See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/370236849

Machine Learning Based Solar Photovoltaic Power Forecasting: A Review and

Comparison

Article in IEEE Access · January 2023

DOI: 10.1109/ACCESS.2023.3270041

CITATIONS READS

5 352

6 authors, including:

Jwaone Gaboitaolelwe Murtala Zungeru Adamu

Botswana International University of Science and Technology Botswana International University of Science and Technology
12 PUBLICATIONS 49 CITATIONS 127 PUBLICATIONS 1,745 CITATIONS

SEE PROFILE SEE PROFILE

Caspar K Lebekwe
Botswana International University of Science and Technology
29 PUBLICATIONS 210 CITATIONS

SEE PROFILE

All content following this page was uploaded by Murtala Zungeru Adamu on 02 May 2023.

The user has requested enhancement of the downloaded file.

Received 22 March 2023, accepted 20 April 2023, date of publication 24 April 2023, date of current version 28 April 2023.
Digital Object Identifier 10.1109/ACCESS.2023.3270041

Machine Learning Based Solar Photovoltaic

Power Forecasting: A Review and Comparison
JWAONE GABOITAOLELWE1 , (Member, IEEE),
ADAMU MURTALA ZUNGERU 1 , (Senior Member, IEEE),
ABID YAHYA 1 , (Senior Member, IEEE), CASPAR K. LEBEKWE 1, (Member, IEEE),
DASARI NAGA VINOD 2 , (Member, IEEE),
AND AYODEJI OLALEKAN SALAU 3,4 , (Member, IEEE)
1 Department of Electrical, Computer and Telecommunications Engineering, Botswana International University of Science and Technology, Palapye 10071,
Botswana
2 Department of Electronics and Communication Engineering, Vel Tech Rangarajan Dr. Sagunthala R&D Institute of Science and Technology, Avadi, Chennai,

Tamil Nadu 600062, India

3 Department of Electrical/Electronic and Computer Engineering, Afe Babalola University, Ado Ekiti 362101, Nigeria
4 Saveetha School of Engineering, Saveetha Institute of Medical and Technical Sciences, Chennai, Tamil Nadu 602105, India

Corresponding author: Adamu Murtala Zungeru ([email protected])

This work was supported by the Office of Research, Development, and Innovation (ORDI) of the Botswana International University of
Science and Technology under Grant S00294.

ABSTRACT The growing interest in renewable energy and the falling prices of solar panels place solar
electricity in a favourable position for adoption. However, the high-rate adoption of intermittent renewable
energy introduces challenges and the potential to create power instability between the available power
generation and the load demand. Hence, accurate solar Photovoltaic (PV) power forecasting is essential
to maintain system reliability and maximize renewable energy integration. The current solar PV power
forecasting approaches are an essential tool to maintain system reliability and maximize renewable energy
integration. This paper presents a comprehensive and comparative review of existing Machine Learning (ML)
based approaches used in PV power forecasting, focusing on short-term horizons. We provide an overview
of factors affecting solar PV power forecasting and an overview of existing PV power forecasting methods
in the literature, with a specific focus on ML-based models. To further enhance the comparison and provide
more insights into the advancement in the area, we simulate the performance of different ML methods used
in solar PV power forecasting and, finally, a discussion on the results of the work.

INDEX TERMS Artificial intelligence, artificial neural networks, photovoltaic, power forecasting, renew-
able energy.

I. INTRODUCTION have the potential to play a key role in the future. According
Electricity has become a significant part of the modern world. to the Global Energy Review in the year 2020, the share
Its demand is predicted to rise as the world becomes more of renewable electricity supply contributed close to 28%
globalized and modernized [1]. Electricity generation from of global electricity generation in the first quarter of 2020.
fossil fuel-based power plants has resulted in significant An increase from 26% in the first quarter of 2019 [2]. The
pollution in terms of Carbon dioxide (CO2 ) emissions and growth is primarily due to the contributions of solar and wind
greenhouse gas (GHG) emissions, resulting in significant farms, spurred on by the decrease in wind turbines and solar
climate change around the planet. Hence alternative renew- panels cost [3] and also many policies adopted by countries
able energy sources (RESs) such as solar, wind, and biomass to reduce and regulate CO2 emissions and global warming
[4]. Encouraged by low, declining prices and large contri-
The associate editor coordinating the review of this manuscript and butions to carbon emission reduction, the power sector is
approving it for publication was Bidyadhar Subudhi . progressing increasingly into integrating and using renewable

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.

40820 For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ VOLUME 11, 2023
J. Gaboitaolelwe et al.: Machine Learning Based Solar Photovoltaic Power Forecasting: A Review and Comparison

energy sources. The use of RE has brought about momentous (based on physical principles), Data driven (statistical and
changes in the field of energy production and has demon- Machine Learning) or hybrid (Combination of one or more).
strated the potential of clean and unlimited energy for the These techniques use various input variables, such as weather
future. data, historical power generation data, and solar irradiance
There exist several RESs, such as hydropower, bioenergy, measurements, to predict the output of solar PV systems.
wind energy, solar thermal, solar photovoltaic (PV) energy,
and many more. Focusing on solar energy, solar photovoltaic 1) PHYSICAL MODELS (WHITE BOX)
(PV) systems have experienced a large growth in electricity Physical models are based on the fundamental physical prin-
generation, with an increased number of installed systems [5], ciples that govern the behavior of solar PV systems. These
[6], [7], [8]. Such PV systems have been installed in open models take into account the basic physical properties of the
spaces or rooftops, integrated into buildings and vehicles, and solar PV system, such as the panel efficiency, the angle of
placed in large arrays in solar power plants. However, besides incidence of the sun’s rays, and the shading of the panels.
solar PV contributing to diversifying power generation and The models also consider the environmental factors that affect
solving global warming solutions, the high-rate adoption of solar energy production, such as temperature, cloud cover,
solar PV introduces challenges. and humidity. Physical models are typically based on math-
Solar energy is an intermittent Renewable energy source ematical equations that describe the physical behavior of the
(RES). Hence, it varies based on both weather conditions solar PV system. These equations can be quite complex and
and site-specific conditions. The possibility of creating power require detailed knowledge of the underlying physics and
instability increases in tandem with the increase in the num- system characteristics (installation details, electrical config-
ber of renewables in the energy mix [9], [10], increasing the uration, geographic location and technical characteristics) of
importance of controlling the flow of renewable power. the PV system being modelled [11]. Furthermore, since they
Solar PV forecasting is the process of predicting the utilize weather conditions as input for forecasting, this results
amount of electrical power that will be generated by a solar in models performance heavily relying on the accuracy of
photovoltaic (PV) system over a specified time. This involves weather forecast data [12].
the use of mathematical models to predict the output of a PV
system based on a range of input parameters. The ambiguity
2) DATA DRIVEN MODELS
of the load consumption as well as the intermittent power pro-
duction of the solar PV production can result in differences Data driven forecasting models refer to models that rely on
between the available power generation and the demanded the use and analysis of historical data to derive insights,
load. Hence, accurate prediction or forecasting of solar PV patterns, and relationships that can be used to create models
power generation capacity is essential. to predict the output of a solar PV system. These models are
Accurate solar PV forecasting is important for several based on statistics and algorithms that can learn patterns and
reasons. First, it allows grid operators or independent systems relationships from data and make accurate predictions. How-
to anticipate changes in the supply of electricity and to adjust ever, they require large amounts of high-quality data to be
the distribution of power accordingly. This helps to maintain effective, and the accuracy of the predictions can be affected
the stability and reliability of the power grid or independent by biases in the data or incomplete data. Data driven can be
systems isolated from the grid. Second, accurate forecasts can further subdivided into two categories, these being statistical
help to optimize the use of renewable energy sources, reduce and machine learning models. The boundaries between sta-
the need for backup power sources, and ultimately lower the tistical modelling and machine learning are often blurry, and
cost of electricity for consumers. some researchers consider them to be the same depending on
Solar PV forecasting is a challenging problem due to the the specific context and perspective of the user.
complex and dynamic nature of solar energy production.
Factors such as cloud cover, humidity, and temperature can a: STATISTICAL MODELS
all affect the output of a PV system. Statistical models are based on statistical analysis of his-
As a result, accurate solar PV forecasting requires the use torical data to identify trends and patterns that can be used
of sophisticated models and the collection of high-quality to forecast future solar PV power output. These models
data. If a forecast is inaccurate, it can result in an imbalance use statistical techniques and mathematical algorithms to
between the available power generation and the demand, fit historical data and derive relationships between differ-
which can lead to power outages or other disruptions. In addi- ent variables, such as solar irradiance, weather conditions,
tion, an inaccurate forecast can lead to underutilization of the and energy production. There are many different types of
solar PV systems, which can limit the amount of renewable statistical models used in solar PV forecasting, including
energy that is integrated into the power grid and increase the autoregressive models (AR), moving average models (MA),
reliance on fossil fuel-based sources of electricity. and Autoregressive Integrated Moving Average (ARIMA).
To address these challenges, researchers and industry pro- ARIMA models are a class of models that are commonly used
fessionals have developed a range of solar PV forecast- for time series analysis and forecasting. They assume that the
ing techniques, which can be broadly classified as physical current value of a time series depends on its past values and

VOLUME 11, 2023 40821

 J. Gaboitaolelwe et al.: Machine Learning Based Solar Photovoltaic Power Forecasting: A Review and Comparison

the errors in previous predictions. The ARIMA model uses tool to maintain system reliability and maximize renewable
these relationships to generate predictions for future values energy integration. This is intended to be achieved by pre-
of the time series. senting a comprehensive and comparative review on exist-
While statistical models such as ARIMA models do ing Machine Learning (ML) based approaches used in PV
involve the use of algorithms to identify patterns and relation- power generation. Though there are related literature surveys
ships, which is a key aspect of machine learning, they are in regarding solar PV power generation forecasting, most do
a separate category from machine learning techniques since not perform simulation comparisons on short-term forecasts.
the models typically rely on statistical methods such as max- Therefore, the presented work is different in that emphasis is
imum likelihood estimation, rather than the iterative learning made on the comparison of short-term horizons and carries
algorithms used in many machine learning techniques. out a comparative review with simulations.
The focus contributions of the paper are as follows:
b: MACHINE LEARNING MODELS (BLACK BOX) 1. We gave an overview of factors affecting solar PV
Machine learning models, on the other hand, are based on power forecasting regarding temporal classifications
a more complex set of algorithms that allow the model to and weather.
learn patterns and relationships in the data on its own. These 2. We gave a detailed of Artificial Intelligence (AI) and
models can handle more complex and nonlinear relationships ML and their application in ML based Solar PV power
between the input variables and output variables and can forecasting.
often generate more accurate forecasts than statistical models. 3. We provided an overview of the performance metrics
As such, the models have been reported to produce accurate used for solar PV power forecasting.
PV forecasting [13]. Examples of machine learning models 4. We provide a comparison of existing solar PV power
used in solar PV forecasting include artificial neural networks forecasting methods in the literature, with a focus on
(ANN), support vector machines (SVM), and random forests ML-based models and a short-term horizon.
(RF). 5. Finally, we provided an experiment setup and simula-
One key difference between statistical models and machine tion of the performance of different ML methods used
learning models is the level of human input required. in solar PV power forecasting.
Statistical models generally require more human input in The remainder of the paper is arranged as follows: Sec-
terms of selecting and engineering features, designing the tion II presents an overview of factors affecting solar PV
model architecture, and tuning hyperparameters. In contrast, power forecasting regarding temporal classifications and
machine learning models can be trained on raw data with min- weather. Section III presents an overview of Solar PV power
imal preprocessing, and the model architecture and hyper- forecasting models with emphasis on ML based models.
parameters can often be automatically optimized using tech- In Section IV, we provide an overview of the performance
niques such as grid search or Bayesian optimization. metrics used for solar PV power forecasting. Section V gives
Another difference between the two approaches is their a comparison of existing PV power forecasting methods in the
ability to handle uncertainty. Statistical models are often more literature, with a focus on ML-based models and a short-term
transparent and can provide estimates of uncertainty in their horizon. In Section VI, experiment setup is provided, while
forecasts, whereas machine learning models are often more Section VII provides results and discussion on the results.
black-box and may not provide as much information about Finally, Section VIII provides the conclusion and future work.
how the model arrived at its predictions. However, techniques
such as Monte Carlo simulation can be used to generate II. FACTORS AFFECTING SOLAR PV POWER
probabilistic forecasts from machine learning models. FORECASTING
Solar PV forecasting methods face several factors that affect
3) HYBRID MODELS their forecasting performance and accuracy. These factors
Hybrid models are models that combine one or more fore- include forecasting horizon and weather conditions [14],
casting methods from either physical or data driven methods [15].
with the goal of overcoming the limitations of individual tech-
niques and incorporating the advantages of each method. This 1) FORECASTING HORIZONS
does come at a cost in the form of increased computational Several methods are used in literature to perform forecasts
time. of solar PV power generation. The methods can be classified
By using a combination of these techniques and continu- based on temporal aspects, spacial aspects or the data used for
ously improving the accuracy of solar PV forecasting, it is the forecast. It is crucial to introduce three ideas with regard
possible to maintain system reliability, maximize renewable to the temporal component of forecasts, these are forecast
energy integration, and ultimately transition to a more sus- interval (fi), forecast resolution (fr) and forecast horizon (fh).
tainable and carbon-free energy system. The temporal span between the current time (t) and the time
In this context, the scope and objective of this paper is to of the next prediction is known as the forecast horizon. [5].
address accurate solar PV power forecasting as an essential The frequency at which forecasts are given is described by

