, Anoune K, Ettalabi N. Comparative study of forecasting [36] Haben S, Arora S, Giasemidis G, Voss M, Vukadinović Greetham D. Review of low
methods for energy demand in Morocco. Energy Rep 2020;6:523–36. https://doi. voltage load forecasting: methods, applications, and recommendations. Appl
org/10.1016/j.egyr.2020.09.030. Energy 2021;304, no. June:117798. https://doi.org/10.1016/j.
[9] Li Z, Qin Y, Hou S, Zhang R, Sun H. Renewable energy system based on IFOA-BP apenergy.2021.117798.
neural network load forecast. Energy Rep 2020;6:1585–90. https://doi.org/ [37] Zsiborács H, Pintér G, Vincze A, Birkner Z, Baranyai NH. Grid balancing
10.1016/j.egyr.2020.10.048. challenges illustrated by two European examples: interactions of electric grids,
[10] Hu X, Tang X, Zhang Q, Chen Z, Zhang Y. Combined machine learning forecasting photovoltaic power generation, energy storage and power generation forecasting.
method for short-term power load based on the dynamic weight adjustment. Energy Rep 2021;7:3805–18. https://doi.org/10.1016/j.egyr.2021.06.007.
Energy Rep 2023;9:866–73. https://doi.org/10.1016/j.egyr.2023.04.082. [38] Plaum F, Ahmadiahangar R, Rosin A, Kilter J. Aggregated demand-side energy
[11] H. Daneshi, M. Shahidehpour, and A. L. Choobbari, “Long-term load forecasting flexibility: a comprehensive review on characterization, forecasting and market
in electricity market,” 2008 IEEE Int. Conf. Electro/Information Technol. IEEE EIT prospects. Energy Rep 2022;8:9344–62. https://doi.org/10.1016/j.
2008 Conf., pp. 395–400, 2008, doi: 10.1109/EIT.2008.4554335. egyr.2022.07.038.
[12] Liu Z, Liu Z. Load characteristics forecasting of Hubei power grid up to year 2030 [39] S. Zan and Q. Zhang, “Short-Term Power Load Forecasting Based on an EPT-VMD-
with the development of electric vehicles. Energy Rep 2022;8:259–68. https://doi. TCN-TPA Model,” Appl. Sci., vol. 13, no. 7, 2023, doi: 10.3390/app13074462.
org/10.1016/j.egyr.2022.02.104. [40] He H, Lu Z, Zhang C, Wang Y, Guo W, Zhao S. A data-driven method for dynamic
[13] M. Hasan, I. J. Khan, P. Dey, and N. Chowdhury, “Analysis of Prospective Wind load forecasting of scraper conveyer based on rough set and multilayered self-
Energy Sites in Bangladesh Utilizing Weibull Distribution Method”. normalizing gated recurrent network. Energy Rep 2021;7:1352–62. https://doi.
[14] Jiang Y, et al. Medium-long term load forecasting method considering industry org/10.1016/j.egyr.2021.09.127.
correlation for power management. Energy Rep 2021;7:1231–8. https://doi.org/ [41] Samuel IA, Emmanuel A, Odigwe IA, Felly-Njoku FC. A comparative study of
10.1016/j.egyr.2021.09.140. regression analysis and artificial neural network methods for medium-term load
[15] Hasan M, Sultana SM, Ferdous MJ, Khan IJ, Nayan MF. “Absorber Layer forecasting. Indian J Sci Technol, 2017;10(10):1–7. https://doi.org/10.17485/ijst/
Thickness Dependent Performance Evaluation of Perovskite Solar Cell for 2017/v10i10/86243.
different Electron Transport Layers,” 3rd Int. Conf Electr Comput Commun Eng [42] Zhang B, Zhao X, Dou Z, Liu L. A new medium and long-term power load
ECCE 2023;2023. https://doi.org/10.1109/ECCE57851.2023.10101620. forecasting method considering policy factors. IEEE Access 2021;9:160021–34.
[16] Wu Z, Mu Y, Deng S, Li Y. Spatial–temporal short-term load forecasting https://doi.org/10.1109/ACCESS.2021.3131237.
framework via K-shape time series clustering method and graph convolutional [43] S. H. Rafi, Short term load forecasting technique based on integration of
networks. Energy Rep 2022;8:8752–66. https://doi.org/10.1016/j. convolutional neural network and long-short-term memory network, ’’Electrical ,
egyr.2022.06.122. Electronic and Communication Department’’ 2020.
