Journal of Electrical Engineering & Technology (2023) 18:719–733

https://doi.org/10.1007/s42835-023-01378-2

ORIGINAL ARTICLE

A Comprehensive Review on Ensemble Solar Power Forecasting

Algorithms
Negar Rahimi1 · Sejun Park1 · Wonseok Choi1 · Byoungryul Oh1 · Sookyung Kim1 · Young‑ho Cho1 · Sunghyun Ahn1 ·
Chulho Chong1 · Daewon Kim1 · Cheong Jin2 · Duehee Lee1

Received: 29 June 2022 / Revised: 25 December 2022 / Accepted: 3 January 2023 / Published online: 12 January 2023
© The Author(s) 2023, corrected publication 2023

Abstract
With increasing demand for energy, the penetration of alternative sources such as renewable energy in power grids has
increased. Solar energy is one of the most common and well-known sources of energy in existing networks. But because of
its non-stationary and non-linear characteristics, it needs to predict solar irradiance to provide more reliable Photovoltaic
(PV) plants and manage the power of supply and demand. Although there are various methods to predict the solar irradiance.
This paper gives the overview of recent studies with focus on solar irradiance forecasting with ensemble methods which are
divided into two main categories: competitive and cooperative ensemble forecasting. In addition, parameter diversity and data
diversity are considered as competitive ensemble forecasting and also preprocessing and post-processing are as cooperative
ensemble forecasting. All these ensemble forecasting methods are investigated in this study. In the end, the conclusion has
been drawn and the recommendations for future studies have been discussed.

Keywords Ensemble methods · Solar forecasting · Cooperative ensemble forecasting

1 Introduction countries with the highest installed solar PV capacity are

shown in Fig. 1 [3].
Over the past decades, the installation and use of renewable According to Renewables 2015-Global Status Report, the
energy resources have been increased due to environmental global capacity of PV was increased from 3.7 GW to 7 GW
issues such as greenhouse gas emission and the reduction between 2004 and 2007 and promoted to 40 GW by 2010. In
of fossil fuels [1]. Solar energy has been considered as one 2015, the global installed capacity of solar PV was around
of the promising renewable energy resources that is able to 177 GW [4]. From 2016 to 2018, based on IRENA, there
supply global energy demand [2]. According to the Inter- was a significant rose in this amount, and it was reached
national Renewable Energy Agency (IRENA), the top ten about 580 GW in 2019 [5]. It has been reported that in the

* Duehee Lee Sunghyun Ahn

[email protected] [email protected]
Negar Rahimi Chulho Chong
[email protected] [email protected]
Sejun Park Daewon Kim
[email protected] [email protected]
Wonseok Choi Cheong Jin
[email protected] [email protected]
Byoungryul Oh 1
Deptartment of Electrical and Electronic Engineering,
[email protected]
Konkuk University, Seoul, South Korea
Sookyung Kim 2
EINS S&C, Seoul, South Korea
[email protected]
Young‑ho Cho
[email protected]

13
Vol.:(0123456789)
720 Journal of Electrical Engineering & Technology (2023) 18:719–733

10.505 due to the substantial increase of solar power genera-

10.562 tion worldwide [9].
China
13.398 Recently, the ensemble forecasting was recommended for
15.928
Japan solar power forecasting. In the ensemble forecasting, many
20.9 different predictions from different forecasting are aver-
USA aged. Averaging predictions can reduce server biases when
34.831
205.072
weather data in outliers, so it can avoid the worst predic-
tions. However, it takes a lot of computational time to run
48.96 Germ
the ensemble forecasting model since several forecasting
any models should be simulated at the same time. However, this
India computational time can be further reduced when the parallel
60.54
computing environment is used. In this study, we organize
61.84
several ensemble solar power forecasting algorithms.
For forecasting methods of PV systems, several review
Fig. 1  Top 10 countries in 2019 based on total PV installed capacity
papers have been published during the last 5 years with
(GW) [3]
different scopes. Their focus was ensemble methods, PV
output power forecasting different PV forecasting methods,
first half of 2020, when the COVID-19 pandemic began, the probabilistic forecasting in solar PV [10], hybrid models
renewable energy sector experienced a turbulent but in the for solar radiation forecasting, post-processing in solar fore-
second half of 2020, the PV sector set a new record for new casting, different methods for forecasting solar irradiance.
capacity which bring the solar sector to its highest level of Hence, there is no updated review with a focus on ensemble
more than 800 Gigawatts [6]. methods only. Therefore, we reviewed the ensemble meth-
However, the stochastic and unpredictable nature of solar ods for solar irradiation forecasting which are divided into
energy led to a number of challenges such as voltage fluctua- main categories competitive (data diversity and parameter
tions and uncertainties in a power grid that make it difficult diversity) and Cooperative forecasting (pre-processing and
to maintain a balance between power generation and load post-processing). The recent research papers that have been
demand [7]. For example, in cloudy days, the amount of selected which have been published in the last four years
solar irradiation received by PV modules has many fluc- (2018–2022) were reviewed. The above-mentioned papers
tuations because of the cloud’s movement, and the random focused on reviewing solar forecasting methods. In this
fluctuations have a significant effect on the solar PV output. paper, the focus was ensemble forecasting methods and their
Various solutions including power system, scheduling bat- classifications in recent years.
tery reserves, demand response, backup generators, and dis- For the ensemble forecasting, there are two topics,
patches have been proposed to prevent the above problems. namely, solar power forecasting and solar irradiance fore-
However, several limitations exist in all these solutions. casting which are known as solar forecasting. Meanwhile,
There is a restriction for the power decline rate of backup they have strongly correlated each other and cannot be sepa-
generators which caused by the unit ramp rate resulted in rated [11]. It should be noted that there is a strong impact
difficulties to meet the incremental power generation need. of solar irradiance on the accuracy of solar power forecasts
However, due to the storage capacity restrictions and the for the power production systems of various sorts. If a fore-
costs of battery reserves, massive-scale energy storage is caster wants to achieve high-quality solar power forecasts,
still difficult to realize. In addition, there is a difficulty to the ability to produce and use irradiance forecasts is essen-
achieve demand response technologies based on the lack tial. In other words, the best solar power prediction is always
information on behavior of electricity consumption by con- obtained by irradiance forecasts generation and conversion
sumers. Besides, the efficiency of the above-mentioned solu- of those irradiance forecasts into power forecasts through
tions largely depends on forecasting accuracy (with different a model chain. Therefore, there is not an intrinsic value for
time horizons) [8]. irradiance forecast [12].
There are two important aspects of accurate forecasting:
reducing the negative effect of random PV power on the
power grid and providing and predicting PV power output
data for grid operators. Hence, there is a need to forecast the 2 Parameters Affecting Solar Power
output power of solar systems for the efficient operation of Forecasting
the power grid. The optimal management of a power system
and scheduling is important for estimation of the reserves. It Forecast horizon, weather classification, error metrics, data
should be noted that the solar forecasting becomes important processing affect the output of solar forecasting.

