Forecasting Wind and Solar Energy Production in the Greek Power
System using ANN Models
GEORGIOS FOTIS1,2, NENAD SIJAKOVIC3, MILETA ZARKOVIC3, VLADAN RISTIC3,
ALEKSANDAR TERZIC4, VASILIKI VITA2, MAGDA ZAFEIROPOULOU5,
EMMANOUIL ZOULIAS5, THEODOROS I. MARIS5
1Independent Power Transmission Operator (IPTO),
Dyrrachiou 89 & Kifissou, 104 43 Athens,
GREECE
2Department of Electrical and Electronics Engineering Educators, ASPETE,
School of Pedagogical and Technological Education of Athens,
15122 Marousi,
GREECE
3Faculty of Electrical engineering,
University of Belgrade,
Bul. Kralja Aleksandra 73, 11000 Belgrade,
SERBIA
4EnergoinfoGroup,
N. Ninkovica 3, 11090 Belgrade,
SERBIA
5Core Department,
National and Kapodistrian University of Athens (NKUA),
Euripus Complex, 34400 Psahna Euboea,
GREECE
Abstract: - Renewable energy sources (RES) like solar and wind are quite uncertain because of the
unpredictable nature of wind and sunlight. As a result, there are at present several issues with system
security and the transformed structure of the energy market due to the increasing utilization of
renewable energy sources (wind and solar). Accurate forecasting of renewable energy production is
extremely important to ensure that the produced energy is equal to the consumed energy. Any
deviations have an impact on the system's stability and could potentially cause a blackout in some
situations. The issue of the high penetration of RES is discussed in this study along with a novel
method of predicting them using artificial neural networks (ANN). The SARIMA prediction model is
contrasted with the ANN approach. The suggested ANN for wind power plants has a mean average
prediction error (MAPE) of 3%–4.3%, whereas the SARIMA model has a MAPE of 5%–6.5%. In
comparison, the present prediction approaches typically have a MAPE of 5%–10%. When the MAPE
of solar power plants was calculated, it was also discovered that the SARIMA model had a MAPE of
2.3%–4% and the suggested ANN had a MAPE of 1.4%–2.3%, whereas the MAPE of the present
prediction methods was often about 9%.
Key-Words: - artificial neural networks; distribution system; renewable energy sources forecast; power
production; solar and wind energy production; transmission system.
Received: March 7, 2023. Revised: December 2, 2023. Accepted: December 23, 2023. Published: December 31, 2023.
WSEAS TRANSACTIONS on POWER SYSTEMS
DOI: 10.37394/232016.2023.18.38
Georgios Fotis, Nenad Sijakovic, Mileta Zarkovic,
Vladan Ristic, Aleksandar Terzic, Vasiliki Vita,
Magda Zafeiropoulou, Emmanouil Zoulias, Theodoros I. Maris
E-ISSN: 2224-350X
373
Volume 18, 2023
1 Introduction
The shortage of fossil fuel resources, [1], [2], as
well as global strategic incentives to reduce carbon
emissions in the environment, have led to the high
penetration of renewable energy sources (RES),
where wind and solar energy are crucial in this
process, [3], [4], [5]. For the stable operation state
of the power system there is a need for accurate
forecast of the load, [6], [7], and the RES
production. RES production is varying in time, as
the weather processes on which it is dependent
(solar radiation and wind speed) are also variant and
difficult to be predicted. By aiding balance, accurate
forecasting can help prevent the severe fluctuations
that could be created in the power grid, [8], [9],
[10], [11], and it can even cause a major blackout
with an enormous impact on society and the
economy, [12], [13], [14].
Numerical weather prediction (NWP) tools, [15],
which originated in meteorology and were directed
to the energy sector after a couple of decades, have
been developed because of collaborations between
electrical engineering and meteorology. The aim of
this collaboration was for these NWP models to
cover the needs of the energy sector. These models
are used to forecast wind, along with algorithms that
give a non-linear transfer of wind speed into power
while also considering other relevant meteorological
and orographic influences, as well as wind turbine
type and/or wind farm architecture, including
shadow impacts. Grid operators employ such
forecasts for intraday and near-real time grid
operations, [16], day-ahead market clearance checks
(24h), [17], and operational planning (many days
ahead) depending on the forecast time horizon, [18].
In terms of RES forecasting, deterministic and
probabilistic methodologies can be distinguished,
[19]. In a deterministic technique, the variable to be
forecasted is estimated with a specific value for each
subsequent time step. A probabilistic strategy
emphasizes providing information about the entire
spectrum of likely power generation events, through
a set of alternative scenarios or a collection of
conditional probability density functions (PDFs).
For example, ensemble models, that a model runs
numerous times from radically altered initial
conditions, [20], or statistical techniques, [21], can
provide the basis for probabilistic predictions. This
gives information about the predicted uncertainty
impacting every single value forecast as well as a
prediction about the probability of the occurrence of
a specific event. While renewable energy
forecasting based on deterministic approaches has
been studied for almost three decades, probabilistic
forecasting has only recently attracted attention. It is
currently becoming more common, particularly in
wind energy.
Following the first published work on wind
energy forecast, [22] many research works have
been published on the subject in the following
decades. The most representative ones in wind
power forecasting are in [23], [24], while, [25], [26],
[27], discuss a range of the most current uses for
deterministic and probabilistic wind power
forecasting.
In terms of solar energy forecasting, the first
published paper is found in [28]. [29], [30], give
thorough assessments of the state of solar irradiance
forecasts for energy generation throughout a range
of time periods, whereas, [31], [32], examine and
compare several forecasting strategies to anticipate
solar power output. In [33], along with these
references, it is offered an interesting overview of a
variety of forecasting techniques as well as
statistical and computational intelligence models,
with a focus on forecasting electricity prices. In
terms of comparing forecasting models and
approaches, [34], [35], [36], offer an insightful
investigation of the progress that has been made in
terms of wind power forecasting.
Plenty of researchers have examined the use of
AI algorithms for solar radiation forecasting, which
is a key factor affecting solar systems' output power,
[37]. In [38], researchers found that ANN was the
most useful method for estimating solar radiation
when compared to other methodologies. In [39] it
was discovered that the Gradient Boosting Tree
(GBT) model performs better than other approaches
regarding both precision as well as accuracy when
estimating solar radiation. In [40], the authors
proved that all the machine learning systems they
evaluated could accurately forecast daily solar
radiation data; the ANN method performed the best.
Wind energy development uses Machine Learning
(ML) and Deep Learning (DL) algorithms, just like
solar energy does. Wind speed data and other
pertinent information are used in this process. [41],
presented hybrid SVM models and argued that the
Support Vector Machine (SVM) model was better
than other models. To improve predicting accuracy,
Xiao proposed employing a self-adaptive Kernel
Extreme Learning Machine (KELM), [42].