[17] M. Hasan, P. Dey, I. J. Khan, and N. Chowdhury, “Assessment of Promising Wind [44] Wang H, Zhang N, Du E, Yan J, Han S, Liu Y. A comprehensive review for wind,
Energy Production Sites in Bangladesh using Energy Pattern Factor Method,” pp. solar, and electrical load forecasting methods. Glob Energy Interconnect, 2022;5
2019–2020, 2022. (1):9–30. https://doi.org/10.1016/j.gloei.2022.04.002.
[18] Chen Y, Chen Z. Short-term load forecasting for multiple buildings: A length [45] Matrenin P, Safaraliev M, Dmitriev S, Kokin S, Ghulomzoda A, Mitrofanov S.
sensitivity-based approach. Energy Rep 2022;8:14274–88. https://doi.org/ Medium-term load forecasting in isolated power systems based on ensemble
10.1016/j.egyr.2022.10.425. machine learning models. Energy Rep 2022;8(April):612–8. https://doi.org/
[19] F. YAPRAKDAL and F. BAL, “Comparison of robust machine-learning and deep- 10.1016/j.egyr.2021.11.175.
learning models for midterm electrical load forecasting,” Eur J Tech no February, [46] Walther J, Spanier D, Panten N, Abele E. Very short-term load forecasting on
2022, doi: 10.36222/ejt.1201977. factory level - A machine learning approach. Procedia CIRP 2019;80:705–10.
[20] I. K. Nti, M. Teimeh, O. Nyarko-Boateng, and A. F. Adekoya, “Electricity load https://doi.org/10.1016/j.procir.2019.01.060.
forecasting: a systematic review,” J Elect Syst Informat Technol vol. 7, no. 1, 2020, [47] Z. Ma et al., “Very Short-Term Renewable Energy Power Prediction Using
doi: 10.1186/s43067-020-00021-8. XGBoost Optimized by TPE Algorithm,” 2020 4th Int. Conf. HVDC, HVDC 2020,
[21] F. M. Butt et al., “Intelligence based Accurate Medium and Long Term Load pp. 1236–1241, 2020, doi: 10.1109/HVDC50696.2020.9292870.
Forecasting System,” Appl Artif Intell 36, (1), 2022, doi: 10.1080/ [48] Munkhammar J, van der Meer D, Widén J. Very short term load forecasting of
08839514.2022.2088452. residential electricity consumption using the Markov-chain mixture distribution
[22] F. Yaprakdal, “An Ensemble Deep-Learning-Based Model for Hour-Ahead Load (MCM) model. Appl Energy 2021;vol. 282, no. PA:116180. https://doi.org/
Forecasting with a Feature Selection Approach: A Comparative Study with State- 10.1016/j.apenergy.2020.116180.
of-the-Art Methods,” Energies, vol. 16, no. 1, 2023, doi: 10.3390/en16010057. [49] R. Gonzalez, S. Ahmed, M. Alamaniotis, “Implementing Very-Short-Term
[23] Manandhar P, Rafiq H, Rodriguez-Ubinas E. Current status, challenges, and Forecasting of Residential Load Demand Using a Deep Neural Network
prospects of data-driven urban energy modeling: a review of machine learning Architecture,” Energies, vol. 16, no. 9, 2023, doi: 10.3390/en16093636.
methods. Energy Rep 2023;9:2757–76. https://doi.org/10.1016/j. [50] H. Joshi, “Multiplicative Neuron Models for Very Short-Term Load Forecasting,”
egyr.2023.01.094. pp. 1–11, 2023.
[24] P. Koponen, J. Ikäheimo, J. Koskela, C. Brester, H. Niska, “Assessing and [51] Walther J, Spanier D, Panten N, Abele E. Very short-term load forecasting on
comparing short term load forecasting performance,” Energies, vol. 13, no. 8, factory level - A machine learning approach. Procedia CIRP 2019;80(March):
2020, doi: 10.3390/en13082054. 705–10. https://doi.org/10.1016/j.procir.2019.01.060.