13
Journal of Electrical Engineering & Technology (2023) 18:719–733 721

2.1 Forecast Horizon appropriate forecast time horizon should be selected because

the accuracy of a predicted model depends on the forecast
The forecast horizon can be considered as the period of time time horizon [17]. It has been proved that by increasing the
in the future (time duration between actual and effective forecasting horizon, the accuracy of forecasting model (both
time) in which the forecasting should be done [13]. Forecast- single and ensemble models) will reduce. This is because
ing horizon can be classified into four categories including the correlation between cloud cover and solar irradiance,
(1) very short-term, (2) short-term, (3) medium-term, and which cannot be accurately predicted for long periods of
(4) long-term [14]. These categories have summarized in time. In addition, for power system planning, long time
Table 1 based on their forecasting period and related appli- horizon forecast is suitable while for PV output forecast-
cations in solar energy systems. However, there is still no ing, intra-hour and intra-day forecast horizons work better
universal classification criterion [15]. [13]. In other words, the model performance decreases with
Furthermore, the researchers usually prefer to use another long-time ahead forecasting (regardless of model type) while
categories to describe forecast horizon including intra-hour, the performance would be increased with short time ahead
intra-day, and day-ahead which have overlapping with short, forecasting [18].
medium, and long-term forecast horizons [16].
2.2 Weather Classification
2.1.1 Intra‑Hour
It is obvious that the PV output is directly related to solar
Intra-hour overlaps with very short-term and short-term irradiance and the accuracy solar irradiance forecasting
horizon categories and it also shows the forecast horizon models is strongly affected by meteorological factors such
from a few seconds to an hour. It is used for operating regu- as cloud cover, temperature, humidity and wind speed. So,
lation reserves, storage system optimization and ensuring climate change and different weather types have significant
grid quality and stability. Such prediction methods can be effect on PV system output power. Therefore, to enhance the
applied in high solar penetration areas such as island grids prediction performance, weather condition is an effective
with spans of 1 to 6 h. step especially for solar irradiance forecasting [19]. Typi-
cally, there are two types of weather condition including
2.1.2 Intra‑Day the normal (ideal) weather type (sunny days), and abnormal
(non-ideal) weather types (rainy, foggy and windy days)
This forecast horizon is used for 1–6 h and also overlaps [20]. According to some studies [19–21], the PV output
with short and medium categories and its application is in power increases in the ideal weather condition (on sunny
electricity trading outside the standard grid and control of days) but it decreases in non-ideal conditions.
electric loads [13].
2.3 Error Metrics
2.1.3 Day‑Ahead
In different steps of model development, evaluation error
Day-ahead Forecasts spanning 6–48 h overlaps with long- metrics are one of the important parameters. In these kinds
term and medium horizons. Similar models have been used of metrics, the comparison of the actual solar irradiance
in unit commitment and utilities planning [13]. In gen- and predicted solar irradiance are considered [22]. It should
eral, before designing PV power forecasting model, the be noticed that there are different units for performance of

Table 1  Different types of forecasting horizon and its applications [15]

Forecasting horizon Forecasting period Applications

Very short-term From seconds to 1 min ahead – Electricity market

– Monitoring of real-time electricity dispatch
– PV storage control
Short-term Several hours Unit commitment
ahead
Medium-term One week or within a month ahead Scheduling of power systems maintenance
(solar energy integrated power or conventional
systems)
Long-term Months or Years ahead – Electricity pricing
– Load forecasting

13
722 Journal of Electrical Engineering & Technology (2023) 18:719–733

√
metrics while W/m2 has been used as unit for the statistical √ N
√1 ∑ ( )2
error of solar radiation [18]. The most used evaluation met- RMSE = √ y − yi(act) (4)
rics for statistical measures are listed as below: N i=1 i(pred)

2.3.1 Mean Bias Error (MBE)

2.3.5 Normalized RMSE (nRMSE)
Mean bias error (MBE): It shows the average bias of a fore-
casting model: Normalized RMSE (nRMSE) is the overall deviations of
larger datasets nRMSE can be calculated.
N
1 ∑( ) �
MBE = yi(pred) − yi(act) (1) 1 ∑N �
yi(pred) − yi(act)
�2
N i=1 N i=1
(5)
nRMSE =
y
The larger value of MBE shows the larger forecast bias.
A positive value of MBE indicates over-forecasting, while a where ( y ) denotes the mean of the actual solar irradiance
negative value means under-forecasting [23]. [27].

2.3.2 Mean Absolute Error (MAE) 2.3.6 Mean Absolute Percentage Error (MAPE)

Mean absolute error (MAE): This error is defined as the Mean absolute percentage error (MAPE) is appropriate to
average of the absolute difference between forecasted and evaluate uniform prediction errors like MAE however it can
actual solar irradiance values. This metric is suitable for be calculated by the difference between each predicted and
uniform forecast errors as equal weight will be given to all actual observation divide by the actual observed value.
discrepancies in the data, and it is also used for both regres-
N |y |
sion problems and evaluation of overall forecast accuracy. y
1 ∑ || i(pred) − i(act) ||
Generally, the smaller MAE is better in forecasting. MAPE = (6)
N i=1 yi(act)
N
1 ∑| |
MAE = |y − yi(act) | (2)
N i=1 | i(pred) | 2.3.7 Determination Coefficient (R2)
where yi(act) is the actual solar irradiance, N is the total num-
Determination coefficient (R^2) is used to extract the infor-
ber of observations and yi(pred) is the predicted solar irradi-
mation of the determination between the forecasted and the
ance [22, 24].
real values and the ranges is between 0 and 1 [28].

2.3.3 Mean Square Error (MSE) 2.3.8 Skill Score (𝐒𝐒)

Mean square error (MSE): It is calculated by averaging the Skill score (SS) measures the performance comparison of a
square of difference between the actual and predicted solar forecast model to a benchmark model is defined as S score
irradiance values [25]. and calculated as follows:
N
1 ∑( )2 RMSEforecast model
MSE = yi(pred) − yi(act) (3) S=1− (7)
N i=1 RMSEbenchmark model

A good forecast model has a SS score of 1. In addition,

2.3.4 Root Mean Square Error (RMSE) SS zero shows a model with forecast error equal to the
benchmark model while a negative SS score is a forecast
Root mean square error (RMSE) is for calculating this met- with higher forecast error [1].
ric, the square root for the average of the squared differences
of predicted and actual solar irradiance values is considered. 2.4 Model Inputs
RMSE is known as the most appreciated performance evalu-
ation metrics which the outliers in the data can be elimi- Inputs are considered as a key factor in forecasting and have a
nated and identified by using this error [25]. Also, as RMSE significant effect on prediction accuracy. Generally, inappropriate
emphasizes the larger errors, it can be used as the main error inputs can cause forecast errors in a system, for example, time
metric [26]. delay, cost, and computational complexity [13]. In addition, the

13
Journal of Electrical Engineering & Technology (2023) 18:719–733 723

correlation of the input and output values affects the performance Group one is statistical models, which are based on his-
of a forecasting model. Therefore, the correlation of PV power torical data as input for prediction and the internal behav-
output with the different meteorological inputs, such as wind ior of the model does not have any effect on it [34]. In
speed and direction, module temperature, atmospheric tempera- other words, to reconstruct the hourly irradiance and met-
ture, solar irradiance, and humidity is important [17]. Among rological parameters statistical models can be used [35].
these parameters, solar irradiance is the most significant input Different techniques are applied for statistical time series
for the forecast and the accuracy of solar irradiance measurement such as support vector machine and Markov chain, regres-
affects the precision of solar power generation [29]. Demon- sion model, and artificial neural network [34].
strated the highest influence in solar power generation related to The second group generated solar power output from
the intensity of solar irradiance. In [30] a SVR-based forecasting external data such as temperature and solar irradiance and
model was proposed for PV power generation forecasting. In this their physical relationship. The generation of irradiance
study, the data of three different PV plants, in Malaysia, includ- forecast between many meteorological parameters would
ing the actual PV power generation data and meteorological data be achieved by using mathematical models. It can be said
(wind speed atmospheric temperature and solar irradiance) were that the results are highly dependent on the accurate input
used and PV power output was received only from 8:00 AM until data of the model [36]. This group can be categorized
19:00 PM. Also, PV output for a specific day and the pattern into two subsections of Empirical models and Numerical
of solar irradiance was presented and the correlation coefficient weather prediction (NWP) [37].
(R^2 = 0.9888) was obtained. The correlation coefficient indi- The last group is hybrid or ensemble methods which
cated that there is a strong correlation between solar irradiance used the combination of physical and statistical methods
and PV power output. It means that if PV power increases, then with unique features to solve the limitations of an indi-
the irradiance will be increased, and vice versa. Other studies vidual model. By these models, at the same time, the fore-
also showed that the PV power generation are correlated with casting performance increases, and the model would be
other meteorological variables, such as clouds, module tempera- improved. Figure 2 shows these forecasting methods [38].
ture, ambient temperature, and wind speed. Also, fluctuations in For forecasting methods of PV systems, several review
sunlight intensity caused by several parameters such as dissipa- papers have been published during the last 5 years with
tion, cloud motion, deformation and birth which affect PV output different scopes. Their focus was ensemble methods, PV
[31]. In another study by [32], PV output variation with solar output power forecasting [14, 32] different PV forecast-
irradiance was analyzed during clear day, partially cloudy and ing methods, probabilistic forecasting in solar PV [40],
cloudy day. Based on the results, in comparison with to partially hybrid models for solar radiation forecasting [41], post-
cloudy and clear day, there was a high fluctuations degree with processing in solar forecasting [42], different methods
solar irradiance were observed in cloudy days.
Another input variables linked together are wind speed,
ambient and module temperature. The temperature of the mod- Artificial Neural
Network
ule and its level of efficiency depend on the ambient tempera-
ture and the amount of absorbed radiation, as well as the wind
Support Vector
speed. When the ambient temperature increases (It is related Mashine
to cloud cover), the module temperature rises and its efficiency Based on Statistical
time series method
also increases, while the wind speed reduces the module tem- Markov Chain
perature and heat loss [33]. It is concluded that there was a Model
weak correlation between atmospheric temperature and PV
power output while there was an extremely weak correlation Autoregressive
Model
between wind speed and PV power output. In contrast, there
was a strong correlation between module temperature and PV Forecasting
techniques
power output [30]. NWP Model
Based on Physical
method
Empirical Model
3 Classification of Solar Forecasting
Algorithms
Combination of
Based on ensemble Statisticsl and
This section classifies solar forecasting algorithms. The method Physical method
solar forecasting algorithms are categorized into three
main models, such as ensemble methods physical and sta-
tistical time-series. Fig. 2  Forecasting techniques based on three major models [39]