40822 VOLUME 11, 2023

J. Gaboitaolelwe et al.: Machine Learning Based Solar Photovoltaic Power Forecasting: A Review and Comparison

TABLE 1. Solar PV power forecasting time horizons. However, since atmospheric conditions often differ from the
STC, the actual energy output of the PV system may differ
significantly from the rated value. Hence weather conditions
are a significant determinant of the PV panels’ variance and
stability [26].
Variations in sun irradiation that reach the earth are what
generate the uncertainty and variability in the solar resource,
hence impacting PV output power. These variations can
be split into deterministic and stochastic components. The
the forecast resolution, while the time span of projections is motion of the Earth as it rotates around its axis with reference
indicated by the forecast interval. The forecasting procedures to the sun (the diurnal cycle) and the earth orbiting around
of PV power generation are generally categorized in the the Sun (Annual cycle) provides an explanation for the deter-
literature based on their forecast horizons [16], [17], [18], ministic aspect. Physical equations are capable of adequately
[19]. These are some of the typical forecasting time horizons describing such predictable features of sun irradiation [5].
[16], [17] shown in TABLE 1. The stochastic behavior of the atmosphere causes greater
Different applications fall under different forecasting hori- and unexpected fluctuations in PV output power. These refer
zons. For example Short term forecast is used in day-ahead to the meteorological or weather conditions. Numerous cli-
energy markets and energy management applications such matic factors, including cloud cover, humidity level, wind
as energy storage and load scheduling and energy market speed, temperature and air pressure contribute to this unpre-
participation [20]. Not considering the implications of time dictability. As such weather forecasting and weather fore-
horizon on forecasting accuracy can result in poor forecasting casted data can be used to improve solar PV systems power
performance and hence impact the activities that rely on forecasting. Furthermore, depending on where the PV plant
the information. In [21], a 15-90 min time horizon PVPF is located, different meteorological variables have different
model was developed. By varying the horizon’s length and correlation factors with PV power. More information on the
maintaining all other parameters at constant; the accuracy of analysis of weather on PV power forecasting can be found
the model was found to vary relative to the horizon’s length. in [27].
On the same dataset, similar finding were observed in [22]
where a forecast error was found to increase from 3.2% to III. SOLAR PV POWER FORECASTING MODELS
15.5% for horizons ranging from 20 to 180 seconds. Other Utilities that highly depend on PV reliability need to predict
studies have adopted multi-scale hybrid forecast models with both the short and long-term production capabilities of solar
different forecast horizons. However, the nRMSE values kept PV units. As such, accurate PV energy harvest forecasting is
on increasing with horizon’s duration regardless of the type necessary. Generally, PV energy forecasting can be carried
of a model employed. This goes to show that forecasting out either using the indirect or direct approach. The direct
horizons does have a negative impact of forecasting accuracy. approach performs forecasts that produce PV power as an
immediate and final output [28], [29]. Whereas the indirect
2) WEATHER CONDITIONS approach first predicts solar irradiation from meteorological
In solar energy harvesting, Photovoltaic (PV) transduc- or weather forecast data as input to PV system models to
ers/cells are the most common energy transducers which produce the predicted PV power harvest [30]. The indirect
convert the solar irradiance received on the earth’s surface approach is preferred where historical PV power data is not
into electrical energy [23], [24]. A solar panel is a transducer available, such as in newly constructed PV systems [31] or
made from a combination of parallel and series electrical where weather forecast data is to be used in other applica-
connections of PV modules, which in turn are made up of tions.
connections of PV cells [24]. The efficiency of the cells and In literature, there are many techniques that have been
hence that of the panels used for solar energy harvesting studied and used to perform solar PV power forecasting.
depends on the amount of solar energy or solar irradiance Below are some of the techniques used for existing solar PV
falling on the panel and the ambient temperature at which power forecasting.
the panel is operating [25]. Several other factors, such as
the geographical location and orientation of the installation, A. ARTIFICIAL INTELLIGENCE (AI) AND MACHINE
the materials and design of the panels and the electrical LEARNING (ML) IN SOLAR PV POWER FORECASTING
components used such as charge controllers and inverters also Artificial intelligence (AI) and machine learning (ML)
influence the amount of energy produced by a PV system. promise to improve modeling and forecasting for a host of
However, once the PV system is installed and operational, problems. In solar PV energy harvests, researchers are adopt-
these factors remain constant, except for atmospheric condi- ing AI and ML techniques much more widely since they
tions which can vary. The output of a PV system is measured demonstrate their power in various applications, including
under standard test conditions (STC) with a cell temperature bias correction, handling large datasets, reducing cognitive
of 25 ◦ C, solar intensity of 1000 W/m2 and an air mass of 1.5. overload, and discovering new knowledge in large data sets.

VOLUME 11, 2023 40823

 J. Gaboitaolelwe et al.: Machine Learning Based Solar Photovoltaic Power Forecasting: A Review and Comparison

AI is a field of computer science that focuses on creating to explore problems and generated labeled data for training
computer systems that can demonstrate intelligent behaviors. where most of the training data is not labeled and is paired
This is done by programming them to be able to simulate and with a few data that are labeled.
imitate the thought processes of humans. AI techniques have Fig. 2 provides a depiction of the learning or training pro-
a range of advantages, such as the ability to quickly general- cess in terms of input data, output data and mapping function
ize data, work with multiple variables, incorporate physical involved in the ML approaches.
understanding into models, and allow for the discovery of
patterns and knowledge from data. 2) ML ALGORITHMS USED IN SOLAR PV POWER
Machine learning (ML) is a subfield under AI involved FORECASTING MODEL DEVELOPMENT
in the study and building of systems that learn and improve
Since solar PV power forecasting is a regression problem,
through experience gained from data. In ML, algorithms
most of the ML approaches used fall under the category of
build models based on sample data, known as training data,
supervised learning. Below are some of the ML techniques
to make predictions or decisions without relying on explicitly
used for solar PV power forecasting:
programmed instructions, formulas, or equations. ML models
are presented with many examples relevant to a task. They
find statistical structure in these examples that eventually a: K-NEAREST NEIGHBOURS (K-NN)
allow the system to develop rules for automating the task. K-nearest neighbors (K-NN) is an ML algorithm that can be
ML-based models are promising solutions for predictions in used for both regression and classification problems. I was
Renewable energy generation. Furthermore, ML-based mod- first developed by Evelyn Fix and J.L. Hodges in 1951 [37]
els can conduct highly nonlinear problems [32]; as such, they and expanded upon by [38]. It is based on an algorithm that
are ideal for fluctuating environments as they can adapt to compares the current state of an unknown data point with the
new data [33]. This gives ML models a significant advantage known states of other data points that are like it. These known
relative to statistical models. states are called ‘‘training samples’’ and are represented in a
feature space, which is a mathematical representation of the
1) CLASSIFICATION OF ML ALGORITHMS data. To make a prediction using the K-NN algorithm, we first
The learning algorithms or approaches in Machine Learning calculate the Euclidean distances between the unknown data
are typically categorized into supervised [34] and unsuper- point and the training samples in the feature space. Then,
vised [35]. In some cases, a third approach, semi-supervised we select the first ‘‘k’’ training samples that are closest to the
learning [36] may be considered. Fig. 1 shows a summarized unknown data point, where ‘‘k’’ is a user-specified parame-
overview and classification of ML algorithms. ter. These ‘‘k’’ training samples are known as the ‘‘nearest
neighbors’’ of the unknown data point.
In classification tasks, the K-NN algorithm uses the class
a: SUPERVISED LEARNING
labels of the nearest neighbors to make a prediction for the
Also known as supervised machine learning, it is an approach
unknown data point. For example, if the nearest neighbors of
to machine learning that uses labelled data to train (generate
an unknown data point are all labeled as ‘‘sunny,’’ then the
models) and predict outcomes using the knowledge learned
K-NN algorithm will predict that the unknown data point is
from the data. The trained models can be used to assign
also a ‘‘sunny.’’ In regression tasks, the K-NN algorithm uses
labels (assign data to specific categories) or make numeri-
the values of the nearest neighbors to make a prediction for
cal predictions on new data from the problem that is being
the unknown data point. For example, if the nearest neighbors
solved. The bulk of supervised learning involves problems of
of an unknown data point all have a value of ‘‘10,’’ then the
classification and regression.
K-NN algorithm will predict that the unknown data point also
has a value of ‘‘10.’’
b: UNSUPERVISED LEARNING The K-NN algorithm weights the contribution of each near-
Contrary to supervised machine learning, unsupervised learn- est neighbor to the prediction. The weight of each neighbor
ing is an approach that uses data that is not labelled to train; is inversely proportional to the distance between the neighbor
therefore, must discover hidden patterns or data groupings and its centroid [39], which is the center of the cluster of
without the aid of labels. The algorithms used in unsupervised nearest neighbors. This means that the closer a neighbor is
learning generate models that can analyze, cluster, and cate- to the centroid, the more it will contribute to the prediction.
gorize unlabeled data as well as detect outliers. Unsupervised In regression tasks, the final prediction is the average of the
learning is used for clustering, association, and dimensional- values of the nearest neighbors.
ity reduction. Several works of literature have used KNN to perform
solar PV power forecasting. For instance in [40], the authors
c: SEMI-SUPERVISED LEARNING explored the impact of classifiers and training data on the
Semi-supervised learning, as the name suggests, mixes super- accuracy of daily weather classification and employed two
vised and unsupervised learning. It is a learning approach commonly used classification methods, KNN and SVM,
that combines the use of both labeled and unlabeled data to classify daily local weather types for DAST solar PV power

40824 VOLUME 11, 2023

J. Gaboitaolelwe et al.: Machine Learning Based Solar Photovoltaic Power Forecasting: A Review and Comparison

FIGURE 1. Types of ML algorithms.

TABLE 2. Kernel functions Used in SVM models.

FIGURE 2. ML training process for supervised, unsupervised, and A Kernel function denoted as k(x, x ′ ), is a symmetrical func-
semi-supervised learning.
tion that in practice describes the similarity between obser-
vations based on their feature values. SVM makes use of
several kernel functions which can be categorized as Linear
forecasting. The study found that SVM performs well with and -Non-Linear hence, by extension, SVMs can be clas-
small sample scale while KNN is more sensitive to the length sified, where L-SVM represent SVM that use linear kernel
of the training dataset and can achieve higher accuracy than function and K-SVM are for those that use nonlinear kernel
SVM with sufficient samples. Overall, this paper provides functions. The Radial basis function is the most widely used
valuable insights into how weather classification can improve and recommended because of its simplicity in adaptation for
solar PV power forecasting. optimization, robustness as well as its high efficiency [43].
Examples of kernel functions used in SVMs are provided in
b: SUPPORT VECTOR MACHINE (SVM) TABLE 2.
Vapnik [41] introduced the support vector machine (SVM) where x represents the input space and x ′ represents the train-
concept which is one of the kernel-based ML techniques ing or test samples computed features vector. D represents the
designed for classification as well as regression tasks. Sup- degree and c the coefficient.
port Vector Regression machines (SVR) are the implementa- The mathematical representation of the prediction calcula-
tion of the SVM concept on regression tasks. SVM operates tion of an SVM can be represented by Equation 5.
by use of kernel functions and the Kernel Trick to map low XN
(ai k xi , x ′ + b)


dimensional data into a higher dimensional feature space, y= (5)
i=1
based on that linear solutions in the higher dimensional fea-
ture space correspond to non-linear solutions in the lower where y is the predicted output value, b is the bias parameter,
dimension original state. Thus, SVM is a viable option for ai is the coefficient and xi is the training set input vector at i
dealing with a variety of naturally non-linear problems [42]. th instance with total n training vectors.

