energies

Article
Solar Power Forecasting Using CNN-LSTM Hybrid Model
Su-Chang Lim 1 , Jun-Ho Huh 2 , Seok-Hoon Hong 1 , Chul-Young Park 1, * and Jong-Chan Kim 3, *

1 R&D Center, TEF Co., Ltd., 60-12 Suncheon-ro, Seo-Myeon, Suncheon 57906, Korea
2 Department of Data Science, Korea Maritime and Ocean University, Busan 49112, Korea
3 Department of Computer Engineering, Sunchon National University, Suncheon 57992, Korea
* Correspondence: [email protected] (C.-Y.P.); [email protected] (J.-C.K.)

Abstract: Photovoltaic (PV) technology converts solar energy into electrical energy, and the PV
industry is an essential renewable energy industry. However, the amount of power generated
through PV systems is closely related to unpredictable and uncontrollable environmental factors
such as solar radiation, temperature, humidity, cloud cover, and wind speed. Particularly, changes in
temperature and solar radiation can substantially affect power generation, causing a sudden surplus
or reduction in the power output. Nevertheless, accurately predicting the energy produced by PV
power generation systems is crucial. This paper proposes a hybrid model comprising a convolutional
neural network (CNN) and long short-term memory (LSTM) for stable power generation forecasting.
The CNN classifies weather conditions, while the LSTM learns power generation patterns based on
the weather conditions. The proposed model was trained and tested using the PV power output data
from a power plant in Busan, Korea. Quantitative and qualitative evaluations were performed to
verify the performance of the model. The proposed model achieved a mean absolute percentage error
of 4.58 on a sunny day and 7.06 on a cloudy day in the quantitative evaluation. The experimental
results suggest that precise power generation forecasting is possible using the proposed model
according to instantaneous changes in power generation patterns. Moreover, the proposed model
can help optimize PV power plant operations.

Citation: Lim, S.-C.; Huh, J.-H.; Keywords: PV system; PV power generation forecasting; AI; deep learning; CNN; LSTM network
Hong, S.-H.; Park, C.-Y.; Kim, J.-C.
Solar Power Forecasting Using
CNN-LSTM Hybrid Model. Energies
2022, 15, 8233. https://doi.org/ 1. Introduction
10.3390/en15218233
The growing energy demand and development of various energy resources world-
Academic Editor: Alon Kuperman wide because of the increasing global population and industrialization have raised the
Received: 25 September 2022
amount of electrical power generated [1]. Renewable energy is considered an alternative
Accepted: 2 November 2022
to fossil fuels. Importantly, renewable energy, such as solar, wind, hydroelectric, biomass,
Published: 4 November 2022
and geothermal energy, is replenishable and is directly procured from nature without
environmental destruction.
Publisher’s Note: MDPI stays neutral
Photovoltaic (PV) energy generation is a crucial component of renewable energy gener-
with regard to jurisdictional claims in
ation. PV energy is abundant, clean, and environment-friendly; further, it has experienced a
published maps and institutional affil-
gradual increase in use in recent years [2–4]. Moreover, there is a significant acceleration in
iations.
the installation of PV power generation systems worldwide. PV power generation involves
directly converting solar energy into electrical energy. When solar rays are irradiated onto
a semiconductor (n- and p-type silicon), electrons move between electrodes, generating
Copyright: © 2022 by the authors.
electricity. Unlike fossil fuel power plants, the construction of a PV power plant is less
Licensee MDPI, Basel, Switzerland. complicated, emits zero pollution, and is environment-friendly [5].
This article is an open access article However, the power generated from a PV power plant can vary since it is affected
distributed under the terms and by the construction location, time, panel capability, and size [6,7]. Moreover, intermittent
conditions of the Creative Commons variability of a PV power plant occurs depending on weather conditions, such as solar
Attribution (CC BY) license (https:// radiation, temperature, humidity, cloud cover, and wind speed, that are unpredictable and
creativecommons.org/licenses/by/ uncontrollable. Notably, changes in temperature and solar radiation can substantially affect
4.0/). power generation [8,9].

Energies 2022, 15, 8233. https://doi.org/10.3390/en15218233 https://www.mdpi.com/journal/energies

Energies 2022, 15, 8233 2 of 17

The nature of such variables can lead to unstable PV power generation, causing a
sudden surplus or reduction in power output. Furthermore, it may cause an imbalance
between power generation and load demand, inducing control and operation problems
in the power grid [10,11]. If the amount of power generation can be accurately forecasted,
operation optimization strategies, such as peak shaving and reducing the uncertainty of
a power generation system, can be effectively applied [12]. Therefore, a method for accu-
rately predicting the amount of energy produced is crucial for PV-based power generation
systems [13–15]. Accurate forecasting can improve the quality of power provided to a
power grid and help reduce the costs related to general variability [16]. In addition, it can
be used in various operation and control activities, including power scheduling in power
distribution and transmission grids [17].
Research on PV power generation forecasting has recently gained considerable inter-
est. Primarily, studies that employ deep learning models are receiving increasing atten-
tion [18–20]. Machine learning (ML)-based deep learning models were developed to solve
complicated problems by extracting meaningful data from big data. Deep learning models
are different from other theoretical ML models. Various hierarchical structures of deep
learning models enable the automatic learning of the methods required to extract semantic
features from raw data and find useful patterns.
Solar power generation has intermittent characteristics and is highly correlated with
dependence on meteorological parameters. The use of various meteorological parameters
can improve the forecasting accuracy of the model. Most conventional methods use
multivariate regression, which requires collecting multiple relevant data such as solar
radiation, temperature, and power generation. However, most PV plants in Korea have no
environmental sensors installed owing to cost considerations; thus, the factors affecting
PV power generation, such as temperature, humidity, cloud cover, and solar radiation, are
challenging to identify. Therefore, only power generation data can be obtained. The Korea
Meteorological Administration provides temperature and solar radiation data but it is not
suitable for use due to location errors of PV plants. Motivated by this, we aim to present a
deep model that forecasts power generation by analyzing time-series patterns in the data
from a PV power plant at which environmental sensors are not installed.
We propose a hybrid model of a long short-term memory (LSTM) and convolutional
neural network (CNN) model specialized for time-series data analysis and forecasting to
improve the accuracy of power generation forecasting. The proposed network adopts a
parallel structure of branched CNN-LSTM. First, a CNN offers the advantage of higher
accuracy since different patterns can be identified depending on weather conditions based
on the pattern extraction schemes. A proposed CNN model classifies the daily weather
as sunny or cloudy by analyzing historical patterns in raw data. Then, the LSTM part
is split into two models trained separately on sunny and cloudy day data and proceeds
to extract long-term dependent features from the raw data. The LSTM model can more
accurately predict the power generation by individually learning the power generation
pattern according to each weather condition type.
The main purpose of this study is to enhance short-term PV forecasting accuracy
by using the proposed hybrid model based on a univariate data approach. To train and
evaluate the proposed method, using the power generation data collected from a PV
power plant in Busan, Korea. The data were collected from 10 September 2019 to 22 July
2021. The data defined as corresponding to sunny and cloudy weather conditions by the
Korea Meteorological Administration were used. The contributions obtained through the
proposed method are as follows. We develop an accurate and simple CNN-LSTM model
for PV forecasting that relies on historical PV data and does not rely on specific sensors or
environment data.
The remaining paper is organized as follows. Section 2 discusses previous studies on
power generation forecasting algorithms. In Section 3, the proposed hybrid CNN-LSTM
model is introduced. The overview of the proposed analytical method, data cleaning,
re-definition, and design verification of the LSTM model is presented. The performance of
Energies 2022, 15, 8233 3 of 17

the proposed model is evaluated using objective indicators in Section 4, and the conclusions
are presented in Section 5.