13
724 Journal of Electrical Engineering & Technology (2023) 18:719–733

for forecasting solar irradiance [24]. Hence, there is no multiple predictors are used to reach an aggregated decision
updated review with a focus on ensemble methods only. in a better format of base predictors. The main advantage of
Therefore, we reviewed the ensemble methods for solar an ensemble model is that it incorporates the suitability of
irradiation forecasting which are divided into main catego- its constituent techniques, which creates a stronger learn-
ries competitive (data diversity and parameter diversity) ing pattern. So, it can enhance the accuracy and strengths
and Cooperative forecasting (pre-processing and post- of individual methods by solving their weakness [46]. The
processing). The recent research papers that have been ensemble methods can be classified into two main groups:
selected which have been published in the last four years competitive and cooperative [39]. In a competitive ensem-
(2018–2022) were reviewed. The above-mentioned papers ble forecasting, the different base predictors train individu-
focused on reviewing solar forecasting methods. In this ally with the same or different data sets by using different
paper, the focus was ensemble forecasting methods and parameters, and the prediction is obtained from the averag-
their classifications in recent years. ing the decisions of all base predictors [47]. Besides, in a
cooperative method, the prediction task will be divided into
several sub-tasks to select the appropriate predictors based
4 Ensemble Forecasting on the characterization of the sub-tasks. Then, the sum of all
the base predictors’ output will be the final decision of the
The main concept of the ensemble technique is training cooperative method [48]. Table 2 shows the classification of
ensemble members (base learners) and combining their pre- the ensemble methods and their perspectives in this paper.
diction into a single output to obtain a better performance
of a model [10]. Figure 3 shows the typical ensemble model 4.1 Competitive Forecasting Methods
construction. As can be seen in the Fig. 3, the training data
is divided into many data sets initially and then several base As mentioned earlier, a competitive forecast model can
learners are generated which can be run in a sequential or construct individual forecast models for the formation of an
parallel format. Finally, the combination of the models is run ensemble forecast model by using multiple predictors using
on the base learners [43, 44]. different initial conditions or parameters. Then, the results
On other words, Ensemble (hybrid, or combined tech- will be obtained from the average of the selected models
nique) [45] is a machine learning-based method in which [50]. The competitive ensemble methods are categorized
into three different perspectives which are data diversity,
parameter diversity and structural diversity [51]. Diversity
is the most important feature of competitive ensemble fore-
casting. For instance, the outputs of base predictors would
be similar if the sub-tasks are similar while the performance
improvement of ensemble predictor will be marginal [50].
Some examples of competitive ensemble forecasting models
are discussed in detail in the following sections.

4.1.1 Data Diversity

Data diversity is another category of cooperative ensemble

forecasting method in which the forecasting system is fed by
more than one input dataset [37]. As shown in the following
equations, two variations have been used in Eq. (9) and (10).
Fig. 3  General construction flow of an ensemble model Equation (9) applies N predictors ( f1 (0) … fN (0)) for N input

Table 2  The classification of Categories Perspectives Methods

ensemble forecasting models
[49] Competitive ensemble methods –Data diversity – Bagging, Boosting
–Parameter diversity – Ensemble Kalman filter
Cooperative ensemble methods –Pre-processing – Wavelet decomposition (WD),
–Post-processing empirical mode decomposition
(EMD)
– ARMA-GARCH, ARMA-ANN,
ARIMA-SVM

13
Journal of Electrical Engineering & Technology (2023) 18:719–733 725

datasets ( x1 … xN ) for forecasting and the weighted average samples are injected into each tree randomly). The output
of all of them will be the final prediction value, but Eq. (10) of RF is obtained by voting of the decision trees. In the case
uses only a single predictor for forecasting. of classification, the majority of voting is used to decide on
N
the predicted result, but in the case of regression, the mean
1∑ value of the predictions of all the estimators (predictors) are
ŷ (t + h) = w f (x (t)) (8)
N i=1 i i i calculated [57]. Figure 5 demonstrates the procedure of ran-
dom forest method graphically. This algorithm does not need
complex calculations like Support Vector Machine (SVM)
wi ≥ 0 (9)
and ANN because the main variables that should be adjusted
N
is the number of trees. Also, compared to ANN and SVM,
∑ training process is faster in the RF. In this algorithm, if suf-
wi = 1 (10)
i=1
ficient trees are used, it is robust to noise and outliers [58].
The application of this RF can be extended to PV produc-
tion forecasting. Tato et al. used real radiation measurements
ŷ (t + h) = f (x1 (t), x2 (t), ..., fN (t)) (11)
by combining simple radiation predictions to forecast the
where ̂y is the predicted value and h is the forecast horizon. solar energy output for short-term temporal horizons [60]. A
The most famous and common approaches are bagging and daily PV power generation forecasting model was proposed
boosting for data diversity [48]. for North China in winter. The proposed forecasting model
was based on the RF algorithm using weather measures [61].
4.1.1.1 Bagging For the first time, bagging or bootstrap The accuracy, extra trees (ET), computational cost, and sta-
aggregation is proposed by Breiman [52] which helps bility of RF were investigated for predicting hourly PV gen-
decrease the forecasting model variance and avoid overfit- eration output. In addition, their performance was compared
ting [51]. The main goal of Bagging is developing several with support vector regression (SVR).The performance
estimators, and results are obtained by aggregation of indi- evaluation of the model was performed by RMSE,MAE
vidual estimators' results with some biasing. Initially, from and R2 and it was concluded that RF performed better on
the original datasets, several subsets of training data are cre-
ated. Then, the samples are selected randomly by replace-
ment of data samples which are called Bootstrap samples. In
the last step, the final prediction is obtained by aggregation
of all the bootstrap predictions [53]. This process is shown
in the Fig. 4. One of the advantages of bagging is the error
reduction of ensemble generation in the baseline predictors.
It also can correlate its estimation with real datasets using
estimates of test sets or cross-validation [54].
The most common algorithm for bagging is random forest
(RF) which can be considered as an extension to the bagging
concept, and it can be used for classification and regres-
sion. It is made up and trained a large number of decision
trees (DT), called predictors and each one produces their
own predictions that can create higher accuracy in predic-
tion [55, 56]. In the RF algorithm, several decision trees are
constructed through training samples (a subset of training Fig. 5  The procedure of random forest method [59]

Fig. 4  Block diagram depicting

the basic principle of Bagging
ensemble learning

13
726 Journal of Electrical Engineering & Technology (2023) 18:719–733

both testing datasets and training in comparison with other based on physical properties and human intuition. They
models [62]. found a 6% increase in RMSE with apply the most important
features of the ensemble method [72]. The several tree-based
4.1.1.2 Boosting To solve the classification and regression methods of DT, GBM, XGBM, bagging, RF, and Cubist [73,
problems, boosting can be also used as a powerful learning 74] were used for forecasting solar irradiation on Jeju Island
strategy. It can be done by the combination of weak learners’ of South Korea. Also, the prediction performance was based
output into a ‘committee’. In addition, boosting improves on the comparison of MAE, RMSE, and nRMSE [75].
the suitability of model to data through bias reduction [63].
In boosting process, several subsets can be created from the 4.1.2 Parameter Diversity
original dataset. Then it sequentially trains the predictors
with the datasets, and at each iteration, smaller weights to In contrast to data diversity, parameter diversity is applied
the data points with smaller error and higher weights to different parameter settings with the same dataset (x). The
the data points with larger error can be assigned by boost- forecasts can be generated by using the following equation
ing. Finally, the results from the weighted average will be [71]:
obtained [64]. Different algorithms can be used in bagging
methods such as AdaBoost [65], extreme gradient boost- 1∑
N

ing (XGBM) [64], gradient boosting machine (GBM) [66, ŷ (t + h) = f (x , 𝜃 ) (12)