VOLUME 11, 2023 40825

 J. Gaboitaolelwe et al.: Machine Learning Based Solar Photovoltaic Power Forecasting: A Review and Comparison

FIGURE 3. Classification of ensemble methods.

SVMs have been used in the works of [44], and [45] to FIGURE 4. Classification of ensemble learning based on learner
generation.
perform solar PV power forecasting. To enhance the accuracy
of PV prediction, [46] introduced a PCA-SVM model. This
model employed PCA to extract the primary features of the
data, which were then used as input for the SVM model. More duce homogeneous base learners, whereas a heterogeneous
works were SVMs have also applied to perform solar PV ensemble uses different learning algorithms to produce het-
power forecasting include research by authors [47], [48]. erogeneous learners. To achieve more accuracy compared to
its individual members, ensemble methods require the base
c: DECISION TREE AND REGRESSION TREE learners to be as accurate and diversified as possible.
A decision tree is a type of ML algorithm that is used to When classifying according to how the base learners
predict the value of a target variable based on the values of are generated, ensemble learning techniques can further be
other input variables. When the target variable is continu- divided into two groups as shown in Fig. 4. These are
ous, meaning that it can take on any real number value, the sequential ensemble and parallel ensemble. It is important
decision tree is called a regression tree. The algorithm works to note that, the ML models used to create the ensembles
by dividing the input data recursively into smaller groups can homogenous or heterogeneous. In a sequential ensemble,
based on the values of the input variables and then making base learners are generated sequentially. The basic motivation
predictions about the target variable by averaging the values of sequential methods is to exploit the dependence between
of the data points in each group. This allows the model to the base learners to improve upon the predictions of the
capture complex non-linear relationships in the data and make prior base learners. Examples of sequential ensemble meth-
more accurate predictions. ods include boosting, blending, and stacking. For Parallel
ensemble, base learners are generated in parallel. The main
d: ENSEMBLE idea is to exploit the independence and variance between
In literature, ensemble learning has been identified as an base learners and reduce the variance of the whole model by
effective way to increase prediction accuracy by combining averaging the errors. A popular method of parallel assemble
and exploiting the advantage of different methods. Ensemble is bagging.
learning is a method of combining the prediction capabilities
of multiple models called weak or base learners to improve e: BOOSTING
predictive performance and accuracy by reducing bias and Boosting is a sequential method used in ML for both regres-
variance exhibited by single-stage models. The key idea is to sion and classification. Boosting is based on using the advan-
appropriately combine models, to generate more accurate and tage of some weak learners and generating them sequentially
resilient models. By combining the efforts of several weak with each learner improving upon the weakness or poor per-
learners, the stability and accuracy are improved as compared formance of the previous weak learner. Boosting algorithms
to single-stage machine learning algorithms. can differ in how they create and aggregate weak learners
Ensemble learning can be described based on the two during the sequential process, however, they are similar in
features, either on the type of base learner or based on how how they use errors and weights to improve the performance
the base learners are connected or generated. Fig. 3 shows of the following learners. This is done by monitoring the
an illustration of how ensemble learning methods can be errors of previous models and using weights based on model
classified. When described according to the base learner, two errors in the generation of new models or the aggregation of
methods exist: homogenous or heterogeneous. A homoge- the results of the weak learners. The redistribution of weights
nous ensemble uses a single base learning algorithm to pro- helps the algorithm identify the parameters that it needs to

40826 VOLUME 11, 2023

J. Gaboitaolelwe et al.: Machine Learning Based Solar Photovoltaic Power Forecasting: A Review and Comparison

FIGURE 5. Structure of boosting process.

focus on to improve its performance. Fig. 5 provides a depic-

tion of the boosting process, whereas equation 6 describes its
mathematical process.
X
f (x) = αt ht (x) (6)
t FIGURE 6. Structure of the bagging process.

where f (x) is the weighted result forming the strong learner,

αt weight of the weak learner and ht (x) is the prediction result
of a weak learner. g: STACKING AND BLENDING
Introduced in 1992, Stacking also known as stacking Gener-
f: BAGGING alization [53] is a hybrid of parallel and sequential ensemble,
Bootstrap aggregating or abbreviated as bagging [49] is a where heterogenous weak learners are trained on the same
method that combines bootstrap sampling, parallel ensemble training data in a parallel manner to produce results or pre-
learning and aggregating of weak learners to create a model. dictions that will be combined with true values to form a
Bootstrapping or bootstrap sampling is a resampling method dataset that is used as training data to train a meta-model
that utilizes row sampling with replacement to randomly or final estimator. The goal is to reduce the generalization
extracts data into several data sets. It is a method popularly error of prediction models by the use of a meta-model. Thus,
used in statistics to estimate statistical values of small datasets the meta-model becomes the strong learner that produces
by creating several data samples called bootstrap samples the final result of the prediction model. It is common that
or bootstraps from the original data set. When calculated 2 training stages or levels are used, where the first stage is
and combined, the average of statistical values of each boot- taken as level 0 and the next is level 1, however more stages
strap achieves a better estimate of the population’s statistics or levels can be added.
as compared to direct estimates from the single original Blending is a variation of Stacking, where instead of using
dataset [50]. Each generated bootstrap sample data set is then the predictions from the whole training dataset or k-fold
used to train a weak learner, hence multiple weak learners cross-validation technique to create training data for the meta-
are created to form a parallel ensemble. The model’s final model, predictions from a holdout set are used to create
output is produced by aggregating the results produced by training data for the meta-model.
the ensemble of trained models, where regression models use
averaging and classification used voting. Thus, resulting in a h: RANDOM FORESTS (RF)
small variance for the forecast results. Fig. 6 provides a depic- Random Forest (RF) is an ensemble machine learning tech-
tion of the bagging process, whereas equation 7 describes its nique that combines multiple decision or regression trees to
mathematical process. create a more accurate predictive model. It was developed by
the works of [54]. Each tree in a Random Forest contributes to
B
1X the mean prediction of individual trees, which can help lower
f (x) = fb (x) (7)
B the variance of the model. As decision or regression trees are
b=1
prone to overfitting, which leads to low bias but high variance,
where f (x) is the aggregated result forming the strong Random Forests capitalize on the growth of several trees for
learner, B is the number of bootstrap samples and fb (x) is each case study, which helps to reduce the variance of the
the prediction result of a weak learner. model by averaging the results from each tree. However, the
Examples works from the literature that have utilized bag- reduction in variance does result in a minor increase in bias
ging in solar PV forecasting include [30], [51], and [52]. while minimizing the variance.

VOLUME 11, 2023 40827

 J. Gaboitaolelwe et al.: Machine Learning Based Solar Photovoltaic Power Forecasting: A Review and Comparison

FIGURE 8. Structure of feed forward neural network.

FIGURE 7. Structure of a neuron.

i: ARTIFICIAL NEURAL NETWORKS (ANNs)

An artificial neural network (ANN) is an ML approach mod-
eled on the function and structure of the biological neural net-
work [55]. Artificial neural networks (ANNs) have been used
for time series forecasting since the late 1980s [56]. They are
comprised of interconnected information-processing units
known as neurons [55], [57], [58]. A neuron is also called
a node or a perceptron and is a computational unit in which FIGURE 9. Structure of recurrent neural network.
decision-making and information processing occurs. A Neu-
ral Network (NN) is formed by the interconnection of the
neurons or nodes with connections in which each neuron
has one or more weighted input connections, an activation
function, and an output connection. Fig. 7 depicts the basic
structure of a neuron.
The general architecture of an ANN is separated into three
sections or layers: input layer, hidden layer and output layer.
Except for input layer neurons, inputs to a given node are mul-
tiplied by the weights and then summed to produce a summed
activation value. The summed activation is then transformed
into the node’s specific output or activation through an acti-
vation function. Other following nodes progressively process FIGURE 10. Structure of single layer perceptron neural network.
this generated output until it reaches the output layer. Such
a method like this has proven to be beneficial in a wide
range of scenarios with a big number of inputs. In both
linear and nonlinear problems, ANN models can respond well A RNN, is a class of NN structure that introduces a form of
[59], outperforming traditional empirical models [60], [61]. memory into a NN through loops or feedback. The feedback
As such ANNs have also been used for solar PV forecasting is such that the output from a node is kept in the network and
as seen other works of literature such as [61], [62], and [63]. used together with the input of the current timestep as input
A variety of topologies exist categorized by the connec- into a node to produce output of the current timestep. Fig. 9
tion forms between input, hidden, and output layers. ANNs shows an illustration of a RNN.
can be classified depending on the direction of information A Single Layer Perceptron is a NN that does not contain
flow (Feed-Forward (FF) or Recurrent (R)) or how many hidden layers, it only has the input and output layer. On the
hidden layers they have (Single Layer Perceptron (SL) or other hand, a Multi-Layer Perceptron is a NN that has one or
Multi-Layer Perceptron (MLP)). A brief description of each more hidden layers together with the input and output layers.
is given. Fig. 10 and Fig. 11 show an illustration of a SLP and MLP
An FF-NN is a class of NN where information only flows in respectively.
one direction. Each neuron in the network is connected such
that it only connects to the neurons in the next layer without j: DEEP NEURAL NETWORK
any loops or feedback in or between the layers. It is one of the Focusing on the number of hidden layers in a NN, NNs
most widely used ANNs [64], [65]. Fig. 8 shows the general with less than 3 hidden layers are considered shallow NN or
structure of an FF-NN. shallow learners whereas NNs with more than 3 hidden layers

40828 VOLUME 11, 2023

J. Gaboitaolelwe et al.: Machine Learning Based Solar Photovoltaic Power Forecasting: A Review and Comparison

forecasted or model predictions. It is suitable for evaluating

uniform forecast errors.
N
1 X
MAE = Pm − Pp (9)
N
i=1

C. MEAN BIAS ERROR (MBE)

Captures a forecasting model’s systematic tendency to over-
estimate or underestimate.
N
1 X

FIGURE 11. Structure of multi-layer perceptron neural network. MBE = Pm − Pp (10)
N
i=1

are referred to as Deep NN or deep learners. The addition of D. ROOT MEAN SQUARE ERROR (RMSE)
more layers to a NN adds complexity to the network structure. Root mean squared error (RMSE) is a measure of how well
a model’s predictions match the true values. RMSE is calcu-
IV. PERFORMANCE METRICS lated by taking the difference between the true values and the
An important aspect of forecasting is model performance and predicted values, squaring this difference, and then averaging
accuracy. Several metrics can provide insight into the accu- the squared differences. By squaring the differences, RMSE
racy and performance of a model. The use of metrics allows gives larger errors a higher weight, which means that outliers
comparisons between different ML models and locations. will have a greater impact on the RMSE value. This means
Each of them emphasizes a different aspect. As a result, the that models with a lower RMSE are generally considered to
model can never be evaluated by one metric that is valid in be more accurate because they make fewer large errors.
all situations. Instead, each of them contributes to measuring v
u
u1 X N
the model’s accuracy.
2
RMSE = t Pm − Pp (11)
In other cases, instead of representing errors as absolute N
i=1
values, normalized performance metrics are used to facilitate
comparisons between models or datasets of different scales E. NORMALIZED OR RELATIVE ROOT MEAN SQUARE
and sizes. The literature describes several methods for accom- ERROR (nRMSE/rRMSE)
plishing performance metric normalization, the majority of q
which involves dividing the target matrices by statistics based PN
2
i=1 Pm − Pp
on observational data. For normalized metrics, the letter ‘‘N’’, nRMSE = (12)
‘‘n’’ or ‘‘r’’ is appended to indicate that they have been Pc
normalized. Some of the common ways to normalize are: where; Pc is the solar PV system power output capacity.
1) Mean (average).
2) Standard deviation. F. MEAN ABSOLUTE PERCENTAGE ERROR (MAPE)
3) Range, i.e., difference between the maximum and min- Mean absolute percentage error (MAPE) is a measure of fore-
imum observations. cast accuracy that is similar to mean absolute error (MAE).
4) Interquartile range and standard deviation. Both measures are used to evaluate the error of a model’s
5) Solar PV system size or estimated capacity. predictions, but MAPE is expressed as a percentage rather
Below are a group of commonly used performance metrics than as an absolute value. MAPE is calculated by taking the
[5]. The matrices stated below are not unique to any field absolute value of the difference between the true value and the
and can be used to evaluate both PV solar energy harvest predicted value and then dividing this value by the true value.
prediction and load demand forecasting. This percentage error is then averaged over all the predictions
made by the model. Like MAE, MAPE is best used when
A. NORMALIZED ERROR (nE) the errors of the model are uniform or consistent across all
predictions.
Pm − Pp
nE = (8) N
max Ppred

100 X Pm − Pp
MAPE = (13)
N Pc
where; Pp is the predicted power output or load demand; and i=0

Pm is the measured power output or load demand. where; Pc is the solar PV system power output capacity.