2. Related Work
A decrease in the PV module price, timely forecasting of power generation, and
policies related to renewable energy have demonstrated a synergistic effect. Renewable-
energy power plants are actively being constructed, thereby increasing energy production.
Power generation forecasting plays a vital role from an economic perspective. Moreover,
PV power generation forecasting helps rationally plan power generation to effectively solve
problems such as system stability and power generation balance. The power generation
forecasting of renewable energy using physical, statistical, ML, and artificial neural network
(ANN)-based methods have been consistently researched.
Physical models forecast PV power generation based on numerical weather prediction
(NWP) and physical principles of PV cells [21]. An input of a physical model consists
of dynamic information, such as NWP and environmental monitoring data, and static
information, such as the installation angle of PV panels and the conversion efficiency of PV
cells [22]. Though the method does not require past information, it depends on geographic
information of PV panels and detailed weather data.
Statistical methods set mapping relationships between past time-series data and PV
energy outputs [23]. The model performance depends on the relationship between past
power generation observations and weather/climate parameters. Generally used statistical
methods include autoregressive moving average, Kalman filter, and Markov chain [24–26].
An ML method forecasts power generation by analyzing the correlations in nonlinear
data. Such a method involves learning the characteristics of sample data and making
predictions using a neural network or training model such as a support vector machine
(SVM). The cross-validated model in [27] used the characteristic statistical parameters of
illumination intensity and surrounding temperature as the input vector. In [28], a dataset
corresponding to four different environments—sunny, cloudy, foggy, and rainy—was used.
The model used an SVM and weather data to classify a specific environment and select an
appropriate model to predict power generation. A hybrid ML model combining an SVM
and the random forest algorithm was developed in [29]. SVM predicts power generation,
while the random forest was applied with an ensemble learning method wherein predicted
values are combined and analyzed. Past and present generated power and weather data
were combined as the input.
Artificial neural networks (ANNs), a subcategory of ML models, are widely used
for forecasting [30]. An ANN is a nonlinear model consisting of completely connected
nodes. The node connections consist of weights used to analyze specific patterns within the
data. An ANN predicts PV power generation through nonlinear approximation. An ANN-
based prediction algorithm obtained favorable results using a shallow neural network
in the initial steps. For the real-time prediction of PV solar energy, a hybrid method
combining ANN and random forest has been proposed [31]. An improved ANN model
was successfully developed to enhance the PV energy forecasting accuracy [32] despite
uncertainty in solar radiation. Another study predicted power generation using solar
radiation and panel temperature as independent variables [33]. Furthermore, [34] used the
nonlinear characteristics of ANNs to predict solar radiation using multiple variables, such
as average temperature, relative humidity, and daytime hours.
In recent years, numerous studies have been conducted on applying deep learning,
which has shown excellent performance in diverse applications to power generation fore-
casting. Unlike ML algorithms, deep learning learns data and patterns using a complicated
neural network architecture. The deep-learning-based recurrent neural network (RNN)
is an ANN-type network that uses sequential or time-series data [35]. This model uses
information from the previous input to determine the current input and output. A feedback
loop for the recurrent layer saves information about a point in time in the past time-series
data of a memory cell. An RNN contains a shared parameter in each layer. A feed-forward
Energies 2022, 15, x FOR PEER REVIEW 4 of 18
Energies 2022, 15, x FOR PEER REVIEW 4 of 18

Energies 2022, 15, 8233 feedback loop for the recurrent layer saves information about a point in time in the past 4 of 17
feedback loop
time-series data offorathe recurrent
memory cell.layer
An RNN savescontains
information about
a shared a point in
parameter intime
eachin the past
layer. A
feed-forward network has different weights for each node, but an RNN shares the same A
time-series data of a memory cell. An RNN contains a shared parameter in each layer.
feed-forward
weight parameter network
in each has different
network weights
layer. This for eachisnode, but an RNN shares the same
network has different weights for each node, butweight
an RNN adjusted
shares thethrough backpropaga-
same weight parameter
weight
tion
in each parameter
and gradient
network in each
descent
layer. This network
duringweight layer.
training. Thischaracteristics
Such
is adjusted weight
through is adjusted
of RNNsthrough backpropaga-
have been
backpropagation and applied
gradient
tion
todescentandduring
PV powergradient training. Such characteristics of RNNs have been applied to PV applied
descent
generation during
forecasting training.
[36,37]. Such characteristics of RNNs have been power
to PV power
generation generation
forecasting forecasting [36,37].
[36,37].
3. Proposed Method
3.3.Proposed
3.1. Proposed
CNN Method
Method
for Feature Extraction
3.1. CNN
CNNfor
3.1.CNN, for Feature
Feature Extraction
Extraction
a deep learning algorithm frequently used for image, text, and signal inputs,
CNN,
CNN,
consists aadeep
deep
of stacked learning
thatalgorithm
learning
layers algorithm frequently
frequently
extract object used
used
features. for
for1image,
Figure showstext,
image, text, and
and signal
the basic signal inputs,
inputs,
architecture
consists
consists
of a CNN. of
ofstacked
stacked layers
layersthat
thatextract
extractobject
objectfeatures.
features.Figure
Figure 11shows
shows the
the basic
basic architecture
architecture
of
ofaaCNN.
CNN.

Figure 1. Standard convolutional neural network (CNN) architecture.

Figure1.1.Standard
Figure Standardconvolutional
convolutionalneural
neuralnetwork
network(CNN)
(CNN)architecture.
architecture.
The model performance is based on the number of stacked layers and the type and
The model performance is based on the number of stacked layers and the type and
size Thekernel.
model performance is basedlayer
on the number of stacked layers and the type and
size ofthe
of the kernel.TheThedata
dataininthe
theinput
input layergoes
goesthrough
throughconvolution
convolutionand andpooling
poolinglayers
layers
size
toto of the
extract kernel.
deep The data
features. These in features
the inputarelayer goes
then through
input in the convolution
fully connectedand layer,
pooling layers
where
extract deep features. These features are then input in the fully connected layer, where
theto extract deep features. These features are then input in the fully connected layer, where
theresult
resultvalues
valuesareareclassified.
classified.
theAresult values are classified.
ACNNCNNisisused
usedfor forextracting
extractinghierarchical
hierarchicalfeatures
featuresfromfroman animage.
image.Accordingly,
Accordingly,aa
CNN A CNN is used for extracting hierarchical features from an image. Accordingly, a
CNN can extract important information from the one-dimensional sequentialand
can extract important information from the one-dimensional sequential andtwo-
two-
CNN can extractdata.
dimensional important information from the one-dimensional sequential and two-
dimensionalinputinput data.
dimensional
1D-CNNs input
are data.
frequently
1D-CNNs are frequently used in
used innatural
naturallanguage
languageoror timetime series
series processing
processing since
since they
can handle sequential data. Unlike in 2D-CNNs, the kernel executing convolution andsince
they 1D-CNNs
can handle are frequently
sequential data. used in
Unlike natural
in language
2D-CNNs, the or time
kernel series
executing processing
convolutionthe
andthey
thecan
sequence handle
sequence
of data ofsequential
dataapplied
being being data. Unlike
applied
have in 2D-CNNs,
have the
a one-dimensional
a one-dimensional form kernel executing
form
in 1D-CNNs. convolution
in 1D-CNNs.
As shown Asin
and the
shown
Figure sequence
in2,Figure
a 1D-CNN2, a of data
1D-CNN
has being aapplied
has moving
a kernel have aa single
kernelalong
moving one-dimensional
alongdimension. form in 1D-CNNs. As
a single dimension.
shown in Figure 2, a 1D-CNN has a kernel moving along a single dimension.

Figure 2. Computation process of 1D-CNN.