[25] Sweet J. Smart grids. Earth 2010;55(5):6. https://doi.org/10.53478/tuba.978- [52] Shahare K, Mitra A, Naware D, Keshri R, Suryawanshi HM. Performance analysis
625-8352-00-9.ch12. and comparison of various techniques for short-term load forecasting. Energy Rep
[26] Ahmad N, Ghadi Y, Adnan M, Ali M. Load forecasting techniques for power 2023;9:799–808. https://doi.org/10.1016/j.egyr.2022.11.086.
system: research challenges and survey. IEEE Access 2022;10(June):71054–90. [53] Zhao Z, Tang J, Liu J, Ge G, Xiong B, Li Y. Short-term microgrid load probability
https://doi.org/10.1109/ACCESS.2022.3187839. density forecasting method based on k-means-deep learning quantile regression.
[27] Manandhar P, Rafiq H, Rodriguez-Ubinas E. Current status, challenges, and Energy Rep 2022;8:1386–97. https://doi.org/10.1016/j.egyr.2022.03.117.
prospects of data-driven urban energy modeling: a review of machine learning [54] Wu X, Wang Y, Bai Y, Zhu Z, Xia A. Online short-term load forecasting methods
methods. Energy Rep 2023;9(February):2757–76. https://doi.org/10.1016/j. using hybrids of single multiplicative neuron model, particle swarm optimization
egyr.2023.01.094. variants and nonlinear filters. Energy Rep 2021;7:683–92. https://doi.org/
[28] Wang M, Yu Z, Chen Y, Yang X, Zhou J. Short-term load forecasting considering 10.1016/j.egyr.2021.01.030.
improved cumulative effect of hourly temperature. Electr Pow Syst Res 2022;205: [55] Hong Y, Zhou Y, Li Q, Xu W, Zheng X. A deep learning method for short-term
107746. https://doi.org/10.1016/j.epsr.2021.107746. residential load forecasting in smart grid. IEEE Access 2020;8:55785–97. https://
[29] Hammad MA, Jereb B, Rosi B, Dragan D. Methods and models for electric load doi.org/10.1109/ACCESS.2020.2981817.
forecasting: a comprehensive review. Logist Sustain Transport 2020;11(1):51–76. [56] Wu K, Peng X, Chen Z, Su H, Quan H, Liu H. A novel short-term household load
https://doi.org/10.2478/jlst-2020-0004. forecasting method combined BiLSTM with trend feature extraction. Energy Rep
[30] S. Akhtar et al., “Progress , and the Road Ahead,” 2023. 2023;9:1013–22. https://doi.org/10.1016/j.egyr.2023.05.041.
[31] E. Mele, “A Review of Machine Learning Algorithms Used for Load Forecasting at [57] B. Ibrahim, L. Rabelo, E. Gutierrez-Franco, and N. Clavijo-Buritica, “Machine
Microgrid Level,” no. October, pp. 452–458, 2019, doi: 10.15308/sinteza-2019- Learning for Short-Term Load Forecasting in Smart Grids,” Energies, vol. 15, no.
452-458. 21, 2022, doi: 10.3390/en15218079.
[32] Al Mamun A, Sohel M, Mohammad N, Haque Sunny MS, Dipta DR, Hossain E. [58] M. Li, H. Tian, Q. Chen, M. Zhou, and G. Li, “A hybrid prediction method for
A comprehensive review of the load forecasting techniques using single and short-term load based on temporal convolutional networks and attentional
hybrid predictive models. IEEE Access 2020;8:134911–39. https://doi.org/ mechanisms,” IET Gener. Transm. Distrib., no. January, pp. 1–14, 2023, doi:
10.1109/ACCESS.2020.3010702. 10.1049/gtd2.12798.
[33] A. A. Mir, M. Alghassab, K. Ullah, Z. A. Khan, Y. Lu, M. Imran, “A review of [59] Yi S, Liu H, Chen T, Zhang J, Fan Y. A deep LSTM-CNN based on self-attention
electricity demand forecasting in low and middle income countries: The demand mechanism with input data reduction for short-term load forecasting. IET Gener
determinants and horizons,” Sustain., 12, (15), 2020, doi: 10.3390/SU12155931. Transm Distrib, 2023;17(7):1538–52. https://doi.org/10.1049/gtd2.12763.