B. MEAN ABSOLUTE ERROR (MAE) G. MEDIAN ABSOLUTE PERCENTAGE ERROR (MdAPE)

It is a metric that represents the average magnitude of fore- Median absolute percentage error (MdAPE) is calculated by
cast errors between the actual or measured values and the taking the median of the absolute percentage errors between

VOLUME 11, 2023 40829

 J. Gaboitaolelwe et al.: Machine Learning Based Solar Photovoltaic Power Forecasting: A Review and Comparison

the true values and the predicted values. MdAPE is similar to TABLE 3. The difference between parallel and sequential ensemble
learning.
MAPE, but it is less sensitive to outliers because it uses the
median instead of the mean to average the errors. MdAPE
is often used as a measure of forecast accuracy, particularly
when the true values and predicted values have different
scales or units of measurement.
Pm − Pp
MdAPE = median 100 (14)
Pm

H. SKILL SCORE (SS)/FORECASTING SKILL

Skill Score is used to compare one forecasting model with
another. It can be approximated as [5];
RMSE RMSEmodel
SSmodel ≈1− (15)
RMSEbaseline
When using a persistence forecast as a reference or base-
line model, the SS score of the persistence model against
itself is equal to 0. A model that exhibits better forecasting
accuracy when compared to the reference model results in
an SS greater than 0. The higher and closer to 1 the SS, the
more accurate and better performing the forecasting model
is. However, models with SS less than 0, indicate a decline in
performance and lower accuracy of the model being tested.

I. COEFFICIENT OF DETERMINATION (R2 )

An R2 is a measure of how much of the variability in a data
set can be explained by a regression or prediction model. For
predictions that are no different to guessing, the R2 score
would be close to zero; indicating that much is unexplained
models. The RMSE and MBE are commonly used metrics
by the regression model. Values close to one mean that the
but comparing the performance of different ML models can
regression or prediction model can capture and or predict the
be challenging when they operate on different scales of PV
dataset well.
Pn
2 systems with varying maximum theoretical production capac-
2 i=1 Pm − Pp 2 ity. To address this issue, normalized or relative error metrics,
R = 1− P
2 0 ≤ R ≤ 1 (16)
n ¯ such as the normalized NRMS, are recommended. However,
i=1 Pm − Pm
few papers from the literature as shown in the comparative
V. COMPARISON OF ML PV POWER FORECASTING table present the accuracy of their models in a normalized or
METHODS AND CORRESPONDING RESEARCHERS relative error form, making it difficult to compare the fore-
In literature, many techniques have been studied and used casting performance of different ML models. One potential
to perform solar PV power forecasting. Each method has solution to this challenge is to normalize error metrics such as
advantages and disadvantages and often, the performances of RMSE and MBE using the maximum theoretical production
the forecasting methods are comparable. However, depending capacity of the PV system, which enables a fair comparison
on the available data, each problem might be solved with of different ML solar PV forecasting models by removing the
a different approach and undergo different optimizations, bias caused by the varying maximum theoretical production
which would result in other PV power forecasting methods capacity of the PV systems. Using a normalized error metric
significantly outperforming other methods. Therefore, there or normalizing RMSE and MBE can help to identify the
is merit in testing and reviewing the different forecasting best performing models for different scales of PV systems.
approaches found in the literature. This section provides an However, some cases arises where some information are not
overview of related works primarily comprised of TABLE available for some papers, and we indicate such situation
4, which shows some works from 2010 – 2022 in PV power using ‘‘NA’’.
forecasting. Based on TABLE 4, the solar PV forecasting model with
the least NRMSE is that by [66], followed by the works of
A. ANALYSIS OF OF ML PV POWER FORECASTING [67]. Both methods realized the models by utilizing the hybrid
METHODS or ensemble approach, where [67] combined ANN and post-
When evaluating the performance of solar PV forecasting processing with a linear regressive model thereby making a
models, it is important to use accurate and consistent perfor- hybrid or series ensemble with NRMSE of 7.53% in a cold
mance metrics that enable fair comparisons between different climate. While [66] using parallel ensembles,introduced a

40830 VOLUME 11, 2023

J. Gaboitaolelwe et al.: Machine Learning Based Solar Photovoltaic Power Forecasting: A Review and Comparison

TABLE 4. ML PV power forecasting methods.

VOLUME 11, 2023 40831

 J. Gaboitaolelwe et al.: Machine Learning Based Solar Photovoltaic Power Forecasting: A Review and Comparison

TABLE 4. (Continued.) ML PV power forecasting methods.

new ensemble, called probabilistic ensemble method (PEM), better understand how their models perform under different
for improving the forecasting performances in cloudy days, environmental conditions and identify the best models for
hence resulting in a NRMSE of 4.79%. various weather conditions and seasonal changes.
Evaluating the forecasting performance of solar PV models Geographic location is a crucial factor in evaluating the
in adverse weather conditions is essential to showcase their performance of solar PV forecasting models. Few works of
robustness and versatility. Atmospheric conditions signifi- literature as shown in the comparative table provided infor-
cantly impact the performance of the solar PV forecasting mation on the geographic location of the PV systems used in
models, and variations can pose significant challenges to their studies, and this can significantly influence the perfor-
the accuracy of the models. As shown in the comparative mance of the solar PV systems. Models trained on data from
table, mostly recent literature included performance analy- PV systems in different locations would have varying degrees
sis of solar PV forecasting models during adverse weather of accuracy when applied to other locations. Therefore, it is
conditions to assess their reliability in providing accurate essential to take into consideration the geographic location
forecasts. By including such information, researchers can of the PV systems used when evaluating and comparing

40832 VOLUME 11, 2023

J. Gaboitaolelwe et al.: Machine Learning Based Solar Photovoltaic Power Forecasting: A Review and Comparison

the performance of solar PV forecasting models. Evaluating TABLE 5. Solar PV power production dataset information.
models based on their performance in the same geographic
location or at least in regions with similar climatic conditions
is necessary to obtain accurate and meaningful results.

B. ML EXPERIMENT SIMULATIONS
To further enhance the comparison and provide more insights
into the advancement in the area, we simulate the perfor-
mance of different ML methods used in solar PV power
forecasting. The ML models selected to be simulated include
Support Vector Regressor (SVR), SVM, Random Forest
(RF), Gradient Boosting (GB), Extreme Gradient Boost
(XGB), Multi-Layer Perceptron and a lasso regressor. The
models were selected based on their effectiveness in regres-
sion tasks and their suitability for handling large datasets with
high dimensionality. Additionally, we considered the ability
of the models to capture the complex non-linear relationships
between the input features and the target variable in solar
PV power forecasting. Based on these criteria, we selected
a range of ML algorithms that have been widely used in
solar PV power forecasting and have shown promising results
Renewable and Sustainable Energy [84]. It contains an exten-
in previous studies. Section VI detailed our experimental
sive 175 time series of rooftop-mounted PV system power
simulations and analysis of the different methods.
measurements of residentials located in the city of Utrecht in
the Netherlands from the year 2014 to 2017. Each rooftop PV
VI. EXPERIMENT SETUP production time series is recorded at 1-minute intervals and
A. EXPERIMENT HARDWARE consists of 4 years of data from December 2014 to November
The experiment is implemented in Windows 10 64-bit Desk- 2017. Information about the solar PV dataset is summarized
top Computer with an Intel(R) Core (TM) i7-10750H CPU @ in TABLE 5.
2.60 GHz processor, 1TB HDD and 8 GB DDR3 RAM. The Apart from the PV power production records, the open
code was written in Python, and ran in a Jupiter notebook access dataset also includes a metadata file of the PV systems,
environment, with Microsoft Visual Studio Code used as an including each PV system’s unique ID, estimated DC and AC
IDE. capacity, annual yield, start and end time of registered power
measurements, azimuth and tilt angle, and mapped grids of
B. EXPERIMENT SOFTWARE PACKAGES the system location. The dataset has been used by several
The main software package used in the experiment is the works in literature to carry out predictions [85], [86], [87].
Scikit Learn library, with NumPy and Pandas as support
packages. Scikit Learn is an open-source python library that is
a simple and efficient tool for statistical modeling, predictive 2) DAY-AHEAD FORECASTED WEATHER DATA
data analysis, and machine learning. It is based on python The weather data used in this work is publicly available
packages NumPy, SciPy, and Matplotlib. It provides efficient from European Centre for Medium-Range Weather Forecasts
methods for dealing with common machine-learning tasks (ECMWF). It consists of 24 hours day ahead, 1-hour resolu-
such as classification, regression, clustering, model selection, tion weather forecasts data corresponding to the same time
and dimensionality reduction. Furthermore, Scikit learn can and location as the solar PV dataset. The weather data has a
be used to create models for both supervised learning and total of 14 weather-related parameters and 2 date and time
unsupervised learning. information parameters, hence a total of 16 variables are
included in the dataset. Information about the meteorological
C. EXPERIMENT DATA dataset is summarized in TABLE 6.
Data used in this work consist of two datasets, these being
PV power production time series and forecasted weather D. EXPERIMENT METHODOLOGY
time series. Further information regarding each dataset The purpose of this experiment is to analyze the dataset
is given. and explore the application of machine learning (ML) meth-
ods to develop regression models to make predictions of
1) ROOFTOP SOLAR PV SYSTEM POWER GENERATION DATA residential Rooftop PV power production. Implementation
This work used a dataset that is an extract from an open- of the work is carried out using the framework as shown
access dataset by Visser et al.; published in the Journal of in Fig. 12.

VOLUME 11, 2023 40833

 J. Gaboitaolelwe et al.: Machine Learning Based Solar Photovoltaic Power Forecasting: A Review and Comparison

TABLE 6. Meteorological dataset information. TABLE 7. Feature variables for model creation.

Furthermore, additional processes such as feature engi-

neering and feature selection may be performed on the dataset
to improve it. The goal of the experiment is to design a
predictive model to perform solar PV power prediction on
FIGURE 12. Machine learning solar PV power production framework. 24h ahead time horizon. Hence, the target variable was set
to be solar PV power production while weather data and time
properties were used as feature variables. The data underwent
1) DATA PREPARATION feature engineering, where additional features were derived
Machine learning algorithms build prediction models based from the existing data and the domain knowledge of solar PV
on data; hence data preparation and cleaning are essential power production. For feature engineering, information was
steps in building ML forecasting models. The use of unpro- extracted from the date and time features, the extracted data
cessed data leads to a loss in the forecasting model accuracy included the day of the month and day of the year. Three extra
The Weather data was clean and had no missing records. The features were generated from the solar PV data by adding
solar PV power dataset underwent multiple processes where 24, 48 and 72 hours of lagged data. TABLE 7 shows the
it is analyzed and cleaned. This involved filling in missing complete list of the feature variables used in the ML model
data and identifying outliers. development stage.

40834 VOLUME 11, 2023

J. Gaboitaolelwe et al.: Machine Learning Based Solar Photovoltaic Power Forecasting: A Review and Comparison

TABLE 8. Grid search hyperparameter search space.