Figure 2. Computation process of 1D-CNN.
Figure
3.2. Long2. Computation process of 1D-CNN.
Short-Term Memory
3.2. Long Short-Term
LSTM Memory
is one of the RNN architectures used for processing time-series data while
3.2. Long Short-Term Memory
solving the problem of gradient exploding or gradient vanishing. This network forgets
unnecessary information while storing information for a prolonged time. Figure 3 shows
the cell structure of LSTM.
 LSTM is one of the RNN architectures used for processing time-series data while
solving the problem of gradient exploding or gradient vanishing. This network forgets
Energies 2022, 15, 8233 5 of 17
unnecessary information while storing information for a prolonged time. Figure 3 shows
the cell structure of LSTM.

Figure 3. Cell diagram of LSTM.

LSTM
LSTM receives
receives current
current input
input data
data and
and long-
long- andand short-term
short-term memory
memory of of the
the previous
previous
cell in each time step. Short-term memory represents the hidden state and represents ℎh𝑡−1
cell in each time step. Short-term memory represents the hidden state and represents t−1,,
while long-term memory represents the
while long-term memory represents the cell state 𝐶t𝑡−1 cell state C −1 . A cell adjusts the information to
. A cell adjusts the information to
be
be maintained or discarded in each time step before delivering short-term
maintained or discarded in each time step before delivering short-term and and long-term
long-term
information
information to tothe
thenext
nextcell
cellusing
using a gate.
a gate.This gate
This is called
gate the the
is called input, forget,
input, or output
forget, gate
or output
and accurately performs filtering through
gate and accurately performs filtering through training. training.
The
The first
first step
step ofof LSTM
LSTM is is identifying
identifying and and removing
removing unnecessary
unnecessary information
information in in aa
memory cell. This process is performed in the forget gate that determines
memory cell. This process is performed in the forget gate that determines an output value an output value
between 0 and 1 based on a sigmoid function. The closer the value is to 1, the more
between 0 and 1 based on a sigmoid function. The closer the value is to 1, the more infor-
information about a previous state is maintained. The information of a previous state is
mation about a previous state is maintained. The information of a previous state is forgot-
forgotten as the value approaches 0, and omitted parts are decided. After passing the
ten as the value approaches 0, and omitted parts are decided. After passing the forget gate,
forget gate, the information to be stored is selected. If the previous time is forgotten, new
the information to be stored is selected. If the previous time is forgotten, new information
information to be remembered is added, and the value of each element is decided as the
to be remembered is added, and the value of each element is decided as the newly added
newly added information. In this case, new information is not stored in the memory cell
information. In this case, new information is not stored in the memory cell uncondition-
unconditionally; instead, an appropriate value is selected using an input gate.
ally; instead, an appropriate value is selected using an input gate.
The sigmoid function is added with the last LSTM cell and the current state and time
The sigmoid function is added with the last LSTM cell and the current state and time
activation feature. The value passed through the sigmoid layer is expressed as a number
activation feature. The value passed through the sigmoid layer is expressed as a number
between 0 and 1 and indicates the degree of new information being updated. The value that
between 0 and 1 and indicates the degree of new information being updated. The value
has passed through the tanh function has a value between −1 and 1. Then, the output is
that has passed
multiplied, through
and the the tanh
final value function
is stored haslong-term
in the a value between
memory. −1The
andfollowing
1. Then, the output
process is
is multiplied, and the final value is stored in the long-term memory.
used to select the output information. The output gate generates a new short-term memory The following process
is used to
(hidden select
state) thedelivered
to be output information.
to the cell in theThenextoutput
step gate
usinggenerates
the currenta new
input,short-term
previous
memory (hidden state) to be delivered to the cell in the next
short-term memory, and newly generated long-term memory. The output of the current step using the current input,
previous short-term memory, and newly generated long-term
time step can also be imported from the hidden state. The short- and long-term values memory. The output of the
current time step can also be imported from the hidden state. The short-
generated by this gate are transferred to the next cell as the process is repeated. The output and long-term
of each time step can be obtained from short-term memory or a hidden state.
 
f t = σ xt w f + ht−1 w f + bias f (1)


gt = tanh xt w g + ht−1 w g + bias g (2)
 values generated by this gate are transferred to the next cell as the process is repeated. The
output of each time step can be obtained from short-term memory or a hidden state.
𝑓𝑡 = 𝜎(𝑥𝑡 𝑤𝑓 + ℎ𝑡−1 𝑤𝑓 + 𝑏𝑖𝑎𝑠𝑓 ) (1)
Energies 2022, 15, 8233 6 of 17
𝑔𝑡 = 𝑡𝑎𝑛ℎ(𝑥𝑡 𝑤𝑔 + ℎ𝑡−1 𝑤𝑔 + 𝑏𝑖𝑎𝑠𝑔 ) (2)

𝑖𝑡 = 𝜎(𝑥𝑡 𝑤𝑖 + ℎ𝑡−1 𝑤𝑖 + 𝑏𝑖𝑎𝑠𝑖 ) (3)

it = σ ( xt wi + ht−1 wi + biasi ) (3)
𝑐 = 𝑓 ⊛ 𝑐 + 𝑔𝑡 ⊛ 𝑖𝑡 (4)
ct 𝑡= f t 𝑡~ ct−𝑡−1
1 + gt ~ i t (4)
𝑜𝑡 = 𝜎(𝑥𝑡 𝑤𝑜 + ℎ𝑡−1 𝑤𝑜 + 𝑏𝑖𝑎𝑠𝑜 ) (5)
ot = σ ( xt wo + ht−1 wo + biaso ) (5)
(6)
h ℎ=
t
𝑡 = 𝑜⊛
o~ tantanh⁡(𝑐
h( c ) 𝑡 )
t (6)
Equations (1)–(6) are applied to each gate. Equation (1) shows the process of the for-
Equations (1)–(6) are applied to each gate. Equation (1) shows the process of the forget
get gate.
gate. Input Input
data data are updated
are updated through
through Equations
Equations (2)–(4).
(2)–(4). Theundergo
The data data undergo an opera-
an operation, as
tion,
in as in Equation
Equation (5), in the(5),
lastinoutput
the lastgate
output
and gate and are as
are updated updated
per theas perEquation
final the final (6).
Equation
Here,
σ(6).
is Here, 𝜎 is the
the sigmoid sigmoidwfunction,
function, 𝑤 is the
is the weight weight
matrix, matrix,
h and ℎ and 𝑐the
ct represent 𝑡 represent
hidden andthe hid-
cell
den and cell states,
states, respectively. respectively.

3.3. Adaptive
3.3. Adaptive Selector
Selector
This paper
This paper proposed
proposed aa hybrid
hybrid CNN-LSTM
CNN-LSTM model model toto forecast
forecastpower
powergeneration.
generation. TheThe
CNN classifies weather type, while the LSTM predicts power generation.
CNN classifies weather type, while the LSTM predicts power generation. The LSTM The LSTM
model learns
model learns the
the power
power generation
generation patterns
patterns according
according to
to weather
weather type,
type, which
which reduces
reduces thethe
complexity and
complexity and variability
variability of data
data fitting
fitting to
to improve
improve prediction
prediction accuracy.
accuracy. Figure
Figure 44 is
is the
the
graph of
graph of power
power generation
generationon on sunny
sunnyandandcloudy
cloudydays.
days.

Figure 4. PV
Figure 4. PV power
power generation
generation graphs;
graphs; (a)
(a) sunny day, (b)
sunny day, (b) cloudy
cloudy day,
day, (c)
(c) scatter
scatter plot
plot of
of sunny
sunny days,
days,
(d) scatter plot of cloudy days.
(d) scatter plot of cloudy days.