There is research interest in the short-term
prediction of RES presented in the current work for
the following reasons. Firstly, while there have been
many publications on load forecasting, the same has
not been done on the forecasting of energy
production from RES. Secondly, this gap is
particularly major regarding the Greek electricity
system, which has a particularly high share of
WSEAS TRANSACTIONS on POWER SYSTEMS
DOI: 10.37394/232016.2023.18.38
Georgios Fotis, Nenad Sijakovic, Mileta Zarkovic,
Vladan Ristic, Aleksandar Terzic, Vasiliki Vita,
Magda Zafeiropoulou, Emmanouil Zoulias, Theodoros I. Maris
E-ISSN: 2224-350X
374
Volume 18, 2023
installed capacity in wind and photovoltaic plants.
This becomes especially necessary given that due to
the policy implemented by the European Union and
the Greek governments, the size of RES will
increase in the coming years (approval has already
been given for the first offshore wind farms in the
Aegean Sea), [43], [44], and Greece will become an
exporter of green energy to Europe (with the new
cross-border connections approved to be
implemented), [45].
The main concern of this research work is the
development and future implementation in the
power systems of a prediction method for wind and
solar power production. This prediction method is
based on Artificial Neural Networks (ANNs) and it
is applied to solar and wind power parks in the
southern region of Greece. The contribution of this
methodology has to do with the exact prediction that
it gives, proving its efficiency with other prediction
methods. Its outcome is even more significant
taking into consideration that it will be a valuable
tool for both the Transmission System Operators
(TSOs) and the Distribution System Operators
(DSOs), especially in Greece, where the penetration
of RES will increase and the demand for an accurate
prediction of the power produced by RES (wind and
solar) will be vital for the power system’s stability.
The proposed ANN methodology
implementation has advantages such as scalability,
interpretability, and the ability to capture non-linear
relationships such as power production from solar or
wind. Also, ANNs have the ability to perform many
calculations simultaneously, which allows them to
process large amounts of data quickly and
efficiently, as data from wind or solar power plants.
From this study, it was found that the accuracy of
the ANN model improved the performance of the
power system. Also, the proposed ANN has a fault
tolerance, needed in the prediction of the RES
production. Corruption of one or more cells of an
ANN does not prevent it from generating output,
making ANNs with this feature fault tolerant, since
there is corrupted data from the Supervisory Control
And Data Acquisition (SCADA), from where the
input data are collected. The main reason for
applying ANNs to our research is that ANNs can
store information on the entire network. Therefore,
the disappearance of a few pieces of information in
one place does not prevent the network from
functioning. Also, ANN is selected because of its
ability to learn and use a non-linear relationship to
map several input parameters to an output
parameter.
The structure of this work is as follows. In
Section 2, the proposed methodology using ANNs
that estimates the power production from solar and
wind parks is presented. Section 3 includes the
results of the proposed methodology, and a
comparison between the forecasted and the exact
production is presented. The concluding notes are
provided in the last section.
2 RES Production Prediction
In this section, the ANN method for the prediction
of power production from RES is presented. There
is also a short introduction to the SARIMA
prediction model and, finally, a comparison between
these two different prediction methods, proving the
better performance of the ANNs.
2.1 ANNs and RES Production Prediction
What is already quite well known about ANNs is
that they represent a relatively young technique that
is based on machine learning principles, [46], [47],
[48]. It is a technique designed to determine the
optimal system output given a predetermined set of
high-quality input data that are necessary to
guarantee the ANN algorithm operates as intended.
This algorithm is usually used when it is challenging
to determine how the input and output values relate
to one another. To determine that transfer function,
the ANN algorithm uses a similar principle as the
human nervous system for learning and
implementing the experiences from the past for the
new tasks, [49], [50], [51], [52]. The human nervous
system contains many neurons that process
information and communicate with one another
through synapses. Basically, what happens there is
that each of the neurons that receives some data
processes it and then determines if it will forward it
to other neurons to which it is connected and, if it
will, to which of those neurons. ANN, on the other
hand, represents the mathematical model of this
system, formed out of the artificial neurons that
transfer the information among themselves, and
then, by trying to find the impact that each piece of
information has on the outcome of the problem,
proposes a solution of the analyzed problem based
on the defined set of inputs.
Figure 1 depicts the mathematical representation
of a neuron in an ANN. Synapses connect the inputs
that each neuron receives to that neuron. These
inputs may come from the outside world or from the
neurons in the previous layer. In Figure 1, xj
represents j input, j=1,..,n. The synapses’ strength is
defined through synaptic weight ωiј, where
excitatory synapses are represented by a positive
value of ωiј, and inhibitory synapses by a negative
value. After multiplying the transfer function by the
appropriate weight coefficients on all inputs, the
WSEAS TRANSACTIONS on POWER SYSTEMS
DOI: 10.37394/232016.2023.18.38
Georgios Fotis, Nenad Sijakovic, Mileta Zarkovic,
Vladan Ristic, Aleksandar Terzic, Vasiliki Vita,
Magda Zafeiropoulou, Emmanouil Zoulias, Theodoros I. Maris
E-ISSN: 2224-350X
375
Volume 18, 2023
output is integrated and compared to the threshold
value. The activation value would be 1 if the
transfer function exceeded the threshold, otherwise
it would be 0. Formally, it can be expressed as
follows:
(2)
The sigmoid function, which is described by (3),
is the most frequent activation function of a neuron.
(3)
However, what can be raised as the first
problematic point here is the matter of determining
the weighting factors that would be assigned to each
of the input values to obtain the result of maximal
accuracy. To resolve this issue, the mechanism can
be established in cases in which both the input data
and the measured output values are known. If that is
the case, the neuron can be fed by the input data,
after which the obtained output could be compared
to the already available measured output value.
Based on the difference between the calculated and
measured values, the weighting factors can be
modified to improve the precision of the described
activity of the neuron. This process is iterative and
can be repeated until there is sufficient accuracy
(usually decided by the difference being low
enough). The schematic of is shown in Figure 2
below.
Fig. 1: An ANN's mathematical representation of a neuron
Fig. 2: Weighting factors tuning for the single neuron
WSEAS TRANSACTIONS on POWER SYSTEMS
DOI: 10.37394/232016.2023.18.38
Georgios Fotis, Nenad Sijakovic, Mileta Zarkovic,
Vladan Ristic, Aleksandar Terzic, Vasiliki Vita,
Magda Zafeiropoulou, Emmanouil Zoulias, Theodoros I. Maris
E-ISSN: 2224-350X
376
Volume 18, 2023
The same logic can be applied to the layers, in
which the neurons are grouped in the ANN. Usually,
there are three types of layers, as can be seen in
Figure 3: the input layers, the hidden layers, and the
output layers. The first layer of the ANN is typically
the input layer, with the only purpose of
transmitting the signals further. The output layer is
the final layer in an ANN. Its goal is to establish the
overall ANN's outcome. The number of hidden
layers, which compose the ANN and affect its
accuracy, forms the pathway between the input and
output layers. When the described method of
determining the weighting factors is applied to the
entire ANN instead of the single neuron, it is called
"back propagation", as shown in Figure 3.
This process of deciding the weighting factors in
the entire ANN is called the training of that ANN.
To pull that off properly, the algorithm needs to be
fed quite a large amount of input data, followed by
the accompanying known output values.