The holdout validation process was used to train and test 24 hours Persistence method and the time of day daily average
the predictive models, hence the data employed in this study method.
was divided into two subsets. The training set contained 70%
of the data and the test set included the remaining 30%.
a: 24H PERSISTENCE
A larger training set was used to ensure the reliability of the
predictive model. The 24H persistence method uses the previous day’s observed
power output as the forecasted value for solar PV output. This
2) MODEL DEVELOPMENT
is expressed in the equation.
In the model development, several ML algorithms were used Pp (t) = Pm (t − 24) (17)
to train and create the prediction models. The ML prediction
models created in the work include Gradient Boost (G Boost), b: TIME OF DAY DAILY AVERAGE (AVERAGE)
extreme gradient Boost (XG boost), Support Vector Regres- This method used the average of previous solar PV power
sion (SVR), Random Forest (RF), Lasso Regression (LR), production based on time as the forecast of the next day at
Ada Boost and A Multi-Layer Perceptron neural network the said time of day.
(MLP NN). The default parameters of regression models in To improve the performance of the ML models, hyper-
the Scikit Learn library were used as a basis for the hyperpa- parameter optimization was carried out on all the models.
rameters of each model. The optimized models’ hyperparam- Grid search hyperparameter optimization was used as the
eters were chosen by conducting a grid search optimization optimization method.
within the range indicated in TABLE 8. The search area was
determined by building upon the default hyperparameters of
Scikit Learn. 3) TESTING/EVALUATION
When evaluating the performance of forecasting pro- Several evaluation metrics are used to compare the perfor-
cedures, baseline (benchmark) methods are used to com- mance of the developed forecasting models. Each metric
pare. Two methods were used to produce the baseline test emphasizes a different aspect of forecasting hence using sev-
results. The baseline methods used in the experiment were eral evaluation metrics provides insight into the accuracy and
based on assumptions of persistence about the underly- performance of a model. The forecast evaluation metrics used
ing process of solar PV power production. This includes in this work are shown in TABLE 9.

VOLUME 11, 2023 40835

 J. Gaboitaolelwe et al.: Machine Learning Based Solar Photovoltaic Power Forecasting: A Review and Comparison

TABLE 9. Forecast evaluation metrics.

VII. RESULTS AND DISCUSSION The 24hr persistence method is selected as the reference
This section presents the results and discussion of the exper- model when calculating skill score (SS), thus resulting in the
iments. SS score of 0 for the 24hr persistence. To make the SS score
easier to read, it has been represented as a percentage. The
A. BASELINE FORECASTS PERFORMANCE average baseline method shows an SS score of 16.6% which
The baseline methods created made simple forecasts based indicates a 16.6% improvement in forecasting accuracy when
on strong assumptions of slow change or persistence of daily compared to the 24hr persistence reference model.
solar PV power production, hence the performance of the Further analysis of the results of the two baseline models
models is not expected to be high. This is evident by the through graphical representations is shown in Fig. 14 and
results shown in TABLE 9, while Fig. 13 which shows the Fig. 15. Where Fig. 14 shows a line plot of the residuals across
nRMSE in percentage, Coefficient of determination and skill the test data of the two-baseline methods Fig. 15 shows a
score bar graph representation of the model performances. localized perspective of the model’s performance through box
Between the two baseline methods, the 24hr persistence plots of the model’s absolute relative error for each hour of
method displays the least accuracy and prediction perfor- the day. From Fig. 14 the superiority of the average baseline
mance across all the performance metrics. Such results are method over the persistence model is further evident as shown
to be expected since it only uses the previous day’s solar by the lesser residuals in most of the instances in the test data.
PV power output as the forecast, however, consecutive days Fig. 15 further shows that the daily prediction error for the
can experience vastly different weather conditions hence persistence model is more than the average baseline model,
resulting in large differences in solar PV power output. hence the higher median values, longer interquartile range,
Furthermore, since the 24hr persistence model only consid- and longer error ranges. Thus, the average model consistently
ers one data point instance, having a few data points sam- outperforms the persistence model across all the performance
pled results in poor forecasting performance. The average metrics as shown in TABLE 10.
(BSL_Average) method shows better performance than the
24hr persistence and this is also expected since it considers B. ML MODEL FORECASTS PERFORMANCE
more data points; as such it can better reflect the underlying The results of the ML model performance are shown in
process, hence the high R2 above 0.5, which indicates that it TABLE 11 and TABLE 12 shows the results of the ML
is able to model close to half of the underlying process for models that were designed using the default Scikit Learn
solar PV power production. library model Hyperparameters, whereas TABLE 12 shows

40836 VOLUME 11, 2023

J. Gaboitaolelwe et al.: Machine Learning Based Solar Photovoltaic Power Forecasting: A Review and Comparison

FIGURE 13. Performance of baseline methods on testing data. (a) Normalized RMSE as percentage. (b) Coefficient of determination. (c) Skill score as
percentage.

FIGURE 14. Residuals of persistence and average baseline forecast methods.

FIGURE 15. Boxplot for absolute relative error for each hour of the day. (a) Results for persistence. (b) Results for average baseline forecast methods.

the performance of the models that were designed with hyper- is evident by both the R2 and SS for all ML models being
parameter optimization using grid search. greater than that of the baseline methods. Furthermore, the
superiority of the ML models over the baseline forecasting
1) ML MODEL WITH DEFAULT HYPERPARAMETER models is also evident by all other error metrics showing
Based on TABLE 11, it can be noted that the ML models better scores for all the ML models when compared to both
perform better than the baseline methods at forecasting. This the average and persistence baseline methods.

VOLUME 11, 2023 40837

 J. Gaboitaolelwe et al.: Machine Learning Based Solar Photovoltaic Power Forecasting: A Review and Comparison

TABLE 10. Performance of baseline forecasts methods.

TABLE 11. Performance of ML models with default hyperparameter.

TABLE 12. Performance of ML models after grid search hyperparameter optimization.

FIGURE 16. Performance of default hyperparameter ML methods on testing data. (a) Normalized RMSE as percentage. (b) Coefficient of determination.
(c) Skill score as percentage.

Focusing on comparing the performance of the ML models between the best model and all the other ML models is not
with default hyperparameters, as shown in TABLE 11 and as significant as compared to the large leap between the ML
Fig. 16, the gradient boost forecasting model is the model and the baseline models. The least accurate model in the ML
with the least forecasting error and highest SS improvement, models is the Ada Boost, with an SS difference of 5.3% when
hence making it the best model. However, the difference compared to the best-performing model.

40838 VOLUME 11, 2023

J. Gaboitaolelwe et al.: Machine Learning Based Solar Photovoltaic Power Forecasting: A Review and Comparison

FIGURE 17. Performance of grid search hyperparameter optimization ML methods on testing data. (a) Normalized RMSE as percentage. (b) Coefficient
of determination. (c) Skill score as percentage.

FIGURE 18. Combined performance evaluation of ML and baseline methods on testing data. (a) Normalized RMSE as percentage.
(b) Coefficient of determination.

FIGURE 19. Combined performance evaluation of ml and baseline methods on testing data; skill score as
percentage.

C. ML MODEL WITH GRID SEARCH HYPERPARAMETER retrained using grid search hyperparameter optimization. The
OPTIMIZATION results of the optimized models can be seen in TABLE 12 and
After training and testing the ML models using default Scikit Fig. 17. TABLE 12 is formatted such that an improvement
Learn library model Hyperparameters, new ML models were in model forecasting performance is highlighted in green

VOLUME 11, 2023 40839

 J. Gaboitaolelwe et al.: Machine Learning Based Solar Photovoltaic Power Forecasting: A Review and Comparison

TABLE 13. Combined methods evaluation.

D. COMBINED MODEL FORECASTS PERFORMANCE

TABLE 13, Fig. 18, and Fig. 19 show the results of all
the tested forecasting models. From TABLE 13, when the
performance of the forecasting models is ranked based on SS
results, it shows that the grid search optimized Random Forest
model is the best model with a SS of 37.33% and nRMSE
of 10.07%. Apart from the baseline models, the worst-
performing forecasting model is the optimized Multi-Layer
Perceptron NN.
Fig. 20 and Fig. 21 provide a graphical representation of
the comparisons of the best-performing model against the
baseline model, where Fig. 20 shows a broad perspective of
FIGURE 20. Residuals of persistence and random forest forecast
methods. the residuals of the models across the test data and Fig. 21
shows a localized perspective of the model’s performance
through box plots of the model’s absolute relative error for
while a decline is highlighted in red. From TABLE 12, it is each hour of the day. It is evident from Fig. 20 that the
shown that the performance of other ML models improved optimized Random Forest model performs better than the
after grid search optimization, such as for Random Forest, baseline model consistently across most of the test data. This
XG Boost and Ada Boost. However, the performance of is shown by the residuals of the Random Forest being less
the other ML models declined, such as that of G Boost than those of the persistence model in most of the forecasts
and the rest. The decline in the performance of the other made by the two models. When comparing the boxplots in
ML models may be caused by overfitting the ML models Fig. 21, though the Random Forest model has more outliers
to the training data, hence the drop in performance when than the persistence model, both the error bars and outliers
the models are tested on new data. When sorted accord- for the Random Forest model are less than those for the
ing to SS, the Random Forest model performed the best, persistence model except for the hours 1 to 7 and 18 to 24.
while the multi-layer perceptron NN was the least performing This indicates that the Random Forest model is performing
model. poorly in nighttime instances.

40840 VOLUME 11, 2023

J. Gaboitaolelwe et al.: Machine Learning Based Solar Photovoltaic Power Forecasting: A Review and Comparison

FIGURE 21. Boxplot for absolute relative error for each hour of the day. (a) Results for persistence baseline forecast method. (b) Results for random
forest ML forecast methods.

VIII. CONCLUSION AND FUTURE WORKS Random Forest (RF), Lasso Regression (LR), Ada Boost
A. CONCLUSION and A Multi-Layer Perceptron neural network (MLP NN).
This paper presented a comprehensive and comparative The baseline methods used in the experiment are 24 hours
review of existing papers on Machine Learning (ML) based Persistence method and the time of day daily average method.
approaches used in photovoltaic (PV) power forecasting. The Using the solar PV power forecasting performance metrics
work provided an overview of factors affecting solar PV presented in TABLE 9, findings from the simulation work
power forecasting. These factors are weather conditions and show that ML models performed better than the baseline
forecasting horizons. From the weather conditions aspect, methods. This is evident from the ML models having lower
the effectiveness of solar PV systems is dependent on many RMSE and higher SS results than the baseline models as
factors including the amount of solar irradiance and ambient shown in TABLE 13. Furthermore, for the ML models with
temperature which are subject to weather conditions. Accu- default hyperparameters, GBoost performed best with a SS
rate weather forecasting can be used to improve the ability of 37.326% while the RF model with optimized hyperparam-
to predict the power output of PV systems, and the meteoro- eters performed the best with a SS of 37.33% also shown in
logical variables that have the greatest impact on PV power TABLE 13.
can vary depending on the location of the PV plant. From the
aspect of forecasting horizon, solar PV power forecasting can
be categorized based on the forecasting time horizon, where B. FUTURE WORKS
the typical forecasting time horizons include very short term, In the future, there are several areas that could be explored
short term, medium term and long term as shown in TABLE 1. further in the field of photovoltaic (PV) forecasting. The
The paper further presented work providing an overview accuracy of PV forecasting is essential for ensuring the reli-
of ML, a description of several ML algorithms used in solar able and efficient operation of the electric grid or off-grid
PV power forecasting and a description of commonly used PV systems. Numerical weather prediction (NWP) models
performance metrics used for solar PV power forecasting. are one of the key factors in the accuracy of PV forecast-
The presented ML algorithms are K-NN, SVM, Decision ing. To improve PV forecasting, future work should focus
trees, Ensemble models such as random forest and ANN. on enhancing NWP models by incorporating additional data
The commonly used ML performance evaluation metrics are sources and utilizing advanced machine learning algorithms.
provided in chapter 4. Furthermore, an overview of related Additionally, developing more sophisticated physical models
works is provided in TABLE 4, which shows some works of atmospheric processes could also improve the accuracy of
from 2010 – 2020 in solar PV power forecasting. PV forecasting.
Finally, a comparison of ML models based on simulations As the use of PV systems for electricity generation contin-
of solar PV forecasting was performed, where several ML ues to increase, there is an increasing need for accurate and
solar PV forecasting models were created and compared reliable PV forecasting. This will require the development of
against each other, and 2 baseline methods based on assump- advanced data preprocessing techniques that can handle large
tions of persistence. In the simulations, the ML prediction and complex datasets. In addition, AI models such as deep
models are trained on historical PV power production data learning algorithms could be explored for use in tasks such
and forecasted weather variables. The ML models created as anomaly detection and outlier removal. Probabilistic fore-
in the work include Gradient Boost (G Boost), extreme gra- casting techniques could also be applied to data preprocessing
dient Boost (XG boost), Support Vector Regression (SVR), in PV forecasting.