The
The power
power generation
generation differs
differs on
on aa sunny
sunny and
and cloudy
cloudy day,
day, as
as shown
shown inin Figure
Figure 4a,b,
4a,b,
respectively.
respectively. The power generation on a cloudy day is low and highly inconsistent. On
The power generation on a cloudy day is low and highly inconsistent. On aa
sunny
sunny day,
day,the
thegraph
graphhas
has aa semi-circular
semi-circular data
data distribution
distribution affected
affected by
by sunrise
sunrise and
and sunset;
sunset;
the
the power generation is the highest around noon, as indicated by the high elevation. On
power generation is the highest around noon, as indicated by the high elevation. On
a cloudy day, the variability of power generation is severe due to the changes in solar
radiation. Figure 4c,d shows that the power generation data on clear and cloudy days are
superimposed and expressed as a scatter plot. In Figure 4c, it can be seen that the power
generation production pattern on a clear day is regularly generated within a certain area. In
this figure, there are cases in which the power generation data is out of range by one point.
This is an outlier due to equipment error or electrical noise. In Figure 4d, it can be seen that
 a cloudy day, the variability of power generation is severe due to the changes in solar
radiation. Figure 4c,d shows that the power generation data on clear and cloudy days are
superimposed and expressed as a scatter plot. In Figure 4c, it can be seen that the power
generation production pattern on a clear day is regularly generated within a certain area.
Energies 2022, 15, 8233 In this figure, there are cases in which the power generation data is out of range by7one of 17
point. This is an outlier due to equipment error or electrical noise. In Figure 4d, it can be
seen that the power generation pattern at the time of cloudy days has high variability
according
the powerto the non-linear
generation characteristics.
pattern at the time of cloudy days has high variability according to
Figure 5a,b show the
the non-linear characteristics. sectional power generation patterns during sunny and cloudy
days,Figure
respectively. On a sunny day,
5a,b show the sectional power the power generation
generation patterns gradually increases
during sunny or de-
and cloudy
creases, but the variability abruptly changes on a cloudy day. Since such
days, respectively. On a sunny day, the power generation gradually increases or decreases, data patterns
also influence
but the continuous
variability abruptly time-series
changes on apatterns, using
cloudy day. thesuch
Since LSTM datamodel for also
patterns training can
influence
complicate
continuousdata convergence.
time-series Therefore,
patterns, theLSTM
using the current weather
model for conditions
training can are first classified
complicate data
according
convergence. Therefore, the current weather conditions are first classified according tothe
to the patterns in Figure 5, and the power generation is forecasted using the
individual
patterns inLSTM
Figuremodel
5, andaccording
the power to generation
the classification results. using
is forecasted Boxesthein the Figure 5 LSTM
individual show
the detailed
model pattern
according form
to the of the collected
classification power
results. Boxesgeneration data.
in the Figure The the
5 show datadetailed
of sunny day
pattern
is generated with a regular period, whereas data of cloudy day has
form of the collected power generation data. The data of sunny day is generated with aa large fluctuation
range
regularand does whereas
period, not show periodicity.
data of cloudyTheday proposed
has a largehybrid CNN-LSTM
fluctuation range and model
doesisnotshown
show
in Figure 6. The proposed hybrid CNN-LSTM model is shown in Figure 6.
periodicity.

Energies 2022, 15, x FOR PEER REVIEW 8 of 17

parameters used for network training are as follows. The optimizer used Adam and set
learning rate = 0.001, the coefficient for primary momentum β1 = 0.9, the coefficient for
secondary momentum β2 = 0.999, and epsilon = 10−8. The weight initialization used kaim-
ing uniform initializer and adopted MSE (Mean Square Error) function as the loss func-
tion. Figure 6c shows the overall structure of the CNN-LSTM hybrid model. Select an
LSTM model according to the output of CNN and predict PV power using time series
5. Sectional patterns of PV power generation; (a) sunny day, (b) cloudy day.
Figure 5.
Figure
data. Sectional patterns of PV power generation; (a) sunny day, (b) cloudy day.

Figure
Figure 6. Proposed CNN-LSTM
6. Proposed CNN-LSTMhybrid
hybridmodel;
model;(a)(a) CNN
CNN architecture
architecture forfor weather
weather classification,
classification, (b)
(b) LSTM
LSTM architecture
architecture forfor
PVPV power
power generation
generation forecasting,
forecasting, (c)(c) Overall
Overall CNN-LSTM
CNN-LSTM hybrid
hybrid model.
model.

4. Experiment
In order to evaluate the effectiveness of the proposed deep model, we designed the
following experiments using collected PV power datasets. The proposed hybrid model
Energies 2022, 15, 8233 8 of 17

Figure 6a shows the CNN architecture used to classify weather conditions. This
architecture consists of two layers of convolution with relu layer and two layers of fully
connected layers. The CNN model is trained to classify sunny and cloudy days based on
the data pattern of weather conditions. The last layer outputs the results of the weather
conditions. In the inner process of the neural network, the difference between the predicted
values and the labels is calculated using a loss function. It then uses a backpropagation
algorithm to minimize the loss function so that the predicted values are as close to the
labels as possible. In the course of training a neural network, parameters, including weights
and biases, are fine-tuned and updated to compare predicted values to labels to produce
better predictions at every epoch. The optimizer used Adam and set learning rate = 0.01,
the coefficient for primary momentum β1 = 0.9, the coefficient for secondary momentum
β2 = 0.999, and epsilon = 10−8 . The weight initialization used kaiming uniform initializer
and adopted the cross-entropy function as the loss function. Figure 6b shows the LSTM
architecture used to forecast PV power generation. This architecture consists of two layers
of LSTM and one layer of the fully connected layer. The last layer outputs the predicted
power generation. The two LSTM models with the same structure are trained on sunny and
cloudy days data, respectively. In the case of the LSTM model, the input, forget and output
gates and cell-state help the model to determine which values to preserve for keeping
long-term memory during the computational process. The hyper-parameters used for
network training are as follows. The optimizer used Adam and set learning rate = 0.001,
the coefficient for primary momentum β1 = 0.9, the coefficient for secondary momentum
β2 = 0.999, and epsilon = 10−8 . The weight initialization used kaiming uniform initializer
and adopted MSE (Mean Square Error) function as the loss function. Figure 6c shows the
overall structure of the CNN-LSTM hybrid model. Select an LSTM model according to the
output of CNN and predict PV power using time series data.

4. Experiment
In order to evaluate the effectiveness of the proposed deep model, we designed the
following experiments using collected PV power datasets. The proposed hybrid model
was trained and tested using the Python 3.8-based Pytorch deep learning framework as a
software platform along with an AMD Ryzen 5 5600X CPU operating at 3.70 GHz, NVIDIA
RTX 2070 GPU, and 16 GB RAM hardware environment.
The proposed hybrid CNN- LSTM model performs PV power generation forecasting,
as shown in Figure 7. In step 1, the power generation data generated from a PV power plant
are classified into sunny and cloudy day data. In step 2, the collected data are pre-processed
to remove noise elements, such as missing values or outliers affecting the experiment.
Normalization is then applied to use the data as a network input. In step 3, the CNN and
LSTM are trained using the training data. Each model only uses power generation data. In
step 4, the results of the forecasting model are tested and verified using various matrices.