Meteorological data is what is most relevant for
wind and solar power plants because the production
power of those sources is directly related to that
kind of data. To improve the prediction further, the
initial database used for training has also been
processed by adding a new entry for the seasons
(summer, winter, autumn, and spring) and for the
time of day (night or day). In this way, the ANN
was enabled to recognize the patterns to predict the
production even better. The dataset used for the
training came from real-life measurements of the
chosen weather and power system parameters for
the current work. Training took up most of that
dataset. Also, one smaller part of the dataset ended
up being used for testing, and another one was used
for the validation of the developed ANN.
Three phases contribute to the ANN's training
and testing process using the MATLAB neural
network toolbox to train and develop neural network
models. The training data is selected from the whole
set of available data. 70% of the database is utilized
for training, 15% is used for testing, and 15% is
used for validation. These database events were
produced and selected at random. This is very
important because a uniform part of the base can
lead to wrong conclusions. Next, to prevent
saturation, the data is normalized. The back
propagation algorithm has been used to train the
artificial neural network.
Fig. 3: Back propagation method of ANN training
WSEAS TRANSACTIONS on POWER SYSTEMS
DOI: 10.37394/232016.2023.18.38
Georgios Fotis, Nenad Sijakovic, Mileta Zarkovic,
Vladan Ristic, Aleksandar Terzic, Vasiliki Vita,
Magda Zafeiropoulou, Emmanouil Zoulias, Theodoros I. Maris
E-ISSN: 2224-350X
377
Volume 18, 2023
In order to evaluate the prediction capability of
the created ANN model, the parameter known as
MAPE (Mean Average Prediction Error), was used,
[53]. For the maximal precision of the created ANN,
this parameter, calculated based on the results of the
ANN and the measured values for the same set of
input data, needs to be as small as possible. The
difference between the measured and predicted
value for the relevant period is what first calculated
as the forecast or residual error (E):
(4)
where Et denotes the variable's period t forecast
error, Yt is the variable's period t measured value,
and Ft denotes the variable's period t forecast
variable. The accuracy metrics are dependent on the
size of the variable because the forecast error Et is
on the same scale as the data. The Mean Absolute
Percentage Error (MAPE) is calculated in equation
(5), to compare forecast performance between
different datasets.
(5)
The best way to achieve minimal error in the
prediction of some power plant production is to
have a good and well-organized database. The better
input data gives better performance for ANN and
better production prediction. Meteorological data
are the most important for wind and solar power
plants because they appear in the physical models of
these renewable energy sources. To improve the
prediction, the database is further processed by
adding an entry for the seasons (summer, winter,
autumn, and spring) and time of day (night or day).
Also, it was tried to present ANN output and
productions binary code, but this was unsuccessful.
The next way to decrease the MAPE is to get the
production from the same hour from the previous
day and to get the production from the previous
hour as inputs.
The database consists of data from 1 January
2018 up to 1 March 2023 (45,242 hourly data) for
solar and wind power plants placed in Southern
Greece, presented in Table 1. The meteorological
data for Wind Power Plants (WPPs) are organized
as: wind direction and wind speed and temperature.
The meteorological data for Solar Power Plants
(SPPs) are organized as: Global horizontal
irradiance, temperature, and wind speed.
In Figure 4 the flowchart of the classification
process is depicted. From validation data, the
optimal ANN hyperparameters are determined:
network architecture, types of activation functions,
regularization parameter, optimal moment for the
end of training, etc. It is not possible to learn
hyperparameters from a test set since this would
result in inconsistent results when estimating
network performance. Cross-validation is the
process of learning hyperparameters by training the
network for several combinations of
hyperparameters on training data and measuring
performance on validation data. The optimal
combination of hyperparameters is determined by
observing which combination yields the greatest
results on the validation data. The measure of
performance for the cross-validation procedure for
regression problems may be the standard deviation.
The maximum number of hyperparameters is
determined by the range of the output variable
(production WPP or SPP) and the number of input
parameters and the size of the database of the
training data set. Then the same number is
distributed among the layers and the number of
neurons in the layers. The optimal result is obtained
based on experiential variation and in accordance
with the minimization of the error (MAPE) on the
test set. It starts with the maximum number of
neurons in a smaller number of layers and comes to
the decision that an ANN with more layers and a
smaller number of neurons is better for predicting
WPP and SPP. Of course, each power plant
represents an ANN model with adjusted
hyperparameters for itself.
WSEAS TRANSACTIONS on POWER SYSTEMS
DOI: 10.37394/232016.2023.18.38
Georgios Fotis, Nenad Sijakovic, Mileta Zarkovic,
Vladan Ristic, Aleksandar Terzic, Vasiliki Vita,
Magda Zafeiropoulou, Emmanouil Zoulias, Theodoros I. Maris
E-ISSN: 2224-350X
378
Volume 18, 2023
Fig. 4: Flowchart of the classification process
Table 1. Solar and Wind Power Plants in selected
regions of Southern Greece
Substation
Name1
Installed
Capacity [MW]
Type of power
plant
SPP-1
2.188
Solar
SPP-2
4.9
Solar
SPP-3
6
Solar
SPP-4
9
Solar
SPP-5
11.963
Solar
WPP-1
7.65
Wind
WPP-2
13.6
Wind
WPP-3
18.4
Wind
WPP-4
28.85
Wind
WPP-5
43.7
Wind
For reasons of information confidentiality, the
names of the power generation substations are given
coded and not with their actual names.
2.2 The SARIMA Prediction Model
The ARIMA model analyzes historical data,
dividing it into three components: autoregressive
(AR), integrated (I), which denotes linear or
polynomial trends, and moving average (MA),
which denotes a weighted moving average over
prior mistakes, [54], [55], [56], [57], [58]. In order
to create the ARIMA(p, d, q) model, it combines the
three model parameters AR(p), I(d), and MA(q).
p = AR order
q = MA order
d = I order (differencing)
The multiplicative Seasonal ARIMA model
namely SARIMA is a variant of the standard
ARIMA model. It is typically written as
SARIMA(p,d,q)(P,D,Q), where, p, d, q and P, D, Q
are positive integers that refer to the polynomial
order of the AR, I, MA parts of the seasonal and
non-seasonal components of the model,
respectively. This is done to account for the wind
speed and the irradiation, which have a seasonal
effect. The SARIMA model is described
mathematically in (6).