VOLUME 11, 2023 40841

 J. Gaboitaolelwe et al.: Machine Learning Based Solar Photovoltaic Power Forecasting: A Review and Comparison

Deep learning algorithms have demonstrated potential for [15] Y. Essam, A. N. Ahmed, R. Ramli, K.-W. Chau, M. S. I. Ibrahim,
use in PV forecasting but have not yet been widely adopted. M. Sherif, A. Sefelnasr, and A. El-Shafie, ‘‘Investigating photovoltaic solar
power output forecasting using machine learning algorithms,’’ Eng. Appl.
Future work could focus on exploring different deep-learning Comput. Fluid Mech., vol. 16, no. 1, pp. 2002–2034, Dec. 2022, doi:
architectures and algorithms to improve the accuracy and 10.1080/19942060.2022.2126528.
reliability of PV forecasting models. As the use of renewable [16] O. Gungor, J. Garnier, T. S. Rosing, and B. Aksanli, ‘‘LENARD:
Lightweight ENsemble LeARner for MeDium-term electricity con-
energy grows, probabilistic forecasting will become increas- sumption prediction,’’ in Proc. IEEE Int. Conf. Commun., Control,
ingly important for power system operations. Future work Comput. Technol. Smart Grids (SmartGridComm), Tempe, AZ, USA,
should focus on developing probabilistic forecasting models Nov. 2020, pp. 1–6. Accessed: May 17, 2021. [Online]. Available:
https://ieeexplore.ieee.org/document/9303012/, doi: 10.1109/SmartGrid-
that can accurately predict the uncertainty of PV power gen- Comm47815.2020.9303012.
eration and incorporate this uncertainty into power system [17] A. Dimovski, M. Moncecchi, D. Falabretti, and M. Merlo, ‘‘PV forecast
decision-making and risk management of solar PV power for the optimal operation of the medium voltage distribution network:
A real-life implementation on a large scale pilot,’’ Energies, vol. 13, no. 20,
forecasting systems. p. 5330, Oct. 2020, doi: 10.3390/en13205330.
[18] S. Sreekumar and R. Bhakar, ‘‘Solar power prediction models: Clas-
sification based on time horizon, input, output and application,’’ in
REFERENCES
Proc. Int. Conf. Inventive Res. Comput. Appl. (ICIRCA), Coimbatore,
[1] IEA. (Jul. 2021). Electricity Market Report—July 2021. [Online]. Avail- India, Jul. 2018, pp. 67–71. Accessed: Apr. 13, 2021. [Online].
able: https://www.iea.org/reports/electricity-market-report-july-2021 Available: https://ieeexplore.ieee.org/document/8597288/, doi: 10.1109/
[2] IEA, Paris, France. (2020). Global Energy Review 2020. ICIRCA.2018.8597288.
Accessed: Mar. 23, 2021. [Online]. Available: https://www.iea.org/reports/ [19] A. Rai, A. Shrivastava, and K. C. Jana, ‘‘An empirical analysis of machine
global-energy-review-2020/electricity learning algorithms for solar power forecasting in a high dimensional
[3] IRENA, International Renewable Energy Agency. (2017). Electric- uncertain environment,’’ IETE Tech. Rev., pp. 1–16, Nov. 2022, doi:
ity Storage and Renewables: Costs and Markets to 2030. Accessed: 10.1080/02564602.2022.2136270.
Apr. 12, 2021. [Online]. Available: https://www.irena.org/publications/ [20] R. Blaga, A. Sabadus, N. Stefu, C. Dughir, M. Paulescu, and V. Badescu,
2017/Oct/Electricity-storage-and-renewables-costs-and-markets ‘‘A current perspective on the accuracy of incoming solar energy forecast-
[4] J. Cifuentes-Faura, ‘‘European union policies and their role in combating ing,’’ Progr. Energy Combustion Sci., vol. 70, pp. 119–144, Jan. 2019, doi:
climate change over the years,’’ Air Qual., Atmos. Health, vol. 15, no. 8, 10.1016/j.pecs.2018.10.003.
pp. 1333–1340, Aug. 2022, doi: 10.1007/s11869-022-01156-5. [21] V. P. A. Lonij, A. E. Brooks, A. D. Cronin, M. Leuthold, and K. Koch,
[5] J. Antonanzas, N. Osorio, R. Escobar, R. Urraca, F. J. Martinez-de-Pison, ‘‘Intra-hour forecasts of solar power production using measurements
and F. Antonanzas-Torres, ‘‘Review of photovoltaic power from a network of irradiance sensors,’’ Sol. Energy, vol. 97, pp. 58–66,
forecasting,’’ Sol. Energy, vol. 136, pp. 78–111, Oct. 2016, doi: Nov. 2013, doi: 10.1016/j.solener.2013.08.002.
10.1016/j.solener.2016.06.069. [22] M. Lipperheide, J. L. Bosch, and J. Kleissl, ‘‘Embedded now-
[6] S. Comello, S. Reichelstein, and A. Sahoo, ‘‘The road ahead for solar PV casting method using cloud speed persistence for a photovoltaic
power,’’ Renew. Sustain. Energy Rev., vol. 92, pp. 744–756, Sep. 2018, doi: power plant,’’ Sol. Energy, vol. 112, pp. 232–238, Feb. 2015, doi:
10.1016/j.rser.2018.04.098. 10.1016/j.solener.2014.11.013.
[23] S. Qazi, ‘‘Fundamentals of standalone photovoltaic systems,’’ in Stan-
[7] U. H. Ramadhani, M. Shepero, J. Munkhammar, J. Widén, and
dalone Photovoltaic (PV) Systems for Disaster Relief and Remote
N. Etherden, ‘‘Review of probabilistic load flow approaches for power
Areas, S. Qazi, Ed. Amsterdam, The Netherlands: Elsevier, 2017, ch. 2,
distribution systems with photovoltaic generation and electric vehicle
pp. 31–82. Accessed: Mar. 19, 2020. [Online]. Available: http://www.
charging,’’ Int. J. Electr. Power Energy Syst., vol. 120, Sep. 2020,
sciencedirect.com/science/article/pii/B9780128030226000022
Art. no. 106003, doi: 10.1016/j.ijepes.2020.106003.
[24] C. V. Nayar, S. M. Islam, H. Dehbonei, K. Tan, and H. Sharma,
[8] R. Ahmed, V. Sreeram, Y. Mishra, and M. D. Arif, ‘‘A review and eval-
‘‘Power electronics for renewable energy sources,’’ in Alternative
uation of the state-of-the-art in PV solar power forecasting: Techniques
Energy in Power Electronics, M. H. Rashid, Ed. Boston, MA,
and optimization,’’ Renew. Sustain. Energy Rev., vol. 124, May 2020,
USA: Butterworth-Heinemann, 2011, ch. 1, pp. 1–79. Accessed:
Art. no. 109792, doi: 10.1016/j.rser.2020.109792.
Mar. 23, 2020. [Online]. Available: http://www.sciencedirect.
[9] S. Impram, S. V. Nese, and B. Oral, ‘‘Challenges of renewable com/science/article/pii/B9780124167148000019
energy penetration on power system flexibility: A survey,’’ Energy [25] K. V. G. Raghavendra, K. Zeb, A. Muthusamy, T. N. V. Krishna,
Strategy Rev., vol. 31, Sep. 2020, Art. no. 100539, doi: 10.1016/j.esr. S. V. S. V. P. Kumar, D.-H. Kim, M.-S. Kim, H.-G. Cho, and H.-J. Kim,
2020.100539. ‘‘A comprehensive review of DC–DC converter topologies and modulation
[10] J. Gao, H. Wang, and H. Shen, ‘‘Smartly handling renewable strategies with recent advances in solar photovoltaic systems,’’ Electronics,
energy instability in supporting a cloud datacenter,’’ in Proc. IEEE vol. 9, no. 1, p. 31, Dec. 2019, doi: 10.3390/electronics9010031.
Int. Parallel Distrib. Process. Symp. (IPDPS), New Orleans, LA, [26] S. Ghazi and K. Ip, ‘‘The effect of weather conditions on the efficiency of
USA, May 2020, pp. 769–778. Accessed: Dec. 12, 2022. [Online]. PV panels in the southeast of U.K.,’’ Renew. Energy, vol. 69, pp. 50–59,
Available: https://ieeexplore.ieee.org/document/9139872/, doi: 10.1109/ Sep. 2014, doi: 10.1016/j.renene.2014.03.018.
IPDPS47924.2020.00084. [27] T. AlSkaif, S. Dev, L. Visser, M. Hossari, and W. van Sark, ‘‘A sys-
[11] A. Dobos, ‘‘PVWatts Version 5 manual,’’ Tech. Rep. NREL/TP-6A20- tematic analysis of meteorological variables for PV output power
62641, Sep. 2014, Art. no. 1158421. Accessed: Mar. 14, 2023. [Online]. estimation,’’ Renew. Energy, vol. 153, pp. 12–22, Jun. 2020, doi:
Available: http://www.osti.gov/servlets/purl/1158421/ 10.1016/j.renene.2020.01.150.
[12] D. Yang, J. Kleissl, C. A. Gueymard, H. T. C. Pedro, and C. F. [28] M. Massaoudi, I. Chihi, H. Abu-Rub, S. S. Refaat, and F. S. Oueslati,
M. Coimbra, ‘‘History and trends in solar irradiance and PV power ‘‘Convergence of photovoltaic power forecasting and deep learning: State-
forecasting: A preliminary assessment and review using text mining,’’ of-art review,’’ IEEE Access, vol. 9, pp. 136593–136615, 2021, doi:
Sol. Energy, vol. 168, pp. 60–101, Jul. 2018, doi: 10.1016/j.solener. 10.1109/ACCESS.2021.3117004.
2017.11.023. [29] U. K. Das, K. S. Tey, M. Seyedmahmoudian, S. Mekhilef, M. Y. I. Idris,
[13] S. Leva, A. Dolara, F. Grimaccia, M. Mussetta, and E. Ogliari, ‘‘Analysis W. Van Deventer, B. Horan, and A. Stojcevski, ‘‘Forecasting of
and validation of 24 hours ahead neural network forecasting of photovoltaic photovoltaic power generation and model optimization: A review,’’
output power,’’ Math. Comput. Simul., vol. 131, pp. 88–100, Jan. 2017, doi: Renew. Sustain. Energy Rev., vol. 81, pp. 912–928, Jan. 2018, doi:
10.1016/j.matcom.2015.05.010. 10.1016/j.rser.2017.08.017.
[14] D. V. Pombo, H. W. Bindner, S. V. Spataru, P. E. Sørensen, and P. Bacher, [30] M. Rana, I. Koprinska, and V. G. Agelidis, ‘‘Univariate and multi-
‘‘Increasing the accuracy of hourly multi-output solar power forecast variate methods for very short-term solar photovoltaic power forecast-
with physics-informed machine learning,’’ Sensors, vol. 22, no. 3, p. 749, ing,’’ Energy Convers. Manag., vol. 121, pp. 380–390, Aug. 2016, doi:
Jan. 2022, doi: 10.3390/s22030749. 10.1016/j.enconman.2016.05.025.