4.1. Data Collection and Pre-Processing

This study used the power generation data collected from a PV power plant in Bu-
san, Korea. The average annual PV power generation time of the region is 3.5 h, the
average wind speed is 4–5 m/s, and the meridian altitude is 53◦ in spring (March–May),
76.5◦ in summer (June–August), 53◦ in fall (September–November), and 29.5◦ in winter
(December–January). The data were collected from 10 September 2019 to 22 July 2021. The
data defined as corresponding to sunny and cloudy weather conditions by the Korea Mete-
orological Administration were used. The Korea Meteorological Administration defines
weather conditions in 11 steps. In this study, steps 1–5 were classified as sunny, while steps
6–11 were classified as cloudy [38]. Table 1 presents the schema of the collected data.
 as shown in Figure 7. In step 1, the power generation data generated from a PV power
plant are classified into sunny and cloudy day data. In step 2, the collected data are pre-
processed to remove noise elements, such as missing values or outliers affecting the ex-
periment. Normalization is then applied to use the data as a network input. In step 3, the
Energies 2022, 15, 8233
CNN and LSTM are trained using the training data. Each model only uses power genera- 9 of 17
tion data. In step 4, the results of the forecasting model are tested and verified using var-
ious matrices.

Figure7.7.Flowchart
Figure Flowchartof
ofthe
theproposed
proposedPV
PVpower
powergeneration
generationforecasting
forecastingmethod.
method.

4.1. Data
Table Collection
1. Schema andcollected
of data Pre-Processing
in Busan.
This study used the power generation data collected from a PV power plant in Busan,
Korea. TheColumn Data Type time of the region isDefault
average annual PV power generation 3.5 h, the average
is 4 − 5⁡m/s, and the meridian
wind speed Date altitude is 53° in spring (March–May),
Date Null 76.5°
Time ° Time °
Nullin winter (De-
in summer (June–August), 53 in fall (September–November), and 29.5
Cumulative Power Double (22, 0)
cember–January). The data were collected from 10 September 2019 to 22 “0”July 2021. The
data defined as corresponding to sunny and cloudy weather conditions by the Korea Me-
teorological
Data andAdministration
time in Table 1 were used. The
represented whenKorea
the Meteorological Administration
data were collected; cumulative defines
power
indicates cumulative power generation. The power generation data of 682 days were used
for the experiment. The dataset consists of 446 and 236 sunny and cloudy day data points,
respectively. Data imbalance may occur since the number of sunny day data is 1.88 times
greater than the cloudy day data. Therefore, 100 data from each class were used for network
training, while the remaining data were used for testing.
Abnormal elements in the input data of a forecasting model can cause high forecasting
errors. Pre-processing the input data can improve model accuracy by reducing computation
costs and solving the inappropriate training problem. PV power data sets can contain
outliers due to solar equipment problems, collection system errors, and software system
problems. Therefore, to eliminate outliers, the datasets are cleansing by calculating the
75% quantile.
Quartile3 + 1.5 × IQR (7)
Energies 2022, 15, 8233 10 of 17

The Cumulative power data in Table 1 is preprocessed using the interquartile range
(IQR). If the value of data was higher than that provided by Equation (7), it was determined
as an outlier and removed.
Missing values in the collected data can occur due to problems in the inverter or data
collection device. Since the proposed model predicts the data of a future point in time using
continuous time-series data, missing values at xt can affect the model’s performance. Thus,
time interpolation using the values at times xt−1 and xt+1 were applied.
The numerical scale of the power generation data used in this study is significant. For
example, power generation is less than 10 W near 06:00 when the amount of solar radiation
is low and more than 2500 W between 12:00–14:00 when the amount of solar radiation is
the highest. A significant difference between the minimum and maximum values can affect
the training speed and convergence accuracy of a network. Therefore, using Equation (8),
normalization was performed to improve computation speed and accelerate the network
training convergence speed.
x − xmin
z= (8)
xmax − xmin
Here, x is the actual data, and z indicates a normalized value.
The data used in this study are the power generation data per minute, obtained by
dividing the cumulative power generation data by power generation time. This study
assumed a power plant without environmental sensors and predicted power generation
using time-series analysis characteristics. Clouds and rain that occur locally reduce the
amount of solar radiation, thus affecting the PV power generation efficiency. However,
they are difficult to predict when environmental sensors are unavailable. Therefore, this
study aimed to predict power generation based only on power generation patterns. The
network receives 20 time-series data and outputs one data. The hyper-parameters used for
training the network are as follows. The mean squared error (MSE) was used as the loss
function, and the Adam optimizer was used for optimization. A learning rate of 0.01 was
applied to network optimization.

4.2. Model Evaluation

In this study, a hybrid model combining CNN and LSTM is established. The general
indicators used for evaluating the performance of a time-series data forecasting model
include mean absolute error (MAE), root mean square error (RMSE), and mean absolute
percentage error (MAPE). Let N be the number of test data, x pred be the value predicted by
the proposed algorithm, and x act be the actual value in the quantitative indicators.
MAE measures the error between predicted and actual values and can vary depending
on the measure of continuous variables, as shown in Equation (9). A smaller value indicates
a higher accuracy.
1 n  
N i∑
MAE = − x (9)

x pred act 
=1
RMSE, defined as in Equation (10), measures the difference between predicted and
actual values; it is a measure of the deviation between predicted and actual values. An
RMSE closer to 0 indicates better performance. The difference between RMSE and MAE is
that RMSE is sensitive to outliers, making it more susceptible to significant deviation. Since
the error is squared, a larger error leads to weight being reflected a greater extent.
s
n 2
1 
RMSE =
n ∑ x predicted − xreal (10)
i =1

MAPE, given by Equation (11), indicates the extent of error in the predicted value. A
value closer to 0 can be interpreted as a more outstanding forecasting model performance.
 Energies 2022, 15, 8233 11 of 17

MAPE is robust to outliers, but in this case, it is more challenging to check for errors
intuitively than in the case of MAE.
n  x predicted − xreal 
 
1
MAPE =
n ∑ 
xreal
 × 100%
 (11)
i =1

The power generation forecasted using the LSTM entails errors. The magnitude of
errors varies depending on the density of the predicted power generation data. The error
between predicted and actual values is smaller when the density of predicted values is
dense. An error becomes more significant when the density is sparse. The accuracy of a
forecasting model is also determined by density. As shown in Equation (12), R2 measures
the strength of a correlation between the predicted and actual values. R2 has the range
of 0 ≤ R2 ≤ 1. A value closer to 0 indicates very low accuracy, and a value closer to 1
indicates higher accuracy of the forecasting model.
 2
∑in=1 x predicted − xreal
R2 = 1 − 2
(12)
∑in=1 ( xreal − xreal )

where xreal is the average of the actual power generation data.

4.3. Experimental Results

Quantitative and qualitative evaluations were performed using sunny and cloudy
ergies 2022, 15, x FOR PEER REVIEW days’ data for model verification. The data of the same month (sunny day: 28 December
2020, cloudy day: 27 December 2020) were used for uniformity of the validation process.
Figure 8 is the graph of the sunny day dataset; certain sections demonstrate power gen-
eration fluctuations due to clouds. Figure 9 is the graph of the cloudy day dataset in which
measured
the maximumdatapower
from output
the device, while
is 400 W. the red
In contrast, on line shows
a sunny themaximum
day, the predictionpowerdata. A q
output is 2500 W. This result implies that the amount of solar radiation
tive evaluation was performed using MAPE, RMSE, and MSE to validate the LST is minimal, resulting
in severe fluctuations in overall power generation. The blue line shows the measured data
casting model. The LSTM model adequately followed the trend of the observati
from the device, while the red line shows the prediction data. A quantitative evaluation
forwas
accurate forecasting.
performed using MAPE, RMSE, and MSE to validate the LSTM forecasting model. The
LSTM model adequately followed the trend of the observation data for accurate forecasting.

Figure 8. Forecasting result of sunny day power generation data.

Figure 8. Forecasting result of sunny day power generation data.
 Energies 2022, 15, 8233 Figure 8. Forecasting result of sunny day power generation data. 12 of 17

Figure 9. Forecasting result of cloudy day power generation data.