(6)
Where: xt is the predicted variable (i.e., wind
speed), is the regular AR polynomial of
order p(), is the regular MA polynomial of
order q(), is the seasonal AR polynomial of
order P(), is the seasonal MA polynomial of
order Q, is the differentiating operator that
eliminate the non-seasonal non-stationarity,
is
the seasonal differentiating operator that eliminate
the seasonal non-stationarity, B is the backshift
operator, making the observation at a specific shift
in time xt (i.e. and finally εt
determines the seasonal period and is subjected to a
white noise technique. These polynomials are
explained in (7-12):
(7)
WSEAS TRANSACTIONS on POWER SYSTEMS
DOI: 10.37394/232016.2023.18.38
Georgios Fotis, Nenad Sijakovic, Mileta Zarkovic,
Vladan Ristic, Aleksandar Terzic, Vasiliki Vita,
Magda Zafeiropoulou, Emmanouil Zoulias, Theodoros I. Maris
E-ISSN: 2224-350X
379
Volume 18, 2023
(8)
(9)
(10)
(11)
(12)
The Akaike Information Criterion (AIC) is a
statistic used to compare models to determine which
one best fits the data. The AIC penalizes some
models for complexity while rewarding those that fit
the data well. It could be written as:
(13)
with n being the overall number of observations
equal to 168, k being the number of free parameters
and RSS is the residual sum of squares. At last, the
forecasting of the intended period can be carried out
utilizing the obtained valid model. Once the model
was formulated, it was used to predict wind speed
and solar radiation for the 1st of July 2023 to the 7th
of July 2023 (the same time period as the proposed
ANN). To evaluate the models' accuracy, the
statistics for the 168-hour forecasting outcomes are
then averaged.
The next section contains the results obtained by
using the developed ANN method in comparison to
the SARIMA prediction model, particularly focused
on the prediction of the power production of wind
and solar power plants in the selected regions of
southern Greece.
3 Forecast Improvement Results
The results given in this section have been obtained
by using both the developed ANN and the SARIMA
model for the forecast of the production of
renewable energy sources, considering the time
horizon of 168 hours into the future. This kind of
generation power forecast has been done for the
period from the 1st of July 2023 to the 7th of July
2023, allowing its further comparison with the
actual measured values of the same parameter.
During this analysis, 10 separate renewable energy
sources have been taken into consideration, as
presented in Table 1.
The obtained findings are crucial for the rest of
the demonstration outcomes, as they provided an
unprecedentedly accurate base for further
investigations of their application in congestion
management, mFRR and aFRR dimensioning and
activations, as well as among the additional
enhanced transmission and distribution system
planning and operation procedures connected to
weather forecasts. For the sake of easier
understanding, the results will first be shown for the
five considered wind power plants and then for the
five solar power plants that were considered.
3.1 Results for the Wind Power Plants
Table 2 gives a comparison of the results obtained
by the ANN forecast and the SARIMA model
compared to the actual production values of the
wind plants. MAPE has been calculated for each of
the 168 hours for each of the WPPs. By that, it was
calculated that the average MAPE for WPP-1 to
WPP-5 was approximately between 3% and 4.3%
for the proposed ANN methodology, while for the
SARIMA model MAPE was approximately between
5% and 6.5%. As a benchmark, the MAPE of WPP
forecasts (market schedules) using GA, [59] or Deep
Learning, [60], is typically around 9% and 7%
respectively, highlighting the improvement made by
the usage of ANN methods. The optimal ANN
structure is with 5 layers with number of neurons:
30 20 10 10 10. More layers lead to over fitting and
increasing the error.
Table 2. Obtained results for the wind power plants
Substation
Name
MAPE [%]
ANN
SARIMA
model
GA
[59]
DL
[60]
WPP-1
3.02
4.92
≈9
≈7
WPP-2
3.27
6.31
WPP-3
3.68
6.52
WPP-4
4.16
5.83
WPP-5
4.28
6.39
To make the examination of the results easier for
the reader, those results have been used to create
diagrams, on which the exact measured production
power is given in blue, whereas the ANN forecast
results, and the SARIMA model are shown in red
and purple, respectively. These diagrams for the
substation names presented in Table 1 can be seen
in Figure 5, Figure 6, Figure 7, Figure 8 and Figure
9. A brief observation of the three curves shown for
WPP also indicates that the results of the ANN
forecast matched the measured production values
well, thus confirming the assumption of ANN being
able to significantly improve the quality of the
WPPs’ power forecast.
WSEAS TRANSACTIONS on POWER SYSTEMS
DOI: 10.37394/232016.2023.18.38
Georgios Fotis, Nenad Sijakovic, Mileta Zarkovic,
Vladan Ristic, Aleksandar Terzic, Vasiliki Vita,
Magda Zafeiropoulou, Emmanouil Zoulias, Theodoros I. Maris
E-ISSN: 2224-350X
380
Volume 18, 2023
Fig. 5: Comparative analysis (7-days period) for WPP-1
Fig. 6: Comparative analysis (7-days period) for WPP-2
0
1
2
3
4
5
6
7
8
020 40 60 80 100 120 140 160
Wind Power Production [MW]
Time [h]
Hourly prediction (7 days period) for the WPP-1
ANN Prediction
Exact production
SARIMA model
0
2
4
6
8
10
020 40 60 80 100 120 140 160
Wind Power Production [MW]
Time [h]
Hourly prediction (7 days period) for the WPP-2
ANN Prediction
Exact production
SARIMA model
WSEAS TRANSACTIONS on POWER SYSTEMS
DOI: 10.37394/232016.2023.18.38
Georgios Fotis, Nenad Sijakovic, Mileta Zarkovic,
Vladan Ristic, Aleksandar Terzic, Vasiliki Vita,
Magda Zafeiropoulou, Emmanouil Zoulias, Theodoros I. Maris
E-ISSN: 2224-350X
381
Volume 18, 2023
Fig. 7: Comparative analysis (7-days period) for WPP-3
Fig. 8: Comparative analysis (7-days period) for WPP-4
0
2
4
6
8
10
020 40 60 80 100 120 140 160
Wind Power Production [MW]
Time [h]
Hourly prediction (7 days period) for the WPP-3
ANN Prediction
Exact production
SARIMA model
0
2
4
6
8
10
12
14
020 40 60 80 100 120 140 160
Wind Power Production [MW]
Time [h]
Hourly prediction (7 days period) for the WPP-4
ANN Prediction
Exact production
SARIMA model
WSEAS TRANSACTIONS on POWER SYSTEMS
DOI: 10.37394/232016.2023.18.38
Georgios Fotis, Nenad Sijakovic, Mileta Zarkovic,
Vladan Ristic, Aleksandar Terzic, Vasiliki Vita,
Magda Zafeiropoulou, Emmanouil Zoulias, Theodoros I. Maris
E-ISSN: 2224-350X
382
Volume 18, 2023
Fig. 9: Comparative analysis (7-days period) for WPP-5
3.2 Results for the Solar Power Plants
Table 3 gives a comparison of the results obtained
by the ANN forecast and the SARIMA model
compared to the actual production values of the
solar plants and other AI techniques. MAPE has
been calculated for the 168 hours for each of the
SPPs, both for the ANN forecast and the SARIMA
method as well. From the aspect of energy
balancing and long-term plans, ANN provided very
good results and was better than the SARIMA
model. ANN can monitor the changes in production
more accurately and therefore generate a more
realistic production plan for solar power plants than
any of the classic planning methodologies. Figure
10, Figure 11, Figure 12, Figure 13 and Figure 14
give the same level of insight but cover the exact
and forecasted production power values for the 5
SPPs. The smallest error MAPE of 1.39% and it is
achieved for a structure with 3 layers with 60, 30
and 20 neurons per layer.
Table 3. Obtained results for the solar power plants.