40842 VOLUME 11, 2023

J. Gaboitaolelwe et al.: Machine Learning Based Solar Photovoltaic Power Forecasting: A Review and Comparison

[31] P. Gupta and R. Singh, ‘‘PV power forecasting based on data-driven [49] W. El-Baz, P. Tzscheutschler, and U. Wagner, ‘‘Day-ahead
models: A review,’’ Int. J. Sustain. Eng., vol. 14, no. 6, pp. 1733–1755, probabilistic PV generation forecast for buildings energy management
Nov. 2021, doi: 10.1080/19397038.2021.1986590. systems,’’ Sol. Energy, vol. 171, pp. 478–490, Sep. 2018, doi:
[32] M. Ryo and M. C. Rillig, ‘‘Statistically reinforced machine learning for 10.1016/j.solener.2018.06.100.
nonlinear patterns and variable interactions,’’ Ecosphere, vol. 8, no. 11, [50] J. Brownlee, Ensemble Learning Algorithms With Python: Make Better
Nov. 2017, Art. no. e01976, doi: 10.1002/ecs2.1976. Predictions With Bagging, Boosting, and Stacking. San Francisco, CA,
[33] A. Géron, Hands-On Machine Learning With Scikit-Learn, Keras, and USA: Machine Learning Mastery, 2021, p. 450. [Online]. Available:
TensorFlow: Concepts, Tools, and Techniques to Build Intelligent Systems, https://books.google.co.bw/books?id=IUkrEAAAQBAJ
2nd ed. Beijing, China: O’Reilly Media, 2019. [51] L. Visser, T. AlSkaif, and W. van Sark, ‘‘Benchmark analysis
[34] Z. Razaghi-Moghadam and Z. Nikoloski, ‘‘Supervised learning of gene of day-ahead solar power forecasting techniques using weather
regulatory networks,’’ Current Protocols Plant Biol., vol. 5, no. 2, predictions,’’ in Proc. IEEE 46th Photovolt. Spec. Conf. (PVSC),
Jun. 2020, Art. no. e20106. Accessed: Aug. 18, 2021. [Online]. Chicago, IL, USA, Jun. 2019, pp. 2111–2116. Accessed: Nov. 22, 2022.
Available: https://onlinelibrary.wiley.com/doi/10.1002/cppb.20106, doi: [Online]. Available: https://ieeexplore.ieee.org/document/8980899/, doi:
10.1002/cppb.20106. 10.1109/PVSC40753.2019.8980899.
[35] D. C. Pagan, T. Q. Phan, J. S. Weaver, A. R. Benson, and A. J. Beaudoin, [52] C. L. Dewangan, S. N. Singh, and S. Chakrabarti, ‘‘Combining forecasts
‘‘Unsupervised learning of dislocation motion,’’ Acta Mater., vol. 181, of day-ahead solar power,’’ Energy, vol. 202, Jul. 2020, Art. no. 117743,
pp. 510–518, Dec. 2019, doi: 10.1016/j.actamat.2019.10.011. doi: 10.1016/j.energy.2020.117743.
[36] J. E. Van Engelen and H. H. Hoos, ‘‘A survey on semi-supervised [53] D. H. Wolpert, ‘‘Stacked generalization,’’ Neural Netw., vol. 5, no. 2,
learning,’’ Mach. Learn., vol. 109, no. 2, pp. 373–440, Feb. 2020, doi: pp. 241–259, Jan. 1992, doi: 10.1016/S0893-6080(05)80023-1.
10.1007/s10994-019-05855-6. [54] L. Breiman, ‘‘Random forests,’’ Mach. Learn., vol. 45, no. 1, pp. 5–32,
[37] B. W. Silverman and M. C. Jones, ‘‘E. Fix and J. L. Hodges (1951): 2001, doi: 10.1023/A:1010933404324.
An important contribution to nonparametric discriminant analysis and [55] N. J. Sairamya, L. Susmitha, S. T. George, and M. S. P. Subathra, ‘‘Hybrid
density estimation: Commentary on fix and hodges (1951),’’ Int. Stat. Rev. approach for classification of electroencephalographic signals using time-
Rev. Int. Stat., vol. 57, no. 3, p. 233, Dec. 1989, doi: 10.2307/1403796. frequency images with wavelets and texture features,’’ in Intelligent Data
Analysis for Biomedical Applications. Amsterdam, The Netherlands: Else-
[38] T. Cover and P. Hart, ‘‘Nearest neighbor pattern classification,’’ IEEE
vier, 2019, pp. 253–273. Accessed: Aug. 17, 2021. [Online]. Available:
Trans. Inf. Theory, vol. IT-13, no. 1, pp. 21–27, Jan. 1967, doi:
https://linkinghub.elsevier.com/retrieve/pii/B9780128155530000136
10.1109/TIT.1967.1053964.
[56] C. Hamzacebi, ‘‘Improving artificial neural networks’ performance in
[39] H. T. C. Pedro and C. F. M. Coimbra, ‘‘Nearest-neighbor methodol-
seasonal time series forecasting,’’ Inf. Sci., vol. 178, no. 23, pp. 4550–4559,
ogy for prediction of intra-hour global horizontal and direct normal
Dec. 2008, doi: 10.1016/j.ins.2008.07.024.
irradiances,’’ Renew. Energy, vol. 80, pp. 770–782, Aug. 2015, doi:
10.1016/j.renene.2015.02.061. [57] Ö. Çelik, A. Teke, and H. B. Yıldırım, ‘‘The optimized artificial neural
network model with Levenberg–Marquardt algorithm for global solar radi-
[40] F. Wang, Z. Zhen, B. Wang, and Z. Mi, ‘‘Comparative study on KNN
ation estimation in eastern Mediterranean region of Turkey,’’ J. Cleaner
and SVM based weather classification models for day ahead short term
Prod., vol. 116, pp. 1–12, Mar. 2016, doi: 10.1016/j.jclepro.2015.12.082.
solar PV power forecasting,’’ Appl. Sci., vol. 8, no. 1, p. 28, Dec. 2017,
[58] B. G. M. Vandeginste, D. L. Massart, L. M. C. Buydens, S. De Jong,
doi: 10.3390/app8010028.
P. J. Lewi, and J. Smeyers-Verbeke, ‘‘Artificial neural networks,’’ in Data
[41] C. Cortes and V. Vapnik, ‘‘Support-vector networks,’’ Mach. Learn., Handling in Science and Technolo, vol. 20. Amsterdam, The Netherlands:
vol. 20, no. 3, pp. 273–297, Sep. 1995, doi: 10.1007/BF00994018. Elsevier, 1998, pp. 649–699. Accessed: Aug. 17, 2021. [Online]. Avail-
[42] S. Chatterjee, D. Dey, and S. Munshi, ‘‘Feature selection and classifi- able: https://linkinghub.elsevier.com/retrieve/pii/S0922348798800543
cation,’’ in Recent Trends in Computer-Aided Diagnostic Systems for [59] J. Díaz-Gómez, A. Parrales, A. Álvarez, S. Silva-Martínez, D. Colorado,
Skin Diseases. Amsterdam, The Netherlands: Elsevier, 2022, pp. 95–135. and J. A. Hernández, ‘‘Prediction of global solar radiation by artificial
Accessed: Oct. 28, 2022. [Online]. Available: https://linkinghub. neural network based on a meteorological environmental data,’’ Desali-
elsevier.com/retrieve/pii/B9780323912112000019 nation Water Treatment, vol. 55, no. 12, pp. 3210–3217, Sep. 2015, doi:
[43] K. Mohammadi, S. Shamshirband, M. H. Anisi, K. A. Alam, 10.1080/19443994.2014.939861.
and D. Petković, ‘‘Support vector regression based prediction [60] L. Zou, L. Wang, A. Lin, H. Zhu, Y. Peng, and Z. Zhao, ‘‘Estimation
of global solar radiation on a horizontal surface,’’ Energy of global solar radiation using an artificial neural network based on an
Convers. Manage., vol. 91, pp. 433–441, Feb. 2015, doi: 10.1016/ interpolation technique in southeast China,’’ J. Atmos. Solar-Terr. Phys.,
j.enconman.2014.12.015. vol. 146, pp. 110–122, Aug. 2016, doi: 10.1016/j.jastp.2016.05.013.
[44] M. Pierro, F. Bucci, M. De Felice, E. Maggioni, D. Moser, A. Perotto, [61] P. A. C. Rocha, J. L. Fernandes, A. B. Modolo, R. J. P. Lima,
F. Spada, and C. Cornaroa, ‘‘Multi-model ensemble for day ahead pre- M. E. V. da Silva, and C. A. D. Bezerra, ‘‘Estimation of daily, weekly and
diction of photovoltaic power generation,’’ Solar Energy, vol. 134, monthly global solar radiation using ANNs and a long data set: A case
pp. 132–146, Sep. 2016, doi: 10.1016/j.solener.2016.04.040. study of Fortaleza, in Brazilian Northeast region,’’ Int. J. Energy Environ.
[45] W. VanDeventer, E. Jamei, G. S. Thirunavukkarasu, M. Eng., vol. 10, no. 3, pp. 319–334, Sep. 2019, doi: 10.1007/s40095-019-
Seyedmahmoudian, T. K. Soon, B. Horan, S. Mekhilef, and A. 0313-0.
Stojcevski, ‘‘Short-term PV power forecasting using hybrid GASVM [62] M. Ding, L. Wang, and R. Bi, ‘‘An ANN-based approach for forecasting
technique,’’ Renew. Energy, vol. 140, pp. 367–379, Sep. 2019, doi: the power output of photovoltaic system,’’ Proc. Environ. Sci., vol. 11,
10.1016/j.renene.2019.02.087. pp. 1308–1315, Jan. 2011, doi: 10.1016/j.proenv.2011.12.196.
[46] S. Qijun, L. Fen, Q. Jialin, Z. Jinbin, and C. Zhenghong, [63] A. Nespoli, S. Leva, M. Mussetta, and E. G. C. Ogliari, ‘‘A selective
‘‘Photovoltaic power prediction based on principal component ensemble approach for accuracy improvement and computational load
analysis and support vector machine,’’ in Proc. IEEE Innov. reduction in ANN-based PV power forecasting,’’ IEEE Access, vol. 10,
Smart Grid Technol.-Asia (ISGT-Asia), Melbourne, VIC, Australia, pp. 32900–32911, 2022, doi: 10.1109/ACCESS.2022.3158364.
Nov. 2016, pp. 815–820. Accessed: Mar. 17, 2023. [Online]. [64] A. Rezrazi, S. Hanini, and M. Laidi, ‘‘An optimisation methodology of
Available: http://ieeexplore.ieee.org/document/7796490/, doi: 10.1109/ artificial neural network models for predicting solar radiation: A case
ISGT-Asia.2016.7796490. study,’’ Theor. Appl. Climatol., vol. 123, nos. 3–4, pp. 769–783, Feb. 2016,
[47] C.-M. Huang, Y.-C. Huang, and K.-Y. Huang, ‘‘A hybrid method for one- doi: 10.1007/s00704-015-1398-x.
day ahead hourly forecasting of PV power output,’’ in Proc. 9th IEEE [65] C. Paoli, C. Voyant, M. Muselli, and M.-L. Nivet, ‘‘Forecasting
Conf. Ind. Electron. Appl., Hangzhou, China, Jun. 2014, pp. 526–531. of preprocessed daily solar radiation time series using neural net-
Accessed: Mar. 17, 2023. [Online]. Available: https://ieeexplore.ieee. works,’’ Sol. Energy, vol. 84, no. 12, pp. 2146–2160, Dec. 2010, doi:
org/document/6931220/, doi: 10.1109/ICIEA.2014.6931220. 10.1016/j.solener.2010.08.011.
[48] M. Pan, C. Li, R. Gao, Y. Huang, H. You, T. Gu, and F. Qin, ‘‘Pho- [66] S. Pretto, E. Ogliari, A. Niccolai, and A. Nespoli, ‘‘A new probabilistic
tovoltaic power forecasting based on a support vector machine with ensemble method for an enhanced day-ahead PV power forecast,’’ IEEE
improved ant colony optimization,’’ J. Cleaner Prod., vol. 277, Dec. 2020, J. Photovolt., vol. 12, no. 2, pp. 581–588, Mar. 2022, doi: 10.1109/JPHO-
Art. no. 123948, doi: 10.1016/j.jclepro.2020.123948. TOV.2021.3138223.