[34] D. Upadhaya, R. Thakur, N. K. Singh, “A systematic review on the methods of [60] Q. Cheng et al., “Short-Term Load Forecasting Method for Low Voltage Users
short term load forecasting,” 2019 2nd Int. Conf. Power Energy Environ. Intell. Based on Deep Belief Neural Network,” iSPEC 2020 - Proc. IEEE Sustain. Power
Control. PEEIC 2019, no. September, pp. 6–11, 2019, doi: 10.1109/ Energy Conf. Energy Transit. Energy Internet, pp. 2564–2569, 2020, doi: 10.1109/
PEEIC47157.2019.8976518. iSPEC50848.2020.9351039.

25
M. Hasan et al. Energy Conversion and Management: X 26 (2025) 100922

[61] X. Zhang, Z. Zeng, P. Wang, J. Song, and Z. Kong, “A hybrid edge-cloud Electron Control, 2020;18(4):2112–7. https://doi.org/10.12928/TELKOMNIKA.
computing method for short-term electric load forecasting based on smart V18I4.14032.
metering terminal,” 2020 IEEE 4th Conf. Energy Internet Energy Syst. Integr. [87] Wang Y, Wang Z, Li X, Wang Z, Liu R, Tong X. Medium and Long-term Load
Connect. Grids Towar. a Low-Carbon High-Efficiency Energy Syst. EI2 2020, pp. Intelligent Forecasting Method Based on the Comprehensive and Multiple Early
3101–3105, 2020, doi: 10.1109/EI250167.2020.9346774. Warning Indicators. J Phys Conf Ser 2020;1449(1):9. https://doi.org/10.1088/
[62] Xuan Y, et al. Multi-Model Fusion Short-Term Load Forecasting Based on Random 1742-6596/1449/1/012124.
Forest Feature Selection and Hybrid Neural Network. IEEE Access 2021;9: [88] B. Jin et al., “Hybrid LSTM–BPNN-to-BPNN Model Considering Multi-Source
69002–9. https://doi.org/10.1109/ACCESS.2021.3051337. Information for Forecasting Medium- and Long-Term Electricity Peak Load,”
[63] Shilpa NG, Sheshadri GS. Electrical Load Forecasting Using Time Series. Analysis Energies, vol. 15, no. 20, 2022, doi: 10.3390/en15207584.
Proc. B-HTC 2020–1st IEEE Bangalore Humanit. Technol. Conf. 2020. https://doi. [89] A. Setiawan, Z. Arifin, B. Sudiarto, F. H. Jufri, Q. Haramaini, and I. Garniwa,
org/10.1109/B-HTC50970.2020.9297986. “Comparison of Medium-Term Load Forecasting Methods (Splitted Linear
[64] Son J, Cha J, Kim H, Wi YM. Day-ahead short-term load forecasting for holidays Regression and Artificial Neural Networks) in Electricity Systems Located in
based on modification of similar days’ load profiles. IEEE Access 2022;10: Tropical Regions,” 2022 3rd Int. Conf. Clean Green Energy Eng. CGEE 2022, no.
17864–80. https://doi.org/10.1109/ACCESS.2022.3150344. August, pp. 84–88, 2022, doi: 10.1109/CGEE55282.2022.9976521.
[65] Zeng P, Jin M, Fazla Elahe MD. Short-term power load forecasting based on cross [90] Inteha A, Nahid-Al-Masood FH, Khan IA. A data driven approach for day ahead
multi-model and second decision mechanism. IEEE Access 2020;8:184061–72. short term load forecasting. IEEE Access 2022;10(August):84227–43. https://doi.
https://doi.org/10.1109/ACCESS.2020.3028649. org/10.1109/ACCESS.2022.3197609.
[66] Rafi SH, Al-Masood N, Deeba SR, Hossain E. A short-term load forecasting method [91] Matrenin P, Safaraliev M, Dmitriev S, Kokin S, Ghulomzoda A, Mitrofanov S.
using integrated CNN and LSTM network. IEEE Access 2021;9:32436–48. https:// Medium-term load forecasting in isolated power systems based on ensemble
doi.org/10.1109/ACCESS.2021.3060654. machine learning models. Energy Rep 2022;8:612–8. https://doi.org/10.1016/j.
[67] Z. Jiang et al., “A Bottom-up Method for Probabilistic Short-Term Load egyr.2021.11.175.