N i=1 i i i
67], and light gradient boosting machine (LGBM) [68]. The
recent advances in boosting algorithms and their applica- where h is the forecast horizon and ̂ y is the predicted
tions in energy research such as solar and wind have been value.𝜃i is the parameter for model fi , i = 1, … , N . It should
reviewed in [69]. They showed that how boosting algorithms be noted that the combination in Eq. (11) uses equal weight-
are effective tools in the performance of prediction models. ing. In solar forecasting studies, once the initial conditions
Recently published studies are using bagging and boost- need to be perturbed, parameter diversity is a useful method
ing methods for forecasting solar irradiation in different as it considers numerical weather prediction models (NWP)
regions. For instance, to improve forecasting accuracy in [76, 77]. Kalman filter is the best method for NWP which
solar energy output, the bagging model was used in [64]. estimates the “true” state of a dynamical system from noisy
The proposed model as an ensemble method, used a based measurement data [76]. The Kalman filter processes in two
learner such as random forest, LGBMs, and XGB by addi- stages including prediction and updating. In the first stage,
tion of the past output data as new features. They found the there is a current state vector contains one or more state vari-
bagging model was successful with higher model accuracy ables, and the state vector with a weighted average will be
using past data features in comparison with a single model updated in the second stage. In addition, without additional
learner with default features [70]. past information the Kalman filter can run in real-time due to
In another study [44], the solar irradiance in five cities of its recursive nature [78]. Recently, studies have been focused
Turkey, was estimated using bagging and boosting ensem- on solar forecasting using Kalman filtering.
bles of ANN, SVR, and DT. Initially, base models (ANN, Yang studied day-ahead NWP forecasts using Kalman fil-
SVR, and DT) were created and examined by using 5 years tering to forecast solar irradiation. The author used multiple
of meteorological data. Then, both bagging and boosting Kalman filters to maintain the original day-ahead horizon
ensembles methods of the base models were constructed [79]. Besides, an ensemble version of Kalman filter is called
and tested with the same data. The results were compared Ensemble Kalman Filter (EnKF) which can correct the fore-
based on two evaluation metrics namely RMSE and R2 . casted value in real-time. It can be done by propagating the
Based on their findings, the proposed model based on bag- uncertainties in time [80]. In another study, the Ensemble
ging and boosting methods improved ANN, DT, and SVR Kalman filter (EnKF) and the state-space models (SSMs) in
in the range of 4.6 and 14.6% in terms of RMSE. Several a short-term PV forecasting experiment which was effective
ensemble models to predict short-term solar irradiation were to forecast solar irradiance [81].
investigated in [71]. The models were RF, Boosted Trees, Jiranantacharoen and Benjapolakul [82], used Kalman
Generalized Random Forest, and Bagged Trees. The perfor- filter and Auto-Regressive Integrated Moving Average
mance of these methods was validated via R2 , and MAPE (ARIMA) for forecasting photovoltaic (PV) power genera-
and compared with SVR and Gaussian process regression. tion. It should be noted that real-time measurement data is
Their result showed a consistent and reliable prediction of needed for Kalman filter to adjust forecast value. Then, they
ensemble methods with high prediction performance in com- proposed a predictor model to apply in the forecasting pro-
parison with the individual regressors. In a recent study in cess once the real-time measurement data is unavailable.
2022, a natural gradient boosting (NGBoost) algorithm was In addition, to estimate the transition matrix for running
used for the short-term solar prediction of PV power systems

13
Journal of Electrical Engineering & Technology (2023) 18:719–733 727

Kalman filter, ARIMA model was applied and the model A method combining the ANN and WD for the forecast-
performance by RMSE and SS. ing of the power output of PV power plants was presented
by [87]. The solar irradiance, and meteorological variables
4.2 Cooperative Forecasting Method such as wind speed humidity, and temperature were chosen
as inputs of the ANN model as the ANN cab address their
The cooperative ensemble forecasting is another type of nonlinear relationships. Then, WD is used to decompose
forecasting method which divides a prediction task into output power of the PV plant resulted in the separation of
several sub-tasks. These tasks can be solved individually. the useful information from disturbances. To build mod-
In addition, the output of one sub-task can be considered els of the decomposed PV output power, ANNs were used.
as input into another sub-task. The overall process of coop- Next, various sky conditions include rainy, sunny, cloudy
erative ensemble forecasting is distributed into two assign- days and overcast were proposed for validation of model.
ments. The first one is to realize the prediction task and Finally, the ANN applied to compare the presented method
the second are some assistant procedures such as parameter with the traditional forecasting method. The results show
optimization, preprocessing, error correction, feature selec- that the method needs lesser calculation time with a better
tion, and postprocessing. forecasting accuracy (MAE of 10.34%, MAPE of 25.37%,
Preprocessing includes reconstruction, decomposition, and RMSE of 19.66% in rainy days).
and transformation which should be performed before fore- Ref [88] was proposed a model called wavelet-coupled
casting [49, 83]. support vector machine (W-SVM) in global incident solar
radiation forecasting based on the minimum and maximum
4.2.1 Pre‑Processing temperature, sunshine hours, evaporation, precipitation
and wind speed as the predictor variables. The merit of the
Raw renewable energy data always has a variety of irregu- W-SVM was benchmarked with the classical SVM model
larities, such as fluctuations. These irregularities have non- to achieve reliable results. Then, in sixteen months from
linear and non-static properties which deteriorate the perfor- 01-March-2014 to 30-June-2015 the data were divided into
mance of the forecasting. Therefore, many pre-processing the test (35%) and train (65%) were set for daily forecasting
techniques have been proposed to break down the renewable in the three metropolitan stations (Townsville Aero, Cairns
energy original signal into several components. These tech- Aero and Brisbane City). The forecast was assessed by pre-
niques have better behavior in terms of outliers and data diction errors (MAE, RMSE, MAPE and RMSE). Based on
variance. With the help of these data processors, the negative the obtained results, the W-SVM model outperformed bet-
impact of irregularities on the accuracy of forecasting can be ter than the classical SVM model for daily forecasts using
properly reduced [84]. In pre-processing forecasting method, optimum input combinations.
the data is decomposed into finite numbers of subseries. It is By referring to Ref [89], the merits of wavelet-ANN mod-
noted that the performance of each subseries is better than els for solar radiation was evaluated. Four different archi-
the original data. Then, to forecast each subseries, a regres- tectures of ANN, namely: adaptive neuro-fuzzy inference
sor and feature extractor is developed independently. In the system (ANFIS), generalized regression neural networks
next step, all the subseries is combined, and the forecast- (GRNN), nonlinear autoregressive recurrent exogenous neu-
ing results will be generated. In pre- processing forecasting, ral network (NARX) and multilayer perceptron (MLP), were
Empirical mode decomposition (EMD) and wavelet decom- used. The for the decomposition of the complex meteoro-
position (WD) are two of the most widely used methods. logical signals into relatively simple parts a wavelet analysis
Time series of temperature data and solar irradiance was used by using wavelet sub-series, and WD transforma-
include seasonal long-term behaviors, and daily informa- tion algorithm. The ANN models were used to model the
tion. Then, for training, it is suggested that to improve the wavelet sub-series and reconstructed to estimate the original
forecasting model performance, the frequency contents of signal. Then, to model the global horizontal irradiation over
those signals can be used instead of the signal values. To do Abu Dhabi city, four meteorological parameters were used
this, the forecasting models can be based on WD of the input including temperature, wind speed, relative humidity, and
data [85]. The WD deal with the solar irradiance fluctuations sunshine duration. The proposed approach was compared to
and used to the input data of a forecasting model and, which ANN models and validated using different metrics such as
resulted in the accuracy improvement. Generally, an efficient R2 , t-statistics RMSE, MAPE, and MBE. The results con-
solution have been made using the wavelet techniques repre- firmed the proposed model improved the performance of
sents resulted in the noise reduction in input datasets before the ANN with a maximum 6.84% in R2 for MLP meanwhile
to implement a prediction model [86]. Some recent studies GRNN had a minimum of 2.78% RMSE.
about WD have summarized (based on the time) as follows. In [90], the information from raw data with better
time–frequency resolutions was extracted and the WD was