VOLUME 11, 2023 40843

 J. Gaboitaolelwe et al.: Machine Learning Based Solar Photovoltaic Power Forecasting: A Review and Comparison

[67] S. Theocharides, G. Makrides, A. Livera, M. Theristis, P. Kaimakis, [85] M. Figueiredo, B. Ribeiro, and A. de Almeida, ‘‘Analysis of trends in
and G. E. Georghiou, ‘‘Day-ahead photovoltaic power production fore- seasonal electrical energy consumption via non-negative tensor factor-
casting methodology based on machine learning and statistical post- ization,’’ Neurocomputing, vol. 170, pp. 318–327, Dec. 2015. Accessed:
processing,’’ Appl. Energy, vol. 268, Jun. 2020, Art. no. 115023, doi: Jul. 13, 2022. [Online]. Available: https://linkinghub.elsevier.com/
10.1016/j.apenergy.2020.115023. retrieve/pii/S0925231215008620, doi: 10.1016/j.neucom.2015.03.088.
[68] P. Ramsami and V. Oree, ‘‘A hybrid method for forecasting the energy [86] D. L. Marino, K. Amarasinghe, and M. Manic, ‘‘Building energy load
output of photovoltaic systems,’’ Energy Convers. Manag., vol. 95, forecasting using deep neural networks,’’ in Proc. 42nd Annu. Conf. IEEE
pp. 406–413, May 2015, doi: 10.1016/j.enconman.2015.02.052. Ind. Electron. Soc. (IECON), Florence, Italy, Oct. 2016, pp. 7046–7051.
[69] A. Mohammed, W. Yaqub, and Z. Aung, ‘‘Probabilistic forecasting Accessed: Jul. 13, 2022. [Online]. Available: http://ieeexplore.ieee.org/
of solar power: An ensemble learning approach,’’ in Proc. Int. Conf. document/7793413/, doi: 10.1109/IECON.2016.7793413.
Intell. Decis. Technol., Jun. 2015, pp. 449–458, doi: 10.1007/978-3- [87] R. Bonetto and M. Rossi, ‘‘Machine learning approaches to energy con-
319-19857-6_38. sumption forecasting in households,’’ 2017, arXiv:1706.09648.
[70] S. Lu, Y. Hwang, I. Khabibrakhmanov, F. J. Marianno, X. Shao, J.
Zhang, B.-M. Hodge, and H. F. Hamann, ‘‘Machine learning based
multi-physical-model blending for enhancing renewable energy forecast—
Improvement via situation dependent error correction,’’ in Proc. Eur.
Control Conf. (ECC), Jul. 2015, pp. 283–290, doi: 10.1109/ECC. JWAONE GABOITAOLELWE (Member, IEEE)
2015.7330558. received the B.Eng. degree in computer engineer-
[71] K.-P. Lin and P.-F. Pai, ‘‘Solar power output forecasting using evo- ing and the M.Eng. degree in computer engineer-
lutionary seasonal decomposition least-square support vector regres- ing and telecommunications from the Botswana
sion,’’ J. Cleaner Prod., vol. 134, pp. 456–462, Oct. 2016, doi: International University of Science and Technol-
10.1016/j.jclepro.2015.08.099.
ogy (BIUST), Palapye, Botswana, in 2018 and
[72] J. Huang and M. Perry, ‘‘A semi-empirical approach using gradient boost- 2020, respectively, where he is currently pursuing
ing and k-nearest neighbors regression for GEFCom2014 probabilistic
the Ph.D. degree with the Department of Electrical,
solar power forecasting,’’ Int. J. Forecasting, vol. 32, no. 3, pp. 1081–1086,
Computer and Telecommunications Engineering.
Jul. 2016, doi: 10.1016/j.ijforecast.2015.11.002.
[73] S. Alessandrini, L. D. Monache, S. Sperati, and G. Cervone, ‘‘An
analog ensemble for short-term probabilistic solar power forecast,’’
Appl. Energy, vol. 157, pp. 95–110, Nov. 2015, doi: 10.1016/j.apenergy.
2015.08.011.
[74] M. P. Almeida, O. Perpiñán, and L. Narvarte, ‘‘PV power forecast using a
nonparametric PV model,’’ Sol. Energy, vol. 115, pp. 354–368, May 2015,
doi: 10.1016/j.solener.2015.03.006.
[75] M. De Felice, M. Petitta, and P. M. Ruti, ‘‘Short-term predictability of
photovoltaic production over Italy,’’ Renew. Energy, vol. 80, pp. 197–204,
Aug. 2015, doi: 10.1016/j.renene.2015.02.010.
[76] M. Zamo, O. Mestre, P. Arbogast, and O. Pannekoucke, ‘‘A bench-
mark of statistical regression methods for short-term forecasting of
photovoltaic electricity production, Part I: Deterministic forecast of
hourly production,’’ Sol. Energy, vol. 105, pp. 792–803, Jul. 2014, doi:
10.1016/j.solener.2013.12.006.
[77] Y. Li, Y. Su, and L. Shu, ‘‘An ARMAX model for forecasting the power ADAMU MURTALA ZUNGERU (Senior Mem-
output of a grid connected photovoltaic system,’’ Renew. Energy, vol. 66, ber, IEEE) received the B.Eng. degree from the
pp. 78–89, Jun. 2014, doi: 10.1016/j.renene.2013.11.067. Federal University of Technology, Minna, Nigeria,
[78] H. Long, Z. Zhang, and Y. Su, ‘‘Analysis of daily solar power predic- the M.Sc. degree from Ahmadu Bello University,
tion with data-driven approaches,’’ Appl. Energy, vol. 126, pp. 29–37, Zaria, Nigeria, and the Ph.D. degree from Notting-
Aug. 2014, doi: 10.1016/j.apenergy.2014.03.084. ham University, U.K. He was a Research Fellow
[79] A. Mellit, A. M. Pavan, and V. Lughi, ‘‘Short-term forecasting of power with the Massachusetts Institute of Technology
production in a large-scale photovoltaic plant,’’ Sol. Energy, vol. 105, (MIT), USA, where he also obtained a Postgradu-
pp. 401–413, Jul. 2014, doi: 10.1016/j.solener.2014.03.018. ate Teaching Certificate, in 2014. He is currently a
[80] M. Zamo, O. Mestre, P. Arbogast, and O. Pannekoucke, ‘‘A bench- Professor and the Head of the Department of Elec-
mark of statistical regression methods for short-term forecasting of trical, Computer, and Telecommunications Engineering, Botswana Inter-
photovoltaic electricity production. Part II: Probabilistic forecast of national University of Science and Technology (BIUST). Before joining
daily production,’’ Sol. Energy, vol. 105, pp. 804–816, Jul. 2014, doi: BIUST, in 2015, he was a Senior Lecturer and the Head of the Electrical and
10.1016/j.solener.2014.03.026. Electronics Engineering Department, Federal University Oye-Ekiti, Nigeria.
[81] M. Gao, J. Li, F. Hong, and D. Long, ‘‘Day-ahead power forecast- He is also a registered Engineer with the Council for The Regulation of
ing in a large-scale photovoltaic plant based on weather classifica- Engineering in Nigeria (COREN), a registered Professional Engineer with
tion using LSTM,’’ Energy, vol. 187, Nov. 2019, Art. no. 115838, doi:
the Botswana Engineers Registration Board (ERB), and an Association
10.1016/j.energy.2019.07.168.
for Computing Machinery (ACM) in the USA. He is an inventor of a
[82] A. Abdellatif, H. Mubarak, S. Ahmad, T. Ahmed, G. M. Shafiullah,
Termite-Hill routing algorithm for wireless sensor networks, a co-inventor of
A. Hammoudeh, H. Abdellatef, M. M. Rahman, and H. M. Gheni,
method and system for sorting of diamonds, and has three of his patent appli-
‘‘Forecasting photovoltaic power generation with a stacking ensem-
ble model,’’ Sustainability, vol. 14, no. 17, p. 11083, Sep. 2022, doi: cations registered with the world intellectual property organization (WIPO).
10.3390/su141711083. He has also authored five academic books and over 60 international research
[83] W. Khan, S. Walker, and W. Zeiler, ‘‘Improved solar photovoltaic articles in reputable journals, including the IEEE SYSTEMS JOURNAL, IEEE
energy generation forecast using deep learning-based ensemble stack- INTERNET OF THINGS JOURNAL, IEEE ACCESS, and JNCA (Elsevier), with over
ing approach,’’ Energy, vol. 240, Feb. 2022, Art. no. 122812, doi: 1600 citations and an H-index of 18. He is also serving as an Associate Editor
10.1016/j.energy.2021.122812. for IEEE ACCESS. He has also served as an international reviewer for IEEE
[84] L. Visser, B. Elsinga, T. AlSkaif, and W. van Sark, ‘‘Open-source quality TRANSACTIONS ON INDUSTRIAL INFORMATICS, IEEE SENSORS, IEEE ACCESS, IEEE
control routine and multi-year power generation data of 175 PV systems,’’ TRANSACTIONS ON MOBILE COMPUTING, and IEEE TRANSACTIONS SUSTAINABLE
Zenodo, Aug. 1, 2022. Accessed: Oct. 14, 2022. [Online]. Available: COMPUTING, and JNCA (Elsevier). He has also served as the Chairperson of
https://zenodo.org/record/6906504 the IEEE Botswana Sub-Section (2019–2020).

40844 VOLUME 11, 2023

J. Gaboitaolelwe et al.: Machine Learning Based Solar Photovoltaic Power Forecasting: A Review and Comparison

ABID YAHYA (Senior Member, IEEE) received DASARI NAGA VINOD (Member, IEEE)
the bachelor’s degree in electrical and electronic received the Bachelor of Technology degree in
engineering majoring in telecommunication from electronics and communication engineering from
the University of Engineering and Technology, Jawaharlal Nehru Technological University, Kak-
Peshawar, Pakistan, and the M.Sc. and Ph.D. inada, Andhra Pradesh, in 2016, the Master of
degrees in wireless and mobile systems from Uni- Technology degree in embedded systems from
versiti Sains Malaysia, Malaysia. He began his Amrita University, Bengaluru, Karnataka, in 2018,
career on an engineering path, which is rare among with a Graduate Aptitude Test in Engineering
other researcher executives. Currently, he is with (GATE) scholarship, and the Ph.D. degree in elec-
the Botswana International University of Science tronics and communication engineering with a
and Technology. He is also a registered Professional Engineer with the GATE scholarship, in 2022. After his highly commendable Ph.D. dissertation
Botswana Engineers Registration Board (ERB). He has many research pub- from the SRM Institute of Science and Technology, Chennai, India. He was
lications to his credit in numerous reputable journals, conference articles, a recipient of GATE score, in 2016, and obtained 84 percentiles. Since
and book chapters. He has received several awards and grants from various February 2022, he has been an Assistant Professor with the Department
funding agencies and supervised several Ph.D. and master’s candidates. His of Electronics and Communication Engineering, Vel Tech Rangarajan Dr.
recent four books, such as Emerging Technologies in Agriculture, Livestock, Sagunthala R&D Institute of Science and Technology, Avadi, Chennai.
and Climate (Springer, 2020), Mobile WiMAX Systems: Performance Anal- He has published two papers in SCI-indexed journals, presented three papers
ysis of Fractional Frequency Reuse (CRC Press | Taylor & Francis, 2019), in international conferences, and has an H-index of three with 68 citations.
Steganography Techniques for Digital Images, LTE-A Cellular Networks: His research interests include the Internet of Things, image processing,
Multi-Hop Relay for Coverage, Capacity, and Performance Enhancement medical image processing, machine learning, and deep learning.
(Springer, July 2018 January 2017), and are being followed in national and
international universities.

AYODEJI OLALEKAN SALAU (Member, IEEE)

received the B.Eng. degree in electrical/computer
engineering from the Federal University of Tech-
nology, Minna, Nigeria, and the M.Sc. and Ph.D.
degrees in electronic and electrical engineering
from Obafemi Awolowo University, Ile-Ife, Nige-
ria. He is currently with the Department of Elec-
trical/Electronics and Computer Engineering, Afe
Babalola University, Ado Ekiti (ABUAD), Nige-
ria. He is also a registered Professional Engineer
with the Council for the Regulation of Engineering in Nigeria (COREN) and
a member of the International Association of Engineers (IAENG). He serves
as a reviewer for numerous reputable international journals and the co-chair
and a technical program committee member for various conferences. He has
authored/coauthored about 125 articles in several reputable international
conferences, book series, and major international journals. He has authored
three books. His research interests include research in the fields of computer
vision, image processing, signal processing, machine learning, and power
systems engineering. He was a recipient of the Quarterly Franklin Member-
CASPAR K. LEBEKWE (Member, IEEE) received ship given, in 2020, by the editorial board of London Journals Press for top
the M.Eng. degree in electronics and communi- quality research output. In addition, his paper was awarded the best paper of
cations engineering and the Ph.D. degree in elec- the year 2019 in cogent engineering. Furthermore, he was a recipient of the
trical and electronics engineering, sponsored by International Research Award on New Science Inventions (NESIN) under
the General Lighthouse Authorities from the Uni- the category of ‘‘Best Researcher Award’’ given by Science Father, in 2020.
versity of Bath, in 2008. His Ph.D. project was He was also a recipient of the International Best Researcher Award given
focused on eLoran Service Volume Coverage Pre- by the ISSN International Science & Technology Awarding body (IISTAC-
diction. He is currently a Senior Lecturer with the 2022), in 2022. He was also awarded the Young Scientist Award by the
Botswana International University of Science and International Research Awards on Science, Technology and Management,
Technology, where he teaches optical communica- INSO, in 2022. Also, he was awarded the Best Researcher Award (2020–
tions, antennas and propagation, discrete mathematics, telemetry, and remote 2022) by the Chancellor of ABUAD.
control and electromagnetic field theory.

VOLUME 11, 2023 40845

View publication stats