Forecasting Based on Medium Voltage Load Patterns,” IEEE Access, vol. 9, no. Mv, [92] R. K. Agrawal, F. Muchahary, M. M. Tripathi, “Long term load forecasting with
pp. 76551–76563, 2021, doi: 10.1109/ACCESS.2021.3082926. hourly predictions based on long-short-term-memory networks,” 2018 IEEE Texas
[68] N. Kuantayev, B. A. Bainiyazov, and A. B. Uakhitova, “Short-Term Electrical Load Power Energy Conf. TPEC 2018, vol. 2018-Febru, no. February, pp. 1–6, 2018, doi:
Forecasting Based on Neural Networks,” IOP Conf. Ser. Earth Environ. Sci., vol. 10.1109/TPEC.2018.8312088.
1154, no. 1, 2023, doi: 10.1088/1755-1315/1154/1/012014. [93] H. Sangrody, N. Zhou, S. Tutun, B. Khorramdel, M. Motalleb, M. Sarailoo, “Long
[69] Zhang G, Wei C, Jing C, Wang Y. Short-term electrical load forecasting based on term forecasting using machine learning methods,” 2018 IEEE Power Energy Conf.
time augmented transformer. Int J Comput Intell Syst, 2022;15(1):1–11. https:// Illinois, PECI 2018, vol. 2018-Janua, no. February, pp. 1–5, 2018, doi: 10.1109/
doi.org/10.1007/s44196-022-00128-y. PECI.2018.8334980.
[70] L. Li, L. Guo, J. Wang, H. Peng, “Short-Term Load Forecasting Based on Spiking [94] Essallah S, Khedher A. A comparative study of long-term load forecasting
Neural P Systems,” Appl. Sci., vol. 13, no. 2, 2023, doi: 10.3390/app13020792. techniques applied to Tunisian grid case. Electr Eng, 2019;101(4):1235–47.
[71] Tang Y, Cai H. Short-term power load forecasting based on VMD-pyraformer- https://doi.org/10.1007/s00202-019-00859-w.
adan. IEEE Access 2023;11(1):61958–67. https://doi.org/10.1109/ [95] Nalcaci G, Özmen A, Weber GW. Long-term load forecasting: models based on
ACCESS.2023.3273596. MARS, ANN and LR methods. Cent Eur J Oper Res, 2019;27(4):1033–49. https://
[72] Li S, Wu H, Wang X, Xu B, Yang L, Bi R. Short-term load forecasting based on AM- doi.org/10.1007/s10100-018-0531-1.
CIF-LSTM method adopting transfer learning. Front Energy Res, 2023;11(March): [96] N. Son, S. Yang, J. Na, “Deep neural network and long short-term memory for
1–10. https://doi.org/10.3389/fenrg.2023.1162040. electric power load forecasting,” Appl. Sci., vol. 10, no. 18, 2020, doi: 10.3390/
[73] Kien DT, Huong PD, Minh ND. Application of Sarima Model in Load Forecasting APP10186489.
in Hanoi City. Int J Energy Econ Policy 2023;13(3):164–70. https://doi.org/ [97] Zhang D, Guan W, Yang J, Yu H, Xiao WC, Yu T. Medium—And Long-Term Load
10.32479/ijeep.14121. Forecasting Method for Group Objects Based on the Image Representation
[74] Zeng W, et al. Ultra short-term power load forecasting based on similar day Learning. Front Energy Res, 2021;9(August):1–15. https://doi.org/10.3389/
clustering and ensemble empirical mode decomposition. Energies 2023;16(4): fenrg.2021.739993.
1–15. https://doi.org/10.3390/en16041989. [98] Jingying Y, et al. Long-term load combination forecasting method considering the
[75] E. Giacomazzi, F. Haag, and K. Hopf, Short-Term Electricity Load Forecasting periodicity and trend of data. E3S Web Conf, 2021;252:4. https://doi.org/
Using the Temporal Fusion Transformer: Effect of Grid Hierarchies and Data 10.1051/e3sconf/202125201057.
Sources, vol. 1, no. 1. Association for Computing Machinery, 2023. doi: 10.1145/ [99] Mohammed NA, Al-Bazi A. An adaptive backpropagation algorithm for long-term
3575813.3597345. electricity load forecasting. Neural Comput Appl, 2022;34(1):477–91. https://doi.