Substation Name
MAPE [%]
ANN
SARIMA model
GA [59]
DL [60]
SPP-1
2.28
3.20
5-10
5-10
SPP-2
1.66
3.95
SPP-3
1.39
2.26
SPP-4
1.85
2.70
SPP-5
1.92
3.21
0
5
10
15
20
25
30
020 40 60 80 100 120 140 160
Wind Power Production [MW]
Time [h]
Hourly prediction (7 days period) for the WPP-5
ANN Prediction
Exact production
SARIMA model
WSEAS TRANSACTIONS on POWER SYSTEMS
DOI: 10.37394/232016.2023.18.38
Georgios Fotis, Nenad Sijakovic, Mileta Zarkovic,
Vladan Ristic, Aleksandar Terzic, Vasiliki Vita,
Magda Zafeiropoulou, Emmanouil Zoulias, Theodoros I. Maris
E-ISSN: 2224-350X
383
Volume 18, 2023
Fig. 10: Comparative analysis (7-days period) for SPP-1
Fig. 11: Comparative analysis (7-days period) for SPP-2
0
0,5
1
1,5
2
020 40 60 80 100 120 140 160
Solar Power Production [MW]
Time [h]
Hourly prediction (7 days period) for the SPP-1
ANN Prediction
Exact production
SARIMA model
0
0,5
1
1,5
2
2,5
3
3,5
4
4,5
020 40 60 80 100 120 140 160
Solar Power Production [MW]
Time [h]
Hourly prediction (7 days period) for the SPP-2
ANN Prediction
Exact production
SARIMA model
WSEAS TRANSACTIONS on POWER SYSTEMS
DOI: 10.37394/232016.2023.18.38
Georgios Fotis, Nenad Sijakovic, Mileta Zarkovic,
Vladan Ristic, Aleksandar Terzic, Vasiliki Vita,
Magda Zafeiropoulou, Emmanouil Zoulias, Theodoros I. Maris
E-ISSN: 2224-350X
384
Volume 18, 2023
Fig. 12: Comparative analysis (7-days period) for SPP-3
Fig. 13: Comparative analysis (7-days period) for SPP-4
0
1
2
3
4
5
020 40 60 80 100 120 140 160
Solar Power Production [MW]
Time [h]
Hourly prediction (7 days period) for the SPP-3
ANN Prediction
Exact production
SARIMA model
0
1
2
3
4
5
6
7
020 40 60 80 100 120 140 160
Solar Power Production [MW]
Time [h]
Hourly prediction (7 days period) for the SPP-4
ANN Prediction
Exact production
SARIMA model
WSEAS TRANSACTIONS on POWER SYSTEMS
DOI: 10.37394/232016.2023.18.38
Georgios Fotis, Nenad Sijakovic, Mileta Zarkovic,
Vladan Ristic, Aleksandar Terzic, Vasiliki Vita,
Magda Zafeiropoulou, Emmanouil Zoulias, Theodoros I. Maris
E-ISSN: 2224-350X
385
Volume 18, 2023
Fig. 14: Comparative analysis (7-days period) for SPP-5
As can be seen from Figure 10, Figure 11, Figure
12, Figure 13 and Figure 14, the forecasted values
follow the exact measured production power well
even for the SPPs. From the above presented
diagrams, it is confirmed that the ANN forecasting
method can be used efficiently and reliably for both
main types of renewable energy sources. In Table 3
the average MAPE for SPP-1 to SPP-5 for the ANN
forecast and the SARIMA model was approximately
1.4%-2.3% and 2.3% - 4%, respectively. MAPE of
the SPP forecast using GA, [56], or DL, [57], is
typically between 5% and 10%, highlighting the
improvement made by the usage of the ANN
methods.
It can be seen from the findings shown in that
there is not much of a difference between the
ARIMA model and the ANN model's predicting
accuracy level. Based on the very modest forecast
errors of both models, one could claim that they
both performed well in terms of forecasting.
Nonetheless, the ANN model consistently
outperforms the ARIMA model in terms of
forecasting accuracy using the test data.
Nevertheless, ANN is superior. As a result, this
research project also contributes to the clarification
of views expressed in the literature about the
advantages of the ANN model over the ARIMA
model for time series prediction, [55].
From the aspect of energy balancing and long-
term plans, ANN gives very good results. ANN can
better and more accurately monitor changes in
production and therefore generate a more realistic
production plan for SPP and WP than classic
planning methodologies.
4 Conclusions
The main goal of this work is to recommend and
evaluate a new production forecasting technique for
wind and solar power plants. It is trying to deal with
the challenges of balance management that System
Operators (SOs) face in the era of renewable energy
sources. The paper presents a technique for
forecasting wind and solar production that presents
extremely high variability, creating problems for the
distribution and transmission systems, which can
lead the system out of its stable operation. The
technique was based on ANN, and the forecast
provided for wind and photovoltaic production was
extremely accurate, proving that it is better than
other forecasting techniques. Calculating the MAPE
for the wind power plants and for the proposed
ANN, it was found between 3% and 4.3%, when
usually the current prediction methods have a
MAPE of 5%–10%. Doing the same for solar power
plants, it was also found that for the proposed ANN,
the MAPE was between 1.4% and 2.3%, when
usually the current prediction methods have a
MAPE of around 9%. Also, the ANN forecast was
almost the same regardless of the size of the
installed wind or solar generation capacity. This
fact, combined with the cooperation between the
0
2
4
6
8
10
020 40 60 80 100 120 140 160
Solar Power Production [MW]
Time [h]
Hourly prediction (7 days period) for the SPP-5
ANN Prediction
Exact production
SARIMA model
WSEAS TRANSACTIONS on POWER SYSTEMS
DOI: 10.37394/232016.2023.18.38
Georgios Fotis, Nenad Sijakovic, Mileta Zarkovic,
Vladan Ristic, Aleksandar Terzic, Vasiliki Vita,
Magda Zafeiropoulou, Emmanouil Zoulias, Theodoros I. Maris
E-ISSN: 2224-350X
386
Volume 18, 2023
operators, will help to deal with the failures in the
forecasting of production from renewable energy
sources and the reliable and stable operation of the
Greek Power System.
In this work, there was also a comparison
between the proposed ANN and the SARIMA
model. From their comparison, the proposed ANN
is better than the SARIMA model since the MAPE
for the SARIMA is approximately twice as high as
the MAPE of the ANN. This proves that the
proposed ANN is more efficient than the other
prediction model (SARIMA). Also, SARIMA
although is less efficient than the proposed ANN, is
more efficient compared to the current prediction
methods. This is something that needs further
examination and perhaps a future hybrid method
combining ANNs and the SARIMA method
probably will show even better remarks.
The SOs, the MOs, and the flexible resources
must work together effectively. Future research
should combine the implementation of the proposed
methodology with energy storage. Since the RES
production prediction will be accurate enough, the
energy storage will be held in an optimal way,
depending on the power system’s needs. Also, for
the better accuracy of the proposed method, more
data is needed, and at this time, this is limited.
However, real-time data in the future can be
collected using the Internet of Things, which will be
available in the Greek TSO and DSO.
