Deep Learning Neural Networks for Short-Term PV Power Forecasting via Sky Image Method

Abstract

Solar photovoltaic (PV) power generation is prone to drastic changes due to cloud cover. The power is easily affected within a very short period of time. Thus, the accuracy of grasping cloud distribution is important for PV power forecasting. This study proposes a novel sky image method to obtain the cloud coverage rate used for short-term PV power forecasting. The authors developed an image analysis algorithm from the sky images obtained by an on-site whole sky imager (WSI). To verify the effectiveness of cloud coverage rate as the parameter for PV power forecast, four different combinations of weather features were used to compare the accuracy of short-term PV power forecasting. In addition to the artificial neural network (ANN) model, long short-term memory (LSTM) and the gated recurrent unit (GRU) were also introduced to compare their applicability conditions. After a comprehensive analysis, the coverage rate is the key weather feature, which can improve the accuracy by about 2% compared to the case without coverage feature. It also indicates that the LSTM and GRU models revealed better forecast results under different weather conditions, meaning that the cloud coverage rate proposed in this study has a significant benefit for short-term PV power forecasting.

Full text

Citation: Kuo, W.-C.; Chen, C.-H.;
Chen, S.-Y.; Wang, C.-C. Deep
Learning Neural Networks for
Short-Term PV Power Forecasting via
Sky Image Method. Energies 2022,15,
4779. https://doi.org/10.3390/
en15134779
Academic Editor: Antonino
Laudani
Received: 30 May 2022
Accepted: 27 June 2022
Published: 29 June 2022
Publisher’s Note: MDPI stays neutral
with regard to jurisdictional claims in
published maps and institutional affil-
iations.
Copyright: © 2022 by the authors.
Licensee MDPI, Basel, Switzerland.
This article is an open access article
distributed under the terms and
conditions of the Creative Commons
Attribution (CC BY) license (https://
creativecommons.org/licenses/by/
4.0/).
energies
Article
Deep Learning Neural Networks for Short-Term PV Power
Forecasting via Sky Image Method
Wen-Chi Kuo 1,* , Chiun-Hsun Chen 2, Sih-Yu Chen 1and Chi-Chuan Wang 1
1Department of Mechanical Engineering, National Yang Ming Chiao Tung University, Hsinchu 300, Taiwan;
h860731abc.me08g@nctu.edu.tw (S.-Y.C.); ccwang@nycu.edu.tw (C.-C.W.)
2Department of Aerospace and Systems Engineering, Feng Chia University, Taichung 407, Taiwan;
chiunhchen@fcu.edu.tw
*Correspondence: wenchi.en08@nycu.edu.tw
Abstract:
Solar photovoltaic (PV) power generation is prone to drastic changes due to cloud cover.
The power is easily affected within a very short period of time. Thus, the accuracy of grasping
cloud distribution is important for PV power forecasting. This study proposes a novel sky image
method to obtain the cloud coverage rate used for short-term PV power forecasting. The authors
developed an image analysis algorithm from the sky images obtained by an on-site whole sky imager
(WSI). To verify the effectiveness of cloud coverage rate as the parameter for PV power forecast,
four different combinations of weather features were used to compare the accuracy of short-term PV
power forecasting. In addition to the artificial neural network (ANN) model, long short-term memory
(LSTM) and the gated recurrent unit (GRU) were also introduced to compare their applicability
conditions. After a comprehensive analysis, the coverage rate is the key weather feature, which can
improve the accuracy by about 2% compared to the case without coverage feature. It also indicates
that the LSTM and GRU models revealed better forecast results under different weather conditions,
meaning that the cloud coverage rate proposed in this study has a significant benefit for short-term
PV power forecasting.
Keywords: deep learning (DL); forecasting; neural network; renewable energy; solar power genera-
tion; sky image
1. Introduction
Solar photovoltaics (PVs) have an intermittent characteristic and are easily affected by
the climate environment, cloud cover, and sunlight intensity. These weather conditions will
cause the instability of renewable energy power generation [
1
]. Therefore, it is imperative
to provide good weather forecasting to master the power generation of renewable energy.
Furthermore, the weather parameters in different forecasting time horizons are also slightly
different. The input meteorological factors of short-term forecasting are usually cloud cover
and wind direction. The weather parameters of medium-term forecasting are temperature,
humidity, irradiance intensity, wind speed, and wind direction from the site of some
meteorological equipment in the local area. For the long-term forecasting, the long-term
observation data of the weather forecast needs to be considered [
2
,
3
]. There have been
many studies that have explored the influence of weather parameters on the accuracy of
PV power forecasting. This research includes finding the best weather influencing factor,
comparing the forecasting deviation under different weather, or adding parameters that
worsen the forecasting results [4–7].
The cloud coverage and solar irradiation prediction are the important weather param-
eters that directly influence the variety of solar photovoltaic power from minutes to hours.
The cloud coverage and solar irradiation can be obtained from the analysis of the sky im-
age. In solar irradiance prediction, image pixels are processed to identify clear/thin/thick
Energies 2022,15, 4779. https://doi.org/10.3390/en15134779 https://www.mdpi.com/journal/energies