[76] Ding Z, Li T, Li X, Cui Z. Enhanced short-term flow prediction in power org/10.1007/s00521-021-06384-x.
dispatching network using a transfer learning approach with GRU-XGBoost [100] Zhang D, Jin X, Shi P, Chew XY. Real-time load forecasting model for the smart
module ding. Front Energy Res, 2024;12(August):1–11. https://doi.org/10.3389/ grid using bayesian optimized CNN-BiLSTM. Front Energy Res, 2023;11(May):
fenrg.2024.1429746. 1–12. https://doi.org/10.3389/fenrg.2023.1193662.
[77] Ahmad A, Javaid N, Guizani M, Alrajeh N, Khan ZA. An accurate and fast [101] Feng Q, Qian S. Research on power load forecasting model of economic
converging short-term load forecasting model for industrial applications in a development zone based on neural network. Energy Rep 2021;7:1447–52. https://
smart grid. IEEE Trans Ind Informatics 2017;13(5):2587–96. https://doi.org/ doi.org/10.1016/j.egyr.2021.09.098.
10.1109/TII.2016.2638322. [102] Kim W, Han Y, Kim KJ, Song KW. Electricity load forecasting using advanced
[78] Khan IU, Javaid N, Taylor CJ, Gamage KAA, Ma X. “Big data analytics based short feature selection and optimal deep learning model for the variable refrigerant
term load forecasting model for residential buildings in smart grids,” IEEE flow systems. Energy Rep 2020;6:2604–18. https://doi.org/10.1016/j.
INFOCOM 2020 - IEEE Conf. Comput Commun Work INFOCOM WKSHPS 2020; egyr.2020.09.019.
2020:544–9. https://doi.org/10.1109/INFOCOMWKSHPS50562.2020.9163031. [103] Chen H, Zhang H. New energy generation forecasting and dispatching method
[79] J. Jing, H. Di, T. Wang, N. Jiang, and Z. Xiang, “Optimization of power system based on big data. Energy Rep 2021;7:1280–8. https://doi.org/10.1016/j.
load forecasting and scheduling based on artificial neural networks,” 2025. egyr.2021.09.136.
[80] A. Zare-Noghabi, M. Shabanzadeh, and H. Sangrody, “Medium-Term Load [104] Syed D, et al. Deep learning-based short-term load forecasting approach in smart
Forecasting Using Support Vector Regression, Feature Selection, and Symbiotic grid with clustering and consumption pattern recognition. IEEE Access 2021;9:
Organism Search Optimization,” IEEE Power Energy Soc. Gen. Meet., vol. 2019- 54992–5008. https://doi.org/10.1109/ACCESS.2021.3071654.
Augus, no. July, 2019, doi: 10.1109/PESGM40551.2019.8973726. [105] B. Deng, Y. Wen, and P. Yuan, “Hybrid Short-Term Load Forecasting using the
[81] Yasin ZM, Aziz NFA, Salim NA, Wahab NA, Rahmat NA. An accurate medium- Hadoop MapReduce Framework,” IEEE Power Energy Soc. Gen. Meet., vol. 2020-
term load forecasting based on hybrid technique. Indones J Electr Eng Comput Sci, Augus, no. 51707017, pp. 1–5, 2020, doi: 10.1109/PESGM41954.2020.9282094.
2018;12(1):161–7. https://doi.org/10.11591/ijeecs.v12.i1.pp161-167. [106] Li L, Ota K, Dong M. When weather matters: IoT-based electrical load forecasting
[82] Butt FM, Hussain L, Mahmood A, Lone KJ. Artificial Intelligence based accurately for smart grid. IEEE Commun Mag, 2017;55(10):46–51. https://doi.org/10.1109/
load forecasting system to forecast short and medium-term load demands. Math MCOM.2017.1700168.
Biosci Eng, 2021;18(1):400–25. https://doi.org/10.3934/MBE.2021022. [107] Xishuang D, Lijun Q, Lei H. “Short-term load forecasting in smart grid: a
[83] S. Aslam, H. Herodotou, S. M. Mohsin, N. Javaid, N. Ashraf, and S. Aslam, “A combined CNN and K-means clustering approach,” 2017 IEEE Int. Conf Big Data
survey on deep learning methods for power load and renewable energy Smart Comput Bigcomp 2017;2017:119–25. https://doi.org/10.1109/
forecasting in smart microgrids,” Renew. Sustain. Energy Rev., vol. 144, no. August BIGCOMP.2017.7881726.