References:
[1] Hou, H.; Lu, W.; Liu, B.; Hassanein, Z.;
Mahmood, H.; Khalid, S. Exploring the Role
of Fossil Fuels and Renewable Energy in
Determining Environmental Sustainability:
Evidence from OECD Countries.
Sustainability 2023, 15, 2048, doi:
10.3390/su15032048.
[2] Rashedi, A.; Khanam, T.; Jonkman, M. On
Reduced Consumption of Fossil Fuels in 2020
and Its Consequences in Global Environment
and Exergy Demand. Energies 2020, 13,
6048, doi: 10.3390/en13226048.
[3] Alves, I.M.; Carvalho, L.M.; Peças Lopes, J.
Modeling demand flexibility impact on the
long-term adequacy of generation systems.
International Journal of Electrical Power &
Energy Systems 2023, 151, 109169.
[4] Alqahtani, S.; Shaher, A.; Garada, A.;
Cipcigan, L. Impact of the High Penetration
of Renewable Energy Sources on the
Frequency Stability of the Saudi Grid.
Electronics 2023, 12, 1470, doi:
10.3390/electronics12061470.
[5] Liu, Y.; You, S.; Tan, J.; Zhang, Y.; Liu, Y.
Frequency Response Assessment and
Enhancement of the U.S. Power Grids Toward
Extra-High Photovoltaic Generation
Penetrations—An Industry Perspective. IEEE
Trans. Power Syst. 2018, 33, 3438–3449.
[6] Pavlatos, C.; Makris, E.; Fotis, G.; Vita, V.;
Mladenov, V. Enhancing Electrical Load
Prediction Using a Bidirectional LSTM
Neural Network. Electronics 2023, 12, 4652,
doi: 10.3390/electronics12224652.
[7] Ekonomou, L. (2010). Greek long-term
energy consumption prediction using artificial
neural networks. Energy, 35(2), pp. 512-517.
doi: 10.1016/j.energy.2009.10.018
[8] Maite Etxandi-Santolaya; Alba Colet-
Subirachs; Mattia Barbero; Cristina Corchero
Development of a platform for the assessment
of demand-side flexibility in a microgrid
laboratory, Applied Energy, Volume 331,
2023, 120359, doi:
10.1016/j.apenergy.2022.120359
[9] M. de Simón-Martín; S. Bracco; M. Rossi; F.
Delfino; A. González-Martínez; J. J. Blanes-
Peiró, A flexible test-bed pilot facility for the
analysis and simulation of Smart Microgrids,
2019 IEEE International Conference on
Environment and Electrical Engineering and
2019 IEEE Industrial and Commercial Power
Systems Europe (EEEIC / I&CPS Europe),
Genova, Italy, 2019, pp. 1-6, doi:
10.1109/EEEIC.2019.8783875.
[10] Zafeiropoulou, M.; Sijakovic, N.; Zarkovic,
M.; Ristic, V.; Terzic, A.; Makrygiorgou, D.;
Zoulias, E.; Vita, V.; Maris, T.I.; Fotis, G. A
Flexibility Platform for Managing Outages
and Ensuring the Power System’s Resilience
during Extreme Weather Conditions.
Processes 2023, 11, 3432, doi:
10.3390/pr11123432.
[11] Zafeiropoulou, M.; Sijakovic, N.; Zarkovic,
M.; Ristic, V.; Terzic, A.; Makrygiorgou, D.;
Zoulias, E.; Vita, V.; Maris, T.I.; Fotis, G.
Development and Implementation of a
Flexibility Platform for Active System
Management at Both Transmission and
Distribution Level in Greece. Appl. Sci. 2023,
13, 11248, doi: 10.3390/app132011248.
[12] Fotis, G.; Vita, V.; Maris, T.I. Risks in the
European Transmission System and a Novel
Restoration Strategy for a Power System after
a Major Blackout. Appl. Sci. 2023, 13, doi:
10.3390/app13010083.
[13] Imai, S.; Novosel, D.; Karlsson, D.;
Apostolov, A. Unexpected Consequences:
Global Blackout Experiences and Preventive
Solutions IEEE Power and Energy Magazine,
WSEAS TRANSACTIONS on POWER SYSTEMS
DOI: 10.37394/232016.2023.18.38
Georgios Fotis, Nenad Sijakovic, Mileta Zarkovic,
Vladan Ristic, Aleksandar Terzic, Vasiliki Vita,
Magda Zafeiropoulou, Emmanouil Zoulias, Theodoros I. Maris
E-ISSN: 2224-350X
387
Volume 18, 2023
2023, 21, pp. 16-29, doi:
10.1109/MPE.2023.3247096.
[14] Kim, D.; Kim, J.-C.; Su, Q.; Joo, S.-K.
Electricity Blackout and Its Ripple Effects:
Examining Liquidity and Information
Asymmetry in U.S. Financial Markets.
Energies 2023, 16, 4939, doi:
10.3390/en16134939.
[15] Donadio, L.; Fang, J.; Porté-Agel, F.
Numerical Weather Prediction and Artificial
Neural Network Coupling for Wind Energy
Forecast. Energies 2021, 14, 338, doi:
10.3390/en14020338.
[16] Heymann, F.; Galus, M. D. Digital platforms
in the energy sector – A menu of regulatory
options for policy makers. In Proceedings of
the 21st Mediterranean Electrotechnical
Conference (MELECON), Palermo, Italy,
2022, pp. 1045-1049, doi:
10.1109/MELECON53508.2022.9843059.
[17] Honkapuro, S.; Jaanto, J.; Annala, S. A
Systematic Review of European Electricity
Market Design Options. Energies 2023, 16,
3704, doi: 10.3390/en16093704.
[18] KU Leuven Energy Institute. EI Fact Sheet:
The Current Electricity Market Design in
Europe. Available online:
https://www.mech.kuleuven.be/en/tme/researc
h/energy_environment/Pdf/ei-factsheet8-
eng.pdf (accessed on 21 July 2023).
[19] Alessandrini, S.; Davò, F.; Sperati, S.; Benini,
M.; Delle Monache, L. Comparison of the
economic impact of different wind power
forecast systems for producers. Adv. Sci. Res.
2014, 11, 49–53.
[20] Tuju, P.E.; Ferrari, F.; Casciaro, G.; Mazzino,
A. The added value of high-resolution
downscaling of the ECMWF-EPS for extreme
precipitation forecasting Atmospheric
Research 2022, 280, 106458, doi:
10.1016/j.atmosres.2022.106458.
[21] Lindberg, O.; Lingfors, D.; Arnqvist, J.; van
der Meer, D.; Munkhammar, J. Day-ahead
probabilistic forecasting at a co-located wind
and solar power park in Sweden: Trading and
forecast verification Advances in Applied
Energy, 2023, 9, 10012, doi:
10.1016/j.adapen.2022.100120.
[22] Brown, B.G.; Katz, R.W.; Murphy, A.M.
Time series models to simulate and forecast
wind speed and wind power. J. Clim. Appl.
Meteorol. 1984, 23, 1184–1195.