Energies 2022,15, 4779 2 of 17
clouds, the function of the solar pixel angle (SPA), image zenith angle (IZA), and solar
zenith angle (SZA) [
8
–
10
]. In addition, the solar irradiation prediction value is obtained
through analyzing the difference in the picture pixel brightness, the difference in the contin-
uous picture pixel and cloud image trajectory, and used as one of the learning parameters
for solar photovoltaic forecasting [11–15].
However, pyrheliometer equipment is installed in most of the PV plants, which can
easily obtain the ultra-violet index (UVI) value, which represents the solar radiation index.
Therefore, the prior research directly used the sky image to analyze the cloud trajectory
or cloud cover and directly used it as a forecasting method for PV power generation. In
particular, Zhang et al. [
16
] used CNN, integrating sky images by stacking four exposures in
grayscale images for the 60 s before the current time t with a 15 s interval. The result showed
that using the LSTM-based model yielded a 21% RMSE skill score that outperformed other
approaches. Kong et al. [
17
] proposed a novel approach based on deep whole-sky-image
learning architectures for 4 to 20 min of solar photovoltaic generation forecasting. The
experimental results showed that the hybrid static image forecaster provided superior
performance, up to an 8.3% improvement in general and up to 32.8% improvement in
the cases of ramp events when compared to that without sky images. Both Wang et al.
and Zhen et al. [
18
,
19
] proposed a hybrid mapping model for solar PV power forecasting,
and their models yielded higher accuracy than the CNN, LSTM, and ANN models. In
the data processing, they filtered the noises of the original sky image and reduced the
image resolution. However, they trained the dataset in three different ways including
color images with a 3D form, stitching of the three R, G, B channels, and all pixel values
of the image. On the other hand, Wen et al. [
20
] used the sky image parameter in solar
photovoltaic prediction for industrial applications. The results showed that it could be used
as a reference basis for smooth photovoltaic power generation control upon cloudy and
sunny days, thereby reducing the energy curtailment. Furthermore, the sky image analysis
methods they used are complicated and difficult to implement. Therefore, the present sky
image analysis method is much simpler and easier to implement in this paper.
This paper proposes a new and interesting image pixel calculation formula for image
processing. It obtains the cloud coverage rate per minute based on the defined threshold as
one of the PV power forecasting parameters. The advantage of the proposed method in
comparison with the previous literature is that these values are easy to parse and sequence
for the time series model. Furthermore, the collected data can be easily integrated with
other weather data quickly and introduced into the PV forecasting deep learning model
to learn and forecast. In particular, the cloud coverage rate can improve the PV power
forecasting accuracy of this study. The structure of this article is as follows. Section 2
describes the solar photovoltaic forecasting framework and the local weather equipment.
Section 3focuses on the proposed sky image coverage analysis method, and the deep
learning models are introduced. Section 4presents and discusses the experimental results
with different weather combinations. Finally, Section 5concludes this article.
2. Forecasting Procedures and Data Set
In this section, the PV forecasting operation procedures is introduced, as shown in
Figure 1, which contains four main processes including the data collection, data processing,
model building and training, and forecasting processes. The details are described as
follows.

Energies 2022,15, 4779 3 of 17
Energies 2022, 15, x FOR PEER REVIEW 3 of 17
Figure 1. The PV power forecasting procedures, introducing Data Collection, Data Preprocessing,
Model Building & Training, and Forecasting Processes.
2.1. Data Collection
The equipment used for data collection in this paper is shown in Figure 2 including
the power data of the 8 kW PV plants, the UVI data from the pyrheliometer, local micro
weather station (contains wind speed and direction, temperature, relative humidity), and
the whole sky imager (WSI). Furthermore, in the overall dataset in this study, the collec-
tion period of the weather and PV data was from 21 February 2021 to 13 June 2021.
Figure 1.
The PV power forecasting procedures, introducing Data Collection, Data Preprocessing,
Model Building & Training, and Forecasting Processes.
2.1. Data Collection
The equipment used for data collection in this paper is shown in Figure 2including
the power data of the 8 kW PV plants, the UVI data from the pyrheliometer, local micro
weather station (contains wind speed and direction, temperature, relative humidity), and
the whole sky imager (WSI). Furthermore, in the overall dataset in this study, the collection
period of the weather and PV data was from 21 February 2021 to 13 June 2021.
Energies 2022, 15, x FOR PEER REVIEW 3 of 17
Figure 1. The PV power forecasting procedures, introducing Data Collection, Data Preprocessing,
Model Building & Training, and Forecasting Processes.
2.1. Data Collection
The equipment used for data collection in this paper is shown in Figure 2 including
the power data of the 8 kW PV plants, the UVI data from the pyrheliometer, local micro
weather station (contains wind speed and direction, temperature, relative humidity), and
the whole sky imager (WSI). Furthermore, in the overall dataset in this study, the collec-
tion period of the weather and PV data was from 21 February 2021 to 13 June 2021.
Figure 2.
The use of the equipment (PV inverter, Weather Station, Whole Sky Imagers) and data
collection schema (using Modbus TCP/IP).

Energies 2022,15, 4779 4 of 17
2.2. Data Processing