2020, p. 110992, 2021, doi: 10.1016/j.rser.2021.110992. [108] Quilumba F, Lee W, Huang H. Using smart meter data to improve the accuracy of
[84] Wang S, Wang S, Wang D. Combined probability density model for medium term intraday load forecasting considering customer behavior similarities.
load forecasting based on quantile regression and kernel density estimation. Ieeexploreieeeorg 2014:1–8.
Energy Procedia 2019;158:6446–51. https://doi.org/10.1016/j. [109] Habbak H, Mahmoud M, Metwally K, Fouda MM, Ibrahem MI. Load forecasting
egypro.2019.01.169. techniques and their applications in smart grids. Energies 2023;16(3):1–33.
[85] A. Zare-Noghabi, M. Shabanzadeh, H. Sangrody, “Medium-Term Load Forecasting https://doi.org/10.3390/en16031480.
Using Support Vector Regression, Feature Selection, and Symbiotic Organism [110] Arvanitidis AI, Bargiotas D, Daskalopulu A, Laitsos VM, Tsoukalas LH. Enhanced
Search Optimization,” IEEE Power Energy Soc Gen Meet vol. 2019-Augus, no. short-term load forecasting using artificial neural networks. Energies 2021;14(22):
August, 2019, doi: 10.1109/PESGM40551.2019.8973726. 1–14. https://doi.org/10.3390/en14227788.
[86] Imam HL, Gaya MS, Galadanci GSM. “Medium term load demand forecast of Kano [111] S. Ungureanu, V. Topa, A. Cziker, “Industrial load forecasting using machine
zone using neural network algorithms,” Telkomnika (Telecommunication Comput. learning in the context of smart grid,” 2019 54th Int. Univ. Power Eng. Conf. UPEC
2019 - Proc., pp. 1–6, 2019, doi: 10.1109/UPEC.2019.8893540.

26
M. Hasan et al. Energy Conversion and Management: X 26 (2025) 100922

[112] Duong-Ngoc H, Nguyen-Thanh H, Nguyen-Minh T. “Short term load forcast using [118] Mishra US, Padhi B. Sustaining the sustainable development goals. Econ Polit
deep learning,” 2019 Innov. Power Adv Comput Technol i-PACT 2019;2019:1–5. Wkly, 2021;56(34):30–1.
https://doi.org/10.1109/i-PACT44901.2019.8960036. [119] Herraiz-Cañete Á, Ribó-Pérez D, Bastida-Molina P, Gómez-Navarro T. Forecasting
[113] Alazab M, Khan S, Krishnan SSR, Pham QV, Reddy MPK, Gadekallu TR. energy demand in isolated rural communities: a comparison between
A Multidirectional LSTM Model for Predicting the Stability of a Smart Grid. IEEE deterministic and stochastic approaches. Energy Sustain Dev, 2022;66(December):
Access 2020;8:85454–63. https://doi.org/10.1109/ACCESS.2020.2991067. 101–16. https://doi.org/10.1016/j.esd.2021.11.007.
[114] J. Zheng, C. Xu, Z. Zhang, X. Li, “2023 57th Annual Conference on Information [120] Liaqat R, Sajjad IA, Waseem M, Alhelou HH. Appliance level energy
Sciences and Systems, CISS 2023,” 2023 57th Annu. Conf. Inf. Sci. Syst. CISS 2023, characterization of residential electricity demand: prospects, challenges and
pp. 1–6, 2023. recommendations. IEEE Access 2021;9(October):148676–97. https://doi.org/
[115] Ryu S, Noh J, Kim H. Deep neural network based demand side short term load 10.1109/ACCESS.2021.3123196.
forecasting. Energies 2017;10(1):1–20. https://doi.org/10.3390/en10010003. [121] Marcillo-Delgado JC, Ortego MI, Pérez-Foguet A. A compositional approach for
[116] Xu Y, et al. Forecasting future development under the interactions among modelling SDG7 indicators: case study applied to electricity access. Renew Sustain
sustainable development goals. Sustainability 2023;15(22):15929. https://doi. Energy Rev, 2019;107(March):388–98. https://doi.org/10.1016/j.
org/10.3390/su152215929. rser.2019.03.028.
[117] Armin Razmjoo A, Sumper A, Davarpanah A. Energy sustainability analysis based
on SDGs for developing countries. Energy Sources, Part A Recover Util. Environ Eff,
2020;42(9):1041–56. https://doi.org/10.1080/15567036.2019.1602215.

27