Figure 9. Forecasting result of cloudy day power generation data.
Table 2 presents the results of a quantitative evaluation of the test data for sunny and
cloudy days. The sunny day data showed a MAPE of 4.58, which is lower than that of the
Tableday
cloudy 2 presents
data, 7.06. the results
In the cloudy of
daya data,
quantitative
peak point evaluation
errors occurredofinthe
thetest data for sun
fluctuation
cloudy days. The sunny day data showed a MAPE of 4.58, which is lower
amplitude according to the power production period, which increased the MAPE than tha
value
compared to that of the sunny day data. The RMSE and MAE were lower
cloudy day data, 7.06. In the cloudy day data, peak point errors occurred in the fluc in the cloudy
day data, thus exhibiting more outstanding results. Figures 10 and 11 plot the residuals
amplitude according
between observed and to the power
forecasted valuesproduction period,
for the sunny and cloudywhich
day test increased the MAPE
data. The dotted
compared to that
line indicates of the sunny
the standard day
deviation (SD)data. The RMSE
corresponding and
to the MAE were lower in the
residuals.
day data, thus exhibiting more outstanding results. Figures 10 and 11 plot the re
Table 2. Quantitative validation results of the power generation forecasting model.
between observed and forecasted values for the sunny and cloudy day test data. T
2
ted line Column
Energies 2022, 15, x FOR PEER REVIEW indicates the standardMAPE deviation RMSE(SD) corresponding
MAE to the Rresiduals.
Sunny 4.58 43.87 34.00 0.99
Table 2. Cloudy
Quantitative 7.21
validation results of 9.09
the power 6.97
generation 0.99
forecasting model.

Column MAPE RMSE MAE 𝑹𝟐

Sunny 4.58 43.87 34.00 0.9
Cloudy 7.21 9.09 6.97 0.9

Figure 10.Residual
Figure 10. Residual graph
graph of forecasting
of forecasting results results using
using power power generation
generation data of sunnydata
days.of sunny days
Energies 2022, 15, 8233 13 of 17
Figure 10. Residual graph of forecasting results using power generation data of sunny days.

Figure 11.Residual
Figure 11. Residual graph
graph of forecasting
of forecasting resultsresults usinggeneration
using power power generation datadays.
data of cloudy of cloudy days.

Table 3 presents the residual comparison results for a sunny day. The SD of residual
Table 3 presents the residual comparison results for a sunny day. The SD of resid
is 38.25, and the mean is 34.00 on a sunny day. Table 4 presents the residual comparison
is 38.25,
results of and the day.
a cloudy mean Theis SD
34.00 on a sunny
of residual day.
is 27.42, andTable 4 presents
the mean thea cloudy
is 17.13 on residualday.compari
results of a large
A relatively cloudy day. of
number ThetheSD of residual
residual is 27.42,
values of and the
sunny days mean
deviate is the
from 17.13 on a cloudy d
standard
A relatively large number of the residual values of sunny days deviate from400
deviation section for the number of data, which can be observed between index 200 and the stand
in Table 3. This
deviation section
section foristhe
thenumber
peak point ofregion
data, where
whichthe power
can generationbetween
be observed is the largest
index 200
in Figure
400 8.
in Table 3. This section is the peak point region where the power generation is
largest in Figure 8.
Table 3. Comparison between observed and hybrid-model estimated values of DC power: Sunny day.

Table 3.Index
Comparison between observed
DC Power and hybrid-model
(Observed) estimated valuesDifference
Model Estimated of DC power: Sunny d
1 9 9 0
Index100 DC Power (Observed)
864 Model Estimated 10
854 Difference
1 200 92272 2216 9 56 0
300 2464 2377 87
100 400 864
1984 1960 854 24 10
200 2272 2216 56
300 2464 2377 87
Table 4. Comparison between observed and hybrid-model estimated values of DC power: Cloudy day.
400 1984 1960 24
Index DC Power (Observed) Model Estimated Difference
Table 4. Comparison
1 between observed
10 and hybrid-model9 estimated values of DC
1 power: Cloudy
100 192 197 −5
Index200 DC Power (Observed)
213 Model
214 Estimated −1 Difference
300 373 360 13
1 400 10266 258 9 8 1
100 192 197 −5
The forecasting model can generate errors in predicting the maximum power gen-
eration value. The error in the cloudy day data is 10.83 less than that for the sunny day
data. The number of data deviating from the standard deviation section is less than the
number of data used in the experiment. Figures 12 and 13 show the frequency distribution
of residuals with an SD of 38.25 and a mean of 34 for sunny days and an SD of 27.42 and a
mean of 17.13 for cloudy days, respectively.
 The
of number
data used of
in data deviating from
the experiment. the standard
Figures 12 and 13deviation
show thesection is lessdistrib
frequency than
of data with
siduals used an
in the
SD experiment.
of 38.25 andFigures
a mean12of and 13 sunny
34 for show the
daysfrequency
and an SD distri
of
siduals
mean of with
17.13an
forSD of 38.25
cloudy and
days, a mean of 34 for sunny days and an SD of
respectively.
Energies 2022, 15, 8233 mean of 17.13 for cloudy days, respectively. 14 of 17

Figure 12. Histogram of residuals between predicted and actual values on a sunny da
12.Histogram
Figure 12.
Figure Histogramof residuals betweenbetween
of residuals predicted and actual values
predicted andonactual
a sunnyvalues
day. on a sunny da

13. Histogram
Figure 13.
Figure Histogram of residuals betweenbetween
of residuals predicted and actual values
predicted andonactual
a cloudy day. on a cloudy d
values
FigureThe13.advantage
Histogram of residuals
of the LSTM model between predicted
is further andwhen
highlighted actual values
using on a cloudy d
unfavorable
Thepatterns.
weather advantage Figureof theshow
14a,b LSTM model is
the qualitative furtherresults
evaluation highlighted when
of various test data using
for The
the advantage
sunny and cloudy of the
day LSTM
data, model
respectively. is
The further
proposed
weather patterns. Figure 14a,b show the qualitative evaluation results of highlighted
algorithm when
adequately using
vario
follows the trends in various power generation patterns for accurate forecasting. However,
weather
for theoccurpatterns.
sunny Figure
andpoints
cloudy 14a,b
day showrespectively.
data, the qualitative Theevaluation
proposed results
algorithmof vari
ade
errors at peak for the cloudy data with irregular patterns and large fluctuations.
for
lows the
thesunny
However, trends andin cloudy
the algorithm is robustday
various powerdata,
to abrupt respectively.
generation
changes Theand
in the patterns
patterns proposed
for accurate
fluctuation algorithm
forecastin
periods, ad
lows
thus the
errors occurtrends
reasonably in various
at following
peak points power
the trend
for the generation
for reliable
cloudy patterns
forecasting.
data with for accurate
A irregular
qualitative forecastin
evaluation
patterns and la
confirms the high efficiency and improved performance of the proposed model in solving
errors
tions. occur
However, at peak points
the algorithm for the cloudy data with irregular
is robust to abrupt changes in the patterns patterns and la
and
the PV power generation forecasting problem.
tions.The
periods, However,
thus the algorithm
powerreasonably
generation is
following
forecasting robust toofabrupt
the trend
performance thefor changes
reliable
proposed model in thedepend-
forecasting.
differs patterns and
A qualit
periods,
ation thethus
ing onconfirms reasonably
weather patterns
the following
highandefficiency the improved
validation datasets.
and trend for reliable
The characteristics forecasting.
of power
performance of the A
generation quali
propos
fluctuation patterns are sufficiently captured by the CNN, while the LSTM finds long-term
ation confirms
solving the PV the high
power efficiencyforecasting
generation and improved performance of the propos
dependencies in the time-series input. In other words, theproblem.
CNN-LSTM hybrid model may
solving the the
not produce PVsamepower resultgeneration forecasting
in an environment problem.
having different weather conditions. For
example, the proposed model produced errors in the power generation peak points for the
sunny day data but made accurate predictions by following the fluctuations pattern for the
cloudy day data having abrupt changes.
Energies 2022,15,
Energies2022, 15,8233
x FOR PEER REVIEW 15 of
15 of 17
18