13
728 Journal of Electrical Engineering & Technology (2023) 18:719–733

applied with a bias compensation Random Forest (BCRF) train and test the BPNN individually after decomposition
to minimize the prediction error. First, eight decomposi- of the time series data by EMD. The proposed EMD-BPNN
tion layers in a stationary WD on all raw input features was model was evaluated with PV power output time series data.
conducted. Then, there was the lower frequency part of the These data were collected from grid-connected photovol-
original signal due to the higher level of decomposition. taic power plants situated in Ghaziabad India at the 100-
Next, to train a random forest regression model, the wavelet kW roof-top. Finally, for the performance evaluation of the
components and time index as input features were used. In developed model, the data set was divided into the four eval-
fact, an additional model created by BCRF for prediction of uation parameters (the symmetric mean absolute percentage
bias to minimize the overall prediction error. Finally, Wave- error (sMAPE), MSE, MAPE and RMSE) and weekly data
let – BCRF technique was evaluated by some error metrics groups (W1-W4) with different forecast horizon of 1, 12
like RMSE, MAE and MAPE. and 24 h ahead. Results indicate that the data decomposition
In another study [91], a multi-level WD based on day- greatly reduces the complexity and evaluation time of the
ahead solar irradiance forecasting method was proposed. back-propagation neural network.
Initially, based on the weather conditions, the daily solar In another study, a hybrid model with combination of
irradiance series were classified into different patterns. Next, ensemble empirical mode decomposition (EEMD) and
the solar irradiance of the next day 24 h was forecasted for variable weights was proposed to overcome the demerits
each weather pattern using decomposed data series at dif- of the EMD-ANN model. Therefore, for decomposition of
ferent WD levels. Then, to fuse the predictions into the final the original PV power generation data, EEMD was used to
forecasting output, a data-driven fusion model correspond- obtain the residual component (RC) and multiple IMF com-
ing to the weather pattern was applied. Simulations showed ponents (IMF_1, IMF_2). Then, residual components and
that the forecasting accuracy using different WD level data IMF components were divided into low-frequency, interme-
depends on the weather conditions (sunny, cloudy, and rainy diate-frequency, and high-frequency sequences. Then for the
days). In sunny days, solar irradiance reached its maximum prediction of these three sequences, variable-weight com-
level, and the day irradiance curves were relatively smooth bination forecasting (VWCF) method was used. The final
but in cloudy and rainy days the value of solar irradiance was prediction results were calculated by summation of the three
limited to lower level and there were more fluctuations on forecasting results. The total number of input variables were
the irradiance curve. To evaluate the forecasting accuracy, 13 but the first six influential variables were top net solar
two error indexes of RMSE and MAE were used. Gener- rad, surface thermal rad down, temperature, surface pres-
ally, the forecasting accuracies all showed the trend of first sure, time, and relative humidity. RF algorithm was used for
increasing and then decreasing with WD. the determination of each variable impact. Then, MAE and
It is usually difficult to design and implement forecast- MSE were used for the evaluation of the forecasting results.
ing for non-stationary and non-linear signals. In these Based on the results, the prediction accuracy of each model
cases, The Fourier decomposition can be used, but it does (in terms of MAE) was reduced [96].
not give information about the time scale characteristics of In [97], the pre-processing technique (EMD) was pre-
the data. For this purpose, and in order to obtain the time- sented to decompose the data. The solar power output using
scale (time–frequency) information of the signal, we need the hybrid design of the SVR model was used the improved
a method that can extract the intrinsic modes embedded in feature selection algorithm that resulted in the selection of
the signal [92]. The EMD method is proposed for this pur- the best input for the next processing. To improve the accu-
pose. In fact, the EMD is a signal analysis method that can racy of the model, the proposed model design was set based
indicate at what moment, at what frequency, and with what on the SVR with PSO optimization. The results showed that
intensity it is present in a signal [93]. By EMD, there is the the proposed algorithm performed better with an average of
decomposition of complex signal into a linear combination 14.55 (%) of MAPE and 0.95 (%) of nRMSE.
of a limited number of intrinsic mode functions (IMFs) with In [98], several multiscale decompositions in methods
different frequencies. Hence, each of the decomposed IMF of time series analysis for one-hour global solar radiation.
components contains local characteristic signals of different Initially, they calculated the time series of the Clear Sky
time scales from the original signal [94]. Some recent stud- Index. Then, EMD and EEMD methods are used to decom-
ies about EMD have been summarized as follows. pose obtained time series data. Next, the data was forecasted
In [95], a hybrid EMD and back-propagation neural net- using a linear model and nonlinear models using the time
work (BPNN) model was developed for photovoltaic power scale fast fluctuation components. Finally, the results were
forecasting. Then, each IMF and each residue were used to

13
Journal of Electrical Engineering & Technology (2023) 18:719–733 729

improved with a combined hybrid model using globally mul- A solar radiation forecasting time series model was pro-
tiscale decomposition. posed for multi-hour forecasting (915 h ahead) and a small-
In [98], several multiscale decompositions in methods scale solar radiation database (30- and 1-s scales) for 1 day
of time series analysis for one-hour global solar radiation. (47,000 s ahead). In the first step, ARMA was used to predict
Initially, they calculated the time series of the Clear Sky future values of the global solar radiation time series. Then, a
Index. Then, EMD and EEMD methods are used to decom- nonlinear autoregressive (NAR) neural network was applied
pose obtained time series data. Next, the data was forecasted for prediction purposes due to the nonlinearity in solar radia-
using a linear model and nonlinear models using the time tion time series. The results showed that the ARMA- NAR
scale fast fluctuation components. Finally, the results were combination had better accuracy. The NRMSE of the hybrid
improved with a combined hybrid model using globally model was equal to 0.2034 compared to 0.3241 for ARMA
multiscale decomposition. Table 3, which summarizes the model and NRMSE equal to 0.2634 for NAR model [103].
above, is in the appendix. An innovative hybrid model was proposed in three differ-
ent cities in Morocco for forecasting the daily global solar
4.2.2 Post‑Processing radiation. There were three steps for estimation including
1) evaluating the linear aspect of the problem by ARIMA
In general, PV data have time-series measurements [99]. model, and 2) building an ANN model to model the residu-
In post-processing as a cooperative ensemble forecasting als of the ARIMA. It was estimated that the error conditions
method, the forecasting will be performed using the time for the ARIMA model can be forecasted by output from
series consecutively by two or more predictors. Moreover, ANN model. The findings showed that the hybrid model
a time series data may have more than one characteristic (ARIMA–ANN) was more accurate in terms of MAPE, R2,
which is suitable for one specific method. For instance, ANN RMSE, MBE, NRMSE, and TS [104]. Table 2 shows the
is usually applied for modeling non-linear time series while findings of recent studies that use ensemble methods for
ARIMA model is suitable for modeling linear time series solar forecasting.
[38]. The main advantage of post-processing methods is
the improvement in the bias of the global solar irradiance
forecasts without the need for a long-term historical data 5 Conclusion
database [100]. Based on post-processing method, there
are several cooperative ensemble forecasting models such In general, the reliability of solar power systems is affected
as ARIMA-ANN, ARIMA-SVM, ARMA_ANN, ARIMA- by the dynamic nature of solar irradiance. Changes in sun-
GARCH, SARIMA-ANN, and ARMA-GARCH. light intensity led to voltage and power fluctuations in solar
Based on the information, post-processing methods are power plants and disruption of power systems. A good way
often used to optimize the output of NWP models. It should to deal with such problems is to predict solar irradiance.
be noted that detailed local weather characteristics generally Accurate forecasting is challenging and involves a variety
cannot be resolved by NWP predictions hence the spatial of methods statistical, physical and ensemble forecast-
resolution has been grown recently. By using the post-pro- ing methods. This paper has reviewed recently published
cessing method, the forecasting will be improved slightly studies (2015–2021) on solar irradiance forecasting using
by consideration of the uncertainty of some forecasts even ensemble models including competitive and cooperative
though most of the forecasting methods contain statistical forecasting methods. The former has been divided into data
components. Therefore, the final forecasting of data using diversity and parameter diversity and the latter has been cat-
this method will be a single value that can be valid for future egorized according to pre-processing and post-processing.
studies [101]. In this paper, recent articles have been discussed according
In a study by David et al., two models of ARMA and to each category. It can be concluded that ensemble mod-
GARCH were combined to provide probabilistic forecasts els perform better than standalone ones. However, a hybrid
of solar irradiance. Meanwhile, to provide a framework that models have a more complex structure, but they provide
can be applied in an operational context easily, a recursive better accuracy. Based on the previous studies, Artificial
estimation of the parameters of the models has been set up. Neural Network (ANN) and Space Vector Machine (SVM)
As they found, higher forecast accuracy has been found by are widely used with ensemble models (WD-ANN, EMD-
the proposed method (ARMA and GARCH) in comparison BPNN and W-SVM) due to their ability in solving complex
to other machine learning-based techniques. In addition, this and non-linear forecasting models. Also, the performance
model gave additional information about the uncertainty of of the ensemble models has been evaluated by some error
the forecasts which was easier to set up [102]. metrics such as RMSE and MAE. It has been indicated that