[23] Tsai, W.-C.; Hong, C.-M.; Tu, C.-S.; Lin, W.-
M.; Chen, C.-H. A Review of Modern Wind
Power Generation Forecasting Technologies.
Sustainability 2023, 15, 10757, doi:
10.3390/su151410757.
[24] Bazionis, I.K.; Karafotis, P.A.; Georgilakis
P.S. A review of short‐term wind power
probabilistic forecasting and a taxonomy
focused on input data IET Renewable Power
Generation 2022, 16, 77-91, doi:
10.1049/rpg2.12330.
[25] He, B.; Ye, L.; Pei, M.; Lu, P.; Dai, B.; Li, Z.;
Wang, K. A combined model for short-term
wind power forecasting based on the analysis
of numerical weather prediction data Energy
Reports, 2022, 8, pp. 929-939, doi:
10.1016/j.egyr.2021.10.102.
[26] Koochali, A.; Schichtel, P.; Dengel, A.;
Ahmed, S. Random Noise vs. State-of-the-Art
Probabilistic Forecasting Methods: A Case
Study on CRPS-Sum Discrimination Ability.
Appl. Sci., 2022, 12, 5104, doi:
10.3390/app12105104.
[27] Wang, Y.; Xu, H.; Zou, R.; Zhang, L.; Zhang,
F. A deep asymmetric Laplace neural network
for deterministic and probabilistic wind power
forecasting Renewable Energy 2022, 196,
497-517, doi: 10.1016/j.renene.2022.07.009.
[28] Jensenius, J.S.; Cotton, G.F. The
Development and Testing of Automated Solar
Energy Forecasts Based on the Model Output
Statistics (MOS) Technique. In Proceedings
of the 1st Workshopon Terrestrial Solar
Resource Forecasting and on the Use on
Satellites for Terrestrial Solar Resource
Assessment, Washington, DC, USA, 2–5
February 1981.
[29] Samu, R.; Calais, M.; Shafiullah, G.M.;
Moghbel, M.; Shoeb, M. A.; Nouri, B.; Blum,
N. Applications for solar irradiance
nowcasting in the control of microgrids: A
review Renewable and Sustainable Energy
Reviews, 2021, 147, 111187, doi:
10.1016/j.rser.2021.111187.
[30] Choi, S.; Hur, J. An Ensemble Learner-Based
Bagging Model Using Past Output Data for
Photovoltaic Forecasting. Energies 2020, 13,
1438, doi: 10.3390/en13061438.
[31] Wu, Y.-K.; Huang, C.-L.; Phan, Q.-T.; Li, Y.-
Y. Completed Review of Various Solar Power
Forecasting Techniques Considering Different
Viewpoints. Energies 2022, 15, 3320, doi:
10.3390/en15093320.
[32] Ali, M.A.; Elsayed, A.; Elkabani, I.; Akrami,
M.; Youssef, M.E.; Hassan, G.E. Optimizing
Artificial Neural Networks for the Accurate
Prediction of Global Solar Radiation: A
Performance Comparison with Conventional
Methods. Energies 2023, 16, 6165, doi:
10.3390/en16176165.
[33] Weron, R. Electricity price forecasting: A
review of the state-of-the-art with a look into
the future. Int. J. For. 2014, 30, 1030–1081.
WSEAS TRANSACTIONS on POWER SYSTEMS
DOI: 10.37394/232016.2023.18.38
Georgios Fotis, Nenad Sijakovic, Mileta Zarkovic,
Vladan Ristic, Aleksandar Terzic, Vasiliki Vita,
Magda Zafeiropoulou, Emmanouil Zoulias, Theodoros I. Maris
E-ISSN: 2224-350X
388
Volume 18, 2023
[34] Wang, Y.; Zou, R.; Liu, F.; Zhang, L.; Liu, Q.
A review of wind speed and wind power
forecasting with deep neural networks
Applied Energy 2021, 304, 117766, doi:
10.1016/j.apenergy.2021.11776.
[35] Tiwari, S.; Ling, J. -M. A Review of Wind
Energy Forecasting Techniques In Proceeding
of the 2021 International Conference on
Technology and Policy in Energy and Electric
Power (ICT-PEP), Jakarta, Indonesia, 2021,
pp. 213-218, doi: 10.1109/ICT-
PEP53949.2021.9600993.
[36] Zhao, L.; Nazir, M.S.; Nazir, H.M.J.; Abdalla,
A.N. A review on proliferation of artificial
intelligence in wind energy forecasting and
instrumentation management. Environmental
Science and Pollution Research 2022, 29,
43690–43709, doi: 10.1007/s11356-022-
19902-8.
[37] Fan, J.; Wu, L.; Zhang, F.; Cai, H.; Zeng, W.;
Wang, X.; Zou, H. Empirical and machine
learning models for predicting daily global
solar radiation from sunshine duration: A
review and case study in China. Renew.
Sustain. Energy Rev. 2019, 100, 186–212.
[38] Gürel, A.E.; A ˘gbulut, Y.; Biçen, Y.
Assessment of machine learning, time series,
response surface methodology and empirical
models in prediction of global solar radiation.
J. Clean. Prod. 2020, 277, 122353.
Sustainability 2023, 15, 7087 27 of 33
[39] Alizamir, M.; Kim, S.; Kisi, O.; Zounemat-
Kermani, M. A comparative study of several
machine learning based non-linear regression
methods in estimating solar radiation: Case
studies of the USA and Turkey regions.
Energy 2020, 197, 117239.
[40] Khosravi, A.; Koury, R.; Machado, L.; Pabon,
J. Prediction of hourly solar radiation in Abu
Musa Island using machine learning
algorithms. J. Clean. Prod. 2018, 176, 63–75.
[41] Zendehboudi, A.; Baseer, M.; Saidur, R.
Application of support vector machine models
for forecasting solar and wind energy
resources: A review. J. Clean. Prod. 2018,
199, 272–285.
[42] Xiao, L.; Shao, W.; Jin, F.; Wu, Z. A self-
adaptive kernel extreme learning machine for
short-term wind speed forecasting. Appl. Soft
Comput. 2020, 99, 106917.
[43] Delagrammatikas, G.; Roukanas, S. Offshore
Wind Farm in the Southeast Aegean Sea and
Energy Security. Energies 2023, 16, 5208,
doi: 10.3390/en16135208.
[44] Zafeiratou, E.; Spataru, C.; Bleischwitz, R.
Wind offshore energy in the Northern Aegean
Sea islanding region In Proceedings of the
IEEE 16th International Conference on
Environment and Electrical Engineering
(EEEIC), Florence, Italy, 2016, pp. 1-7, doi:
10.1109/EEEIC.2016.7555518.
[45] Ziozas, N.; Tsoutsos, T. Clean Energy
Transition in Southeast Europe: The Paradigm
of Greece from a Fossil Fuel Mediator to a
Community Energy Hub. 2021, In: Coenen,
F.H.J.M., Hoppe, T. (eds) Renewable Energy
Communities and the Low Carbon Energy
Transition in Europe. Palgrave Macmillan,
Cham., doi: 10.1007/978-3-030-84440-0_4.
[46] Mohsin, S.M.; Maqsood, T.; Madani, S.A.
Solar and Wind Energy Forecasting for Green
and Intelligent Migration of Traditional
Energy Sources. Sustainability 2022, 14,
16317, doi: 10.3390/su142316317.
[47] Shukri, N. b. M.; Jumaat, S. A. b. ANN -
Based Model for Prediction Electricity from
Wind Energy. In Proceedings of the 11th
IEEE Symposium on Computer Applications
& Industrial Electronics (ISCAIE), Penang,
Malaysia, 2021, pp. 36-41, doi:
10.1109/ISCAIE51753.2021.9431825.
[48] Siddhant, K.; Garg, H.; Saha, A.; Singh, N.;
Choudhary N. K.; Singh, D. K. Solar Energy
Forecasting using Artificial Neural Network.
In Proceedings of the IEEE Students
Conference on Engineering and Systems
(SCES), Prayagraj, India, 2022, pp. 1-5, doi:
10.1109/SCES55490.2022.9887754.
[49] Pavlatos, C.; Makris, E.; Fotis, G.; Vita, V.;
Mladenov, V. Utilization of Artificial Neural
Networks for Precise Electrical Load
Prediction. Technologies 2023, 11, 70, doi:
10.3390/technologies11030070.
[50] Rojek, I.; Jasiulewicz-Kaczmarek, M.;
Piechowski, M.; Mikołajewski, D. An
Artificial Intelligence Approach for
Improving Maintenance to Supervise Machine
Failures and Support Their Repair. Appl. Sci.
2023, 13, 4971, doi: 10.3390/app13084971.
[51] Tziolis, G.; Spanias, C.; Theodoride, M.;
Theocharides, S.; Lopez-Lorente, J.; Livera,
A.; Makrides, G.; Georghiou, G.E. Short-term
electric net load forecasting for solar-
integrated distribution systems based on
Bayesian neural networks and statistical post-
processing. Energy 2023, 271, 127018.
[52] Zhao, Z.; Tang, J.; Liu, J.; Ge, G.; Xiong, B.;
Li, Y. Short-term microgrid load probability
density forecasting method based on k-means-
deep learning quantile regression. Energy
Reports 2022, 8, 1386.
[53] Armstrong, J.S.; Collopy, F. Error measures
for generalizing about forecasting methods:
Empirical comparisons. Int. J. Forecast. 1992,
1, 69–80.
WSEAS TRANSACTIONS on POWER SYSTEMS
DOI: 10.37394/232016.2023.18.38
Georgios Fotis, Nenad Sijakovic, Mileta Zarkovic,
Vladan Ristic, Aleksandar Terzic, Vasiliki Vita,
Magda Zafeiropoulou, Emmanouil Zoulias, Theodoros I. Maris
E-ISSN: 2224-350X
389
Volume 18, 2023
[54] Islam, M.K.; Hassan, N.M.S.; Rasul, M.G.;
Emami, K.; Chowdhury, A.A. Forecasting of
Solar and Wind Resources for Power
Generation. Energies 2023, 16, 6247, doi:
10.3390/en16176247.
[55] Maroua Haddad, Jean Nicod, Yacouba
Boubacar Mainassara, Landy Rabehasaina,
Zeina Al Masry, Marie Péra Wind and Solar
Forecasting for Renewable Energy System
using SARIMA-based Model. International
conference on Time Series and Forecasting,
Sep 2019, Gran Canaria, Spain. hal-02867736
[56] Yang, S.; Yang, H.; Li, N.; Ding, Z. Short-
Term Prediction of 80–88 km Wind Speed in
Near Space Based on VMD–PSO–LSTM.
Atmosphere 2023, 14, 315, doi:
10.3390/atmos14020315.
[57] Ryu, J.-Y.; Lee, B.; Park, S.; Hwang, S.; Park,
H.; Lee, C.; Kwon, D. Evaluation of Weather
Information for Short-Term Wind Power
Forecasting with Various Types of Models.
Energies 2022, 15, 9403, doi:
10.3390/en15249403.
[58] Luca Pinciroli, Piero Baraldi, Guido Ballabio,
Michele Compare, Enrico Zio, Optimization
of the Operation and Maintenance of
renewable energy systems by Deep
Reinforcement Learning, Renewable Energy,
Volume 183, 2022, pp. 752-763, doi:
10.1016/j.renene.2021.11.052.
[59] Prema, V.; Bhaskar, M. S.; Almakhles, D.;
Gowtham, N.; Rao, K. U. Critical Review of
Data, Models and Performance Metrics for
Wind and Solar Power Forecast. IEEE
Access, 2022, 10, pp. 667-688, doi:
10.1109/ACCESS.2021.3137419.
[60] Lima, M.A.; Carvalho, P.C.M.; Fernández-
Ramírez, L.M.; Braga, P.S. Improving solar
forecasting using Deep Learning and Portfolio
Theory integration, Energy 2020, 195,
117016, doi: 10.1016/j.energy.2020.117016.
List of Abbreviations
aFRR Automatic Frequency Restoration Reserve
AIC Akaike Information Criterion
ANN Artificial Neural Network
ARIMA Auto Regressive Integrated Moving Average
DL Deep Learning
DSO Distribution System Operator
GBT Gradient Boosting Tree
KELM Kernel Extreme Learning
MAPE Mean Average Prediction Error
mFRR Manual Frequency Restoration Reserve
ML Machine Learning
PDF Probability Density Function
RES Renewable Energy Sources
RSS Residual Sum of Squares
SCADA Supervisory Control And Data Acquisition
SO System Operator
SPP Solar Power Plant
SVM Support Vector Machine
TSO Transmission System Operator
WPP Wind Power Plant
WSEAS TRANSACTIONS on POWER SYSTEMS
DOI: 10.37394/232016.2023.18.38
Georgios Fotis, Nenad Sijakovic, Mileta Zarkovic,
Vladan Ristic, Aleksandar Terzic, Vasiliki Vita,
Magda Zafeiropoulou, Emmanouil Zoulias, Theodoros I. Maris
E-ISSN: 2224-350X
390
Volume 18, 2023
Contribution of Individual Authors to the
Creation of a Scientific Article (Ghostwriting
Policy)
The authors equally contributed in the present
research, at all stages from the formulation of the
problem to the final findings and solution.
Sources of Funding for Research Presented in a
Scientific Article or Scientific Article Itself
No funding was received for conducting this study.
Conflict of Interest
The authors have no conflicts of interest to declare.
Creative Commons Attribution License 4.0
(Attribution 4.0 International, CC BY 4.0)
This article is published under the terms of the
Creative Commons Attribution License 4.0
https://creativecommons.org/licenses/by/4.0/deed.en
_US
WSEAS TRANSACTIONS on POWER SYSTEMS
DOI: 10.37394/232016.2023.18.38
Georgios Fotis, Nenad Sijakovic, Mileta Zarkovic,
Vladan Ristic, Aleksandar Terzic, Vasiliki Vita,
Magda Zafeiropoulou, Emmanouil Zoulias, Theodoros I. Maris
E-ISSN: 2224-350X
391
Volume 18, 2023