Figure 14.Qualitative
Figure14. Qualitativeevaluation
evaluationof
oftrends
trendsininvalidation
validationdata;
data;(a)
(a)sunny
sunnyday,
day,(b)
(b)cloudy
cloudyday.
day.
5. Conclusions
The power generation forecasting performance of the proposed model differs de-
A PVon
pending power
the generation forecasting
weather patterns andmodel can improve
validation datasets.prediction accuracy according
The characteristics of powerto
weather conditions and enhance the planning, operation, and stability of
generation fluctuation patterns are sufficiently captured by the CNN, while the LSTMPV power systems.
However, PV power generation forecasting can be challenging owing to intermittency in
finds long-term dependencies in the time-series input. In other words, the CNN-LSTM
weather conditions. A statistical method for inferring dependency between past and short-
hybrid model may not produce the same result in an environment having different
term observed data is required to build an effective forecasting model that depends only on
weather conditions. For example, the proposed model produced errors in the power gen-
past data, excluding the solar radiation data highly correlated with power production.
eration peak points for the sunny day data but made accurate predictions by following
This paper proposes a CNN-LSTM hybrid model for PV power generation forecasting.
the fluctuations pattern for the cloudy day data having abrupt changes.
The proposed model overcomes the drawbacks of the individual models while preserving
their advantages. Because training one LSTM model using different time-series data may
5. Conclusions
affect network convergence, separate models were built according to weather conditions.
WeatherA PV power were
patterns generation forecasting
classified using CNN,modelandcanthe
improve
LSTMprediction
model wasaccuracy
appliedaccord-
to the
ing to weather conditions and enhance the planning, operation, and stability
classification results. The power generation patterns can clearly distinguish between of PV sunny
power
systems.
and cloudy However,
days. ThePV power generation
proposed forecastingforecasting
model can can be challenging
sufficiently owing
reflect the to inter-
fluctuations
mittency
in in weatherfor
power generation conditions. A statisticalAmethod
accurate forecasting. for inferring
qualitative evaluation dependency
confirmed between
that the
past and short-term
forecasted observed
power output data
signals is required
react to build and
to fluctuations an effective
adequatelyforecasting
follow model that
the actual
depends only on past
power output signal trend. data, excluding the solar radiation data highly correlated with
power production.
Moreover, a quantitative evaluation demonstrated that the proposed model has RMSEs
of 4.58This
andpaper
27.55,proposes
MAEs ofa 34.00
CNN-LSTM
and 17.13,hybrid
andmodel
MAPEs forofPV power
4.58 and generation
7.06 for sunnyforecast-
and
ing. The
cloudy dayproposed model overcomes
data, respectively. the drawbacks
The residual of had
distributions the individual
an SD of 38.25models
and awhile
meanpre-of
serving
34.00 for their
sunnyadvantages.
days and anBecause training
SD of 27.42 and aone LSTM
mean model
of 17.13 forusing
cloudy different time-series
days, respectively,
data may affect network convergence, separate models were built according to weather
Energies 2022, 15, 8233 16 of 17

thus establishing the validity of the proposed model. The characteristics of collected data
and power generation values may vary depending on the inverter manufacturer. Since the
power generation capacity of a PV power plant system varies depending on installation
scale and weather conditions, the model requires reconfiguration according to the power
generation capacity. Future research should focus on methods for forecasting power
generation by adaptively reflecting the power generation capacity of a PV power plant.
This includes the application of optimization techniques that automatically perform model
fitting according to the data accumulated in the system to improve the prediction accuracy;
In addition, it is necessary to analyze factors affecting power generation efficiencies such
as PV panel characteristics and solar radiation. Additionally, the forecasting results are
expected to be used to understand the decrease in inverter efficiency over time. In cases
where inverter power generation is less than the predicted generation, it will help determine
whether the phenomenon is an efficiency decline due to the aging of the inverter.

Author Contributions: Conceptualization, S.-C.L., C.-Y.P. and J.-C.K.; Data curation, S.-C.L., J.-H.H.,
S.-H.H. and J.-C.K.; Formal analysis, S.-C.L. and J.-C.K.; Funding acquisition, C.-Y.P. and J.-C.K.;
Investigation, S.-C.L., J.-H.H., S.-H.H., C.-Y.P. and J.-C.K.; Methodology, S.-C.L., J.-H.H., S.-H.H.,
C.-Y.P. and J.-C.K.; Project administration, S.-H.H., C.-Y.P. and J.-C.K.; Resources, S.-C.L., J.-H.H.,
S.-H.H., C.-Y.P. and J.-C.K.; Software, S.-C.L., J.-H.H., S.-H.H., C.-Y.P. and J.-C.K.; Supervision, S.-H.H.,
C.-Y.P. and J.-C.K.; Validation, S.-C.L., S.-H.H., C.-Y.P. and J.-C.K.; Visualization, S.-C.L., J.-H.H.
and J.-C.K.; Writing—original draft, S.-C.L., J.-H.H., S.-H.H., C.-Y.P. and J.-C.K.; Writing—review &
editing, S.-C.L., J.-H.H., C.-Y.P. and J.-C.K. All authors have read and agreed to the published version
of the manuscript.
Funding: This work was supported by the Korea Institute of Energy Technology Evaluation and
Planning (KETEP) grant funded by the Korean government (MOTIE) (20203040010130, Development
and demonstration of remote intelligent operation and maintenance (O&M) technology of MW-class
solar power plant using 5G technology).
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: Not applicable.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Antonanzas, J.; Osorio, N.; Escobar, R.; Urraca, R.; Martinez-De-Pison, F.J.; Antonanzas-Torres, F. Review of Photovoltaic Power
Forecasting. Sol. Energy 2016, 136, 78–111. [CrossRef]
2. Kim, J.C.; Huh, J.H.; Ko, J.S. Improvement of MPPT Control Performance Using Fuzzy Control and VGPI in the PV System for
Micro Grid. Sustainability 2019, 11, 5891. [CrossRef]
3. Eseye, A.T.; Zhang, J.; Zheng, D. Short-Term Photovoltaic Solar Power Forecasting Using a Hybrid Wavelet-PSO-SVM Model
Based on SCADA and Meteorological Information. Renew. Energy 2018, 118, 357–367. [CrossRef]
4. Wang, H.; Yi, H.; Peng, J.; Wang, G.; Liu, Y.; Jiang, H.; Liu, W. Deterministic and Probabilistic Forecasting of Photovoltaic Power
Based on Deep Convolutional Neural Network. Energy Convers. Manag. 2017, 153, 409–422. [CrossRef]
5. Park, C.Y.; Hong, S.H.; Lim, S.C.; Song, B.S.; Park, S.W.; Huh, J.H.; Kim, J.C. Inverter Efficiency Analysis Model Based on Solar
Power Estimation Using Solar Radiation. Processes 2020, 8, 1225. [CrossRef]
6. Pelland, S.; Remund, J.; Kleissl, J.; Oozeki, T.; Brabandere, K.D. Photovoltaic and Solar Forecasting: State of the Art. IEA PVPS
Task 2013, 14, 1–36.
7. Pedro, H.T.C.; Coimbra, C.F.M. Assessment of Forecasting Techniques for Solar Power Production with no Exogenous Inputs. Sol.
Energy 2012, 86, 2017–2028. [CrossRef]
8. Rettger, P.; Keshner, M.; Pligavko, K.A.; Moore, J.; Littmann, W.B. Dynamic Management of Power Production in a Power System
Subject to Weather-Related Factors. U.S. Patent 2010/0198420 A, 5 August 2010.
9. Qing, X.; Niu, Y. Hourly Day-Ahead Solar Irradiance Prediction Using Weather Forecasts by LSTM. Energy 2018, 148, 461–468.
[CrossRef]
10. Ko, J.S.; Huh, J.H.; Kim, J.C. Overview of Maximum Power Point Tracking Methods for PV System in Micro Grid. Electronics 2020,
9, 816. [CrossRef]
11. Sobri, S.; Koohi-Kamali, S.; Rahim, N.A. Solar Photovoltaic Generation Forecasting Methods: A review. Energy Convers. Manag.
2018, 156, 459–497. [CrossRef]
Energies 2022, 15, 8233 17 of 17