13
730 Journal of Electrical Engineering & Technology (2023) 18:719–733

EMD is more powerful, and it has more accuracy than WD correlated with temperature and negatively correlated with
in solar forecasting methods. wind speed. Other weather variables have low correlation
Moreover, in this article, the different model inputs, and values.
their effects on the prediction of solar radiation have been
discussed. Common inputs are solar irradiance, atmospheric
and module temperature, wind speed and direction, and Appendix
humidity. Among these, solar irradiance is most positively
correlated with PV output. Solar irradiance is positively See Table 3.

Table 3  Summary of recent studies for solar forecasting using perspectives

Ref Time ahead Input variables Output variable Perspectives Forecasting Error Comparison
method

[53] Daily −Wind Speed Wind power Data diversity Bagging RMSE = 10.6% Bagging > ANN >
−Wind k-NN
Direction
−Air pressure
[61] Daily −Temperature PV power Data diversity Random MAPE = 8.5% RF > GBDT
−Air Forest
Pressure
−Wind
[63] Daily −Temperature PV power Data diversity Boosting RMSE = 9.48% Boosting > AR
−Relative humid-
ity
−Cloud cover
−Precipitation
[65] Hour −Wind Wind power Data diversity Ada Boosting MAE = 1.0581 AB > ANN
Speed MAPE = 7.63%
[76] Daily −PV power PV power Parameter Time series MSE = 16.24 TS > SARIMA
Diversity
[77] Hour −Irradiation PV power Parameter SVR MAE = 37.04 SVR > NAM >
Diversity SP
[81] Hour −Irradiation Purchased PV Parameter Polynomial MAPE = 10.51% –
−Temperature power Diversity Regression
[87] Daily − Humidity PV power Pre- WD-ANN RMSE = 19.663% WD-ANN > ANN
− Temperature processing MAE = 10.349%
− Wind Speed
[88] Daily − Sunshine hours Global solar irra- Pre- W-SVM RMSE = 2.317 W-SVM > SVM
− Temperature diation processing MJ m−2
− Wind speed— MAE = 1.819
evaporation— MJ m−2
precipitation
[90] 6h − Solar irradiance PV power Pre- WD-BCRF RMSE = 32.12% WD-BCRF > WD-
− Temperature processing MAE = 20.64% SVM > RF
− PV output
[96] 1h − surface thermal PV power Pre- EEMD-VWCF RMSE = 0.0107 EEMD-
rad down processing MAE = 0.0622 VWCF > BPNN > RF
− Top net solar rad
− surface pressure
− Humidity
−Temperature
[95] 1,12 and 24 h − PV power Pre- EMD-BPNN RMSE = 0.019 EMD-BPNN > BPNN
processing MAPE = 5.47%
[38] 10 min − Irradiation PV power Post- ANN RMSE = 6% ANN > ARIMA
processing

13
Journal of Electrical Engineering & Technology (2023) 18:719–733 731

Acknowledgements This work was supported by the Korea Electric 16. Das UK et al (2018) Forecasting of photovoltaic power genera-
Power Corporation (KEPCO) as “Study of Distribution Market Opera- tion and model optimization: a review. Renew Sustain Energy
tor (DMO) Platform and Electricity Market” under Grant R19XO02-02. Rev 81:912–928
17. Singla P, Duhan M, Saroha S (2022) A comprehensive review
Funding Konkuk University Researcher Fund, 2021, (A)Duehee and analysis of solar forecasting techniques. Front Energy
Lee, Korea Electric Power Corporation (KEPCO), R19X002-02, 16(2):187–223
(A)Duehee Lee. 18. Wang F et al (2019) Generative adversarial networks and con-
volutional neural networks based weather classification model
Open Access This article is licensed under a Creative Commons Attri- for day ahead short-term photovoltaic power forecasting. Energy
bution 4.0 International License, which permits use, sharing, adapta- Convers Manage 181:443–462
tion, distribution and reproduction in any medium or format, as long 19. Chen C et al (2011) Online 24-h solar power forecasting based
as you give appropriate credit to the original author(s) and the source, on weather type classification using artificial neural network. Sol
provide a link to the Creative Commons licence, and indicate if changes Energy 85(11):2856–2870
were made. The images or other third party material in this article are 20. Gao M et al (2019) Day-ahead power forecasting in a large-scale
included in the article's Creative Commons licence, unless indicated photovoltaic plant based on weather classification using LSTM.
otherwise in a credit line to the material. If material is not included in Energy 187:115838
the article's Creative Commons licence and your intended use is not 21. Behera MK, Majumder I, Nayak N (2018) Solar photovoltaic
permitted by statutory regulation or exceeds the permitted use, you will power forecasting using optimized modified extreme learning
need to obtain permission directly from the copyright holder. To view a machine technique. Eng Sci Technol, Int J 21(3):428–438
copy of this licence, visit http://​creat​iveco​mmons.​org/​licen​ses/​by/4.​0/. 22. AlSkaif T et al (2020) A systematic analysis of meteorologi-
cal variables for PV output power estimation. Renew Energy
153:12–22
23. Kumar DS et al (2020) Solar irradiance resource and fore-
References casting: a comprehensive review. IET Renew Power Gener
14(10):1641–1656
1. Yagli GM, Yang D, Srinivasan D (2019) Automatic hourly solar 24. Kumari P, Toshniwal D (2021) Deep learning models for solar
forecasting using machine learning models. Renew Sustain irradiance forecasting: a comprehensive review. J Clean Prod
Energy Rev 105:487–498 318:128566
2. Vyas S et al. (2022) Forecasting solar power generation on the 25. Pedro HT et al (2018) Assessment of machine learning tech-
basis of predictive and corrective maintenance activities. arXiv niques for deterministic and probabilistic intra-hour solar fore-
preprint arXiv:​2205.​08109 casts. Renew Energy 123:191–203
3. (IRENA) I.R.E.A. (2020) Renewable Capacity Statistics 2020. 26. Fernández-Peruchena CM et al (2015) Increasing the temporal
Abu Dhabi. resolution of direct normal solar irradiance series in different
4. Zervos A (2015) Renewables 2015 Global Status Report. climatic zones. Sol Energy 115:255–263
5. REN21 (2021) Renewables 2021 Global Status Report. (Paris: 27. Lai CS et al (2021) A deep learning based hybrid method
REN21 Secretariat) for hourly solar radiation forecasting. Expert Syst Appl
6. Notton G et al (2018) Intermittent and stochastic character of 177:114941
renewable energy sources: consequences, cost of intermittence 28. Vignola F et al (2016) Effects of changing spectral radiation dis-
and benefit of forecasting. Renew Sustain Energy Rev 87:96–105 tribution on the performance of photodiode pyranometers. Sol
7. Wang F et al (2020) A minutely solar irradiance forecasting Energy 129:224–235
method based on real-time sky image-irradiance mapping model. 29. Das UK et al (2017) SVR-based model to forecast PV power gen-
Energy Convers Manage 220:113075 eration under different weather conditions. Energies 10(7):876
8. Voyant C et al (2017) Machine learning methods for solar radia- 30. Sun Y et al. (2015) Research on short-term module temperature
tion forecasting: a review. Renew Energy 105:569–582 prediction model based on BP neural network for photovoltaic
9. Van der Meer DW, Widén J, Munkhammar J (2018) Review power forecasting. In: 2015 IEEE power & energy society gen-
on probabilistic forecasting of photovoltaic power produc- eral meeting. 2015
tion and electricity consumption. Renew Sustain Energy Rev 31. Raza MQ, Nadarajah M, Ekanayake C (2016) On recent advances
81:1484–1512 in PV output power forecast. Sol Energy 136:125–144
10. Yang D, van der Meer D (2021) Post-processing in solar forecast- 32. Reddy SS (2017) Optimal scheduling of thermal-wind-solar
ing: ten overarching thinking tools. Renew Sustain Energy Rev power system with storage. Renew Energy 101:1357–1368
140:110735 33. Gupta A, Gupta K, Saroha S (2021) A review and evaluation of
11. Zhang G et al (2022) Solar forecasting with hourly updated solar forecasting technologies. Mater Today: Proc 47:2420–2425
numerical weather prediction. Renew Sustain Energy Rev 34. Kim I (2017) Master Markov chain Monte Carlo and acceptance-
154:111768 rejection algorithms for synthesising short-term variations in the
12. Ahmed R et al (2020) A review and evaluation of the state-of-the- generation output of the photovoltaic system. IET Renew Power
art in PV solar power forecasting: techniques and optimization. Gener 11(6):878–888
Renew Sustain Energy Rev 124:109792 35. Panamtash H, Zhou Q (2018) Coherent probabilistic solar power
13. Mellit A et al (2020) Advanced methods for photovoltaic output forecasting. In 2018 IEEE International Conference on Probabil-
power forecasting: a review. Appl Sci 10(2):487 istic Methods Applied to Power Systems (PMAPS). 2018. IEEE
14. Kumari P, Toshniwal D (2021) Extreme gradient boosting and 36. Rodríguez F et al (2018) Predicting solar energy generation
deep neural network based ensemble learning approach to fore- through artificial neural networks using weather forecasts for
cast hourly solar irradiance. J Clean Prod 279:123285 microgrid control. Renew Energy 126:855–864
15. Hong T et al (2020) Energy forecasting: a review and outlook. 37. Sobri S, Koohi-Kamali S, Rahim NA (2018) Solar photovoltaic
IEEE Open Access J Power Energy 7:376–388 generation forecasting methods: a review. Energy Convers Man-
age 156:459–497

13
732 Journal of Electrical Engineering & Technology (2023) 18:719–733

38. Gürtürk M, Ucar F, Erdem M (2022) A novel approach to investi- 59. Meng M, Song C (2020) Daily photovoltaic power generation
gate the effects of global warming and exchange rate on the solar forecasting model based on random forest algorithm for north
power plants. Energy 239:122344 China in winter. Sustainability 12(6):2247
39. Li B, Zhang J (2020) A review on the integration of probabilistic 60. Ahmad MW, Mourshed M, Rezgui Y (2018) Tree-based ensem-
solar forecasting in power systems. Sol Energy 210:68–86 ble methods for predicting PV power generation and their com-
40. Guermoui M et al (2020) A comprehensive review of hybrid parison with support vector regression. Energy 164:465–474
models for solar radiation forecasting. J Clean Prod 258:120357 61. Persson C et al (2017) Multi-site solar power forecasting using
41. Polikar R (2012) Ensemble learning in ensemble machine learn- gradient boosted regression trees. Sol Energy 150:423–436
ing: methods and applications. Zhang C, Ma Y, Eds. Springer: 62. Taieb SB, Hyndman RJ (2014) A gradient boosting approach
Berlin to the Kaggle load forecasting competition. Int J Forecast
42. Basaran K, Özçift A, Kılınç D (2019) A new approach for predic- 30(2):382–394
tion of solar radiation with using ensemble learning algorithm. 63. Liu H et al (2015) Comparison of four Adaboost algorithm based
Arab J Sci Eng 44(8):7159–7171 artificial neural networks in wind speed predictions. Energy Con-
43. Diagne M et al (2013) Review of solar irradiance forecasting vers Manage 92:67–81
methods and a proposition for small-scale insular grids. Renew 64. Fan J et al (2018) Comparison of Support Vector Machine and
Sustain Energy Rev 27:65–76 Extreme Gradient Boosting for predicting daily global solar
44. Arora I, Gambhir J, and Kaur T (2020) Solar irradiance forecast- radiation using temperature and precipitation in humid subtropi-
ing using decision tree and ensemble models. In: 2020 Second cal climates: a case study in China. Energy Convers Manage
International Conference on Inventive Research in Computing 164:102–111
Applications (ICIRCA). 2020. IEEE 65. Kim H, Lee D (2021) Probabilistic solar power forecasting based
45. Bracale A, Carpinelli G, De Falco P (2016) A probabilistic com- on bivariate conditional solar irradiation distributions. IEEE
petitive ensemble method for short-term photovoltaic power fore- Trans Sustain Energy 12(4):2031–2041
casting. IEEE Trans Sustain Energy 8(2):551–560 66. Park J et al (2020) Multistep-ahead solar radiation forecasting
46. Ren Y, Suganthan P, Srikanth N (2015) Ensemble methods for scheme based on the light gradient boosting machine: a case
wind and solar power forecasting—A state-of-the-art review. study of Jeju Island. Remote Sens 12(14):2271
Renew Sustain Energy Rev 50:82–91 67. Tyralis H, Papacharalampous G (2021) Boosting algorithms
47. He Y, Wang Y (2021) Short-term wind power prediction based on in energy research: a systematic review. Neural Comput Appl
EEMD–LASSO–QRNN model. Appl Soft Comput 105:107288 33(21):14101–14117
48. Gensler A, and Sick B (2016) Forecasting wind power-an 68. Choi S, Hur J (2020) An ensemble learner-based bagging
ensemble technique with gradual coopetitive weighting based model using past output data for photovoltaic forecasting.
on weather situation. In: 2016 International Joint Conference on Energies 13(6):1438
Neural Networks (IJCNN). 2016. IEEE 69. Lee J et al (2020) Reliable solar irradiance prediction using
49. Ribeiro MHDM et al (2022) Efficient bootstrap stacking ensem- ensemble learning-based models: a comparative study. Energy
ble learning model applied to wind power generation forecasting. Convers Manage 208:112582
Int J Electr Power Energy Syst 136:107712 70. Mitrentsis G, Lens H (2022) An interpretable probabilistic
50. Breiman L (1996) Bagging predictors. Mach Learn model for short-term solar power forecasting using natural
24(2):123–140 gradient boosting. Appl Energy 309:118473
51. Vaish J, Datta SS, and Seethalekshmi K (2020) Short term load 71. McCandless T, Dettling S, Haupt SE (2020) Comparison of
forecasting using ANN and ensemble models considering solar implicit vs explicit regime identification in machine learning
irradiance. In: 2020 International Conference on Electrical and methods for solar irradiance prediction. Energies 13(3):689
Electronics Engineering (ICE3). 2020. IEEE 72. Sharma A, Kakkar A (2018) Forecasting daily global solar
52. Ribeiro MHDM, dos Santos Coelho L (2020) Ensemble irradiance generation using machine learning. Renew Sustain
approach based on bagging, boosting and stacking for short- Energy Rev 82:2254–2269
term prediction in agribusiness time series. Appl Soft Comput 73. Moon J et al. (2021) A Comparative analysis of tree-based
2020(86):105837 models for day-ahead solar irradiance forecasting. In 2021
53. Srivastava R, Tiwari A, Giri V (2019) Solar radiation forecasting International Conference on Platform Technology and Service
using MARS, CART, M5, and random forest model: a case study (PlatCon). 2021
for India. Heliyon 5(10):e02692 74. Yang D, Dong Z (2018) Operational photovoltaics power
54. Munawar U, Wang Z (2020) A framework of using machine forecasting using seasonal time series ensemble. Sol Energy
learning approaches for short-term solar power forecasting. J 166:529–541
Electr Eng Technol 15(2):561–569 75. Kaur A et al (2016) Benefits of solar forecasting for energy
55. Ali M et al (2021) Variational mode decomposition based ran- imbalance markets. Renew Energy 86:819–830
dom forest model for solar radiation forecasting: new emerging 76. Flynn H, Larsen G (2022) Investigating the application of
machine learning technology. Energy Rep 7:6700–6717 Kalman Filters for real-time accountancy in fusion fuel cycles.
56. Meenal R et al (2022) Weather forecasting for renewable energy Fusion Eng Des 176:113037
system: a review. Arch Comput Methods Eng. https://​doi.​org/​10.​ 77. Yang D (2019) On post-processing day-ahead NWP forecasts
1007/​s11831-​021-​09695-3 using Kalman filtering. Sol Energy 182:179–181
57. Shaik, AB and Srinivasan S (2019) A brief survey on random 78. Nilam B, Ram ST (2022) Forecasting geomagnetic activity
forest ensembles in classification model. In: International Con- (Dst Index) using the ensemble kalman filter. Monthly Notices
ference on Innovative Computing and Communications. 2019. Royal Astron Soc 511(1):723–731
Springer 79. Takeda H (2017) Short-term ensemble forecast for purchased
58. Huertas Tato J, Centeno Brito M (2018) Using smart persistence photovoltaic generation. Sol Energy 149:176–187
and random forests to predict photovoltaic energy production. 80. Jiranantacharoen P, and Benjapolakul W (2019) Photovol-
Energies 12(1):100 taic power generation forecast by using estimator model and
kalman filter. In 2019 International Conference on Power,
Energy and Innovations (ICPEI). 2019. IEEE

13
Journal of Electrical Engineering & Technology (2023) 18:719–733 733

81. Hajirahimi Z, Khashei M (2022) Hybridization of hybrid struc- 101. Mukaram MZ, Yusof F (2017) Solar radiation forecast using
tures for time series forecasting: a review. Artif Intell Rev. hybrid SARIMA and ANN model: a case study at several loca-
https://​doi.​org/​10.​1007/​s10462-​022-​10199-0 tions in Peninsular Malaysia. Malays J Fundam Appl Sci Spec
82. Bokde N, Feijóo A, Kulat K (2018) Analysis of differencing Issue Some Adv Ind Appl Math 13:346–350
and decomposition preprocessing methods for wind speed pre- 102. Belmahdi B, Louzazni M, Bouardi AE (2020) A hybrid ARIMA–
diction. Appl Soft Comput 71:926–938 ANN method to forecast daily global solar radiation in three
83. De Giorgi MG et al (2015) Error analysis of hybrid photovol- different cities in Morocco. Eur Phys J Plus 135(11):1–23
taic power forecasting models: a case study of mediterranean
climate. Energy Convers Manage 100:117–130 Publisher's Note Springer Nature remains neutral with regard to
84. Malvoni M, De Giorgi MG, Congedo PM (2017) Forecasting of jurisdictional claims in published maps and institutional affiliations.
PV Power Generation using weather input data-preprocessing
techniques. Energy Procedia 126:651–658
85. Zhu H et al (2015) A power prediction method for photovol-
taic power plant based on wavelet decomposition and artificial Negar Rahimi she is currently studying for a Ph.D. degree at Konkuk
neural networks. Energies 9(1):11 University. Her research interests are electricity market and solar power
86. Deo RC, Wen X, Qi F (2016) A wavelet-coupled support forecasting.
vector machine model for forecasting global incident solar
radiation using limited meteorological dataset. Appl Energy Sejun Park he is currently studying for a master’s degree at Konkuk
168:568–593 University’s AI Energy System Laboratory. His research interests are
87. Hussain S, AlAlili A (2017) A hybrid solar radiation modeling electric vehicles, ESS, and prediction.
approach using wavelet multiresolution analysis and artificial
neural networks. Appl Energy 208:540–550 Wonseok Choi he has been in Konkuk University AI system Labora-
88. Chiang P-H et al. (2017) Forecasting of solar photovoltaic sys- tory since 2020. His Majors are system operation optimization analysis
tem power generation using wavelet decomposition and bias- and algorithm production.
compensated random forest. In 2017 Ninth Annual IEEE Green
Technologies Conference (GreenTech) Byoungryul Oh he received the B.S. and M.S. degrees in Electrical
89. Zhen Z et al (2018) Multi-level wavelet decomposition based Engineering from Konkuk University, Seoul, South Korea in 2019
day-ahead solar irradiance forecasting. In 2018 IEEE Power & and 2022 respectively. His research interests include data science and
Energy Society Innovative Smart Grid Technologies Conference distributed optimization utilized in distribution system and electricity
(ISGT). 2018. IEEE market.
90. Prasad R et al (2019) Designing a multi-stage multivariate empir-
ical mode decomposition coupled with ant colony optimization Sookyung Kim she received the B.S. and M.S. degrees in electrical
and random forest model to forecast monthly solar radiation. engineering from Konkuk University, Seoul, South Korea, in 2019 and
Appl Energy 236:778–792 2021 respectively. Her research interests include forecasting and opti-
91. Zhang N et al (2022) Short-term PV output power forecasting mization utilized in power systems, smart grids, and electricity market.
based on CEEMDAN-AE-GRU. J Electr Eng Technol 17:1–12
92. Bokde N et al (2019) A review on hybrid empirical mode decom- Youngho Cho he received the B.S. degree from Konkuk University,
position models for wind speed and wind power prediction. Ener- Seoul, Korea, in 2022. He is currently working toward the Ph.D. degree
gies 12(2):254 from The University of Texas at Austin, TX, USA. His research inter-
93. Yadav HK, Pal Y, Tripathi MM (2020) Short-term PV power ests include machine learning and data analytics for renewable energy,
forecasting using empirical mode decomposition in integra- power system operations, and energy management systems.
tion with back-propagation neural network. J Inf Optim Sci
41(1):25–37 Sunghyun Ahn he is currently studying at Konkuk University’s AI
94. Wang H, Sun J, Wang W (2018) Photovoltaic power forecasting Energy System Laboratory. His research interests are power system
based on EEMD and a variable-weight combination forecasting flexibility and real-time power system evaluation.
model. Sustainability 10(8):2627
95. Zhang W, Dang H, Simoes R (2018) A new solar power output Chulho Chong he is currently studying for B.S. degree in Electri-
prediction based on hybrid forecast engine and decomposition cal and Electronics Engineering from Konkuk university. His research
model. ISA Trans 81:105–120 interests are machine learning, solar power forecasting, and capacity
96. Monjoly S et al (2017) Hourly forecasting of global solar radia- market.
tion based on multiscale decomposition methods: a hybrid
approach. Energy 119:288–298 Daewon Kim he is currently studying for B.S. degree in Computer
97. Acharya SK, Lee J (2018) An Ensemble Learning Method for Engineering from Konkuk university. His research interests are
Solar Power Forecasting. In: Proceedings of Symposium of the machine learning, solar power forecasting, and reinforcement learning.
Korean Institute of communications and Information Sciences,
vol 8, no 3, pp 1310–1311 Cheong Jin he received B.S degree in electrical engineering from
98. Blaga R et al (2019) A current perspective on the accuracy of Pohang University of Science and Technology, Pohang, South Korea, in
incoming solar energy forecasting. Prog Energy Combust Sci 2006. His research interests are digital twin and artificial intelligence.
70:119–144
99. Yang B et al (2021) Classification and summarization of solar Duehee Lee he received the B.S. degree in electronic and electrical
irradiance and power forecasting methods: a thorough review. engineering from the Pohang University of Science and Technology,
CSEE J Power Energy Syst. https://d​ oi.o​ rg/1​ 0.1​ 7775/C
​ SEEJP
​ ES.​ Pohang, South Korea, in 2004, and the M.S. and Ph.D. degrees in
2020.​04930 electrical and computer engineering from The University of Texas at
100. David M et al (2016) Probabilistic forecasting of the solar irra- Austin, Austin, TX, USA, in 2009 and 2015, respectively. He is cur-
diance with recursive ARMA and GARCH models. Sol Energy rently an associate professor with the Electrical Engineering Depart-
133:55–72 ment, Konkuk University, Seoul, South Korea.

13