12. Husein, M.; Chung, I.-Y. Day-Ahead Solar Irradiance Forecasting for Microgrids Using a Long Short-Term Memory Recurrent
Neural Network: A Deep Learning Approach. Energies 2019, 12, 1856. [CrossRef]
13. European Photovoltaic Industry Association. Connecting the Sun: Solar Photovoltaics on the Road to Large-Scale Grid Integration;
Technical Report; EPIA: Brussels, Belgium, 2012.
14. Porter, K.; Rogers, J. Survey of Variable Generation Forecasting in the West: August 2011–June 2012; National Renewable Energy
Laboratory (NREL): Golden, CO, USA, 2012.
15. Tapakis, R.; Charalambides, A. Equipment and Methodologies for Cloud Detection and Classification: A Review. Sol. Energy
2013, 95, 392–430. [CrossRef]
16. Yang, H.T.; Huang, C.M.; Huang, Y.C.; Pai, Y.S. A Weather-Based Hybrid Method for 1-Day Ahead Hourly Forecasting of PV
Power Output. IEEE Trans. Sustain. Energy 2014, 5, 917–926. [CrossRef]
17. Eftekharnejad, S.; Heydt, G.T.; Vittal, V. Optimal Generation Dispatch with High Penetration of Photovoltaic Generation. IEEE
Trans. Sustain. Energy 2015, 6, 1013–1020. [CrossRef]
18. Zhang, Q.; Tian, X.; Zhang, P.; Hou, L.; Peng, Z.; Wang, G. Solar Radiation Prediction Model for the Yellow River Basin with Deep
Learning. Agronomy 2022, 12, 1081. [CrossRef]
19. Brahma, B.; Wadhvani, R. Solar Irradiance Forecasting Based on Deep Learning Methodologies and Multi-Site Data. Symmetry
2020, 12, 1830. [CrossRef]
20. Aslam, M.; Lee, J.M.; Kim, H.S.; Lee, S.J.; Hong, S. Deep Learning Models for Long-Term Solar Radiation Forecasting Considering
Microgrid Installation: A Comparative Study. Energies 2020, 13, 147. [CrossRef]
21. Bakker, K.; Whan, K.; Knap, W.; Schmeits, M. Comparison of Statistical Post-Processing Methods for Probabilistic NWP Forecasts
of Solar Radiation. Sol. Energy 2019, 191, 138–150. [CrossRef]
22. Yeom, J.M.; Deo, R.C.; Adamwoski, J.F.; Chae, T.; Kim, D.S.; Han, K.S.; Kim, D.Y. Exploring Solar and Wind Energy Resources
in North Korea with COMS MI Geostationary Satellite Data Coupled with Numerical Weather Prediction Reanalysis Variables.
Renew. Sustain. Energy Rev. 2020, 119, 109570. [CrossRef]
23. Das, U.K.; Tey, K.S.; Seyedmahmoudian, M.; Mekhilef, S.; Idris, M.Y.I.; Van Deventer, W.; Stojcevski, A. Forecasting of Photovoltaic
Power Generation and Model Optimization: A review. Renew. Sustain. Energy Rev. 2018, 81, 912–928. [CrossRef]
24. Kushwaha, V.; Pindoriya, N.M. A SARIMA-RVFL Hybrid Model Assisted by Wavelet Decomposition for Very Short-Term Solar
PV Power Generation Forecast. Renew. Energy 2019, 140, 124–139. [CrossRef]
25. Lamsal, D.; Sreeram, V.; Mishra, Y.; Kumar, D. Kalman Filter Approach for Dispatching and Attenuating the Power Fluctuation of
Wind and Photovoltaic Power Generating Systems. IET Gener. Transm. Distrib. 2018, 12, 1501–1508. [CrossRef]
26. Miao, S.; Ning, G.; Gu, Y.; Yan, J.; Ma, B. Markov Chain Model for Solar Farm Generation and its Application to Generation
Performance Evaluation. J. Clean. Prod. 2018, 186, 905–917. [CrossRef]
27. Wang, F.; Mi, Z.; Su, S.; Zhao, H. Short-Term Solar Irradiance Forecasting Model Based on Artificial Neural Network Using
Statistical Feature Parameters. Energies 2012, 5, 1355–1370. [CrossRef]
28. Liu, J.; Fang, W.; Zhang, X.; Yang, C. An Improved Photovoltaic Power Forecasting Model with the Assistance of Aerosol Index
Data. IEEE Trans. Sustain. Energy 2015, 6, 434–442. [CrossRef]
29. Abuella, M.; Chowdhury, B. Random Forest Ensemble of Support Vector Regression Models for Solar Power Forecasting. In
Proceedings of the IEEE Power & Energy Society Innovative Smart Grid Technologies Conference (ISGT), Washington, DC, USA,
23–26 April 2017; pp. 1–5.
30. Voyant, C.; Notton, G.; Kalogirou, S.; Nivet, M.L.; Paoli, C.; Motte, F.; Fouilloy, A. Machine Learning Methods for Solar Radiation
Forecasting: A Review. Renew. Energy 2017, 105, 569–582. [CrossRef]
31. Benali, L.; Notton, G.; Fouilloy, A.; Voyant, C.; Dizene, R. Solar Radiation Forecasting Using Artificial Neural Network and
Random Forest Methods: Application to Normal Beam, Horizontal Diffuse and Global Components. Renew. Energy 2019, 132,
871–884. [CrossRef]
32. Ghimire, S.; Deo, R.C.; Downs, N.J.; Raj, N. Global Solar Radiation Prediction by ANN Integrated with European Centre for
Medium Range Weather Forecast Fields in Solar Rich Cities of Queensland Australia. J. Clean. Prod. 2019, 216, 288–310. [CrossRef]
33. Roumpakias, E.; Stamatelos, T. Prediction of a Grid-Connected Photovoltaic Park’s Output with Artificial Neural Networks
Trained by Actual Performance Data. Appl. Sci. 2022, 12, 6458. [CrossRef]
34. Ahmed, R.; Sreeram, V.; Mishra, Y.; Arif, M.D. A Review and Evaluation of the State-of-the-Art in PV Solar Power Forecasting:
Techniques and Optimization. Renew. Sustain. Energy Rev. 2020, 124, 109792. [CrossRef]
35. Noh, S.H. Analysis of Gradient Vanishing of RNNs and Performance Comparison. Information 2021, 12, 442. [CrossRef]
36. Beigi, M.; Beigi Harchegani, H.; Torki, M.; Kaveh, M.; Szymanek, M.; Khalife, E.; Dziwulski, J. Forecasting of Power Output of a
PVPS Based on Meteorological Data Using RNN Approaches. Sustainability 2022, 14, 3104. [CrossRef]
37. Son, N. Comparison of the Deep Learning Performance for Short-Term Power Load Forecasting. Sustainability 2021, 13, 12493.
[CrossRef]
38. Available online: https://www.kma.go.kr/kma/biz/forecast05.jsp (accessed on 26 August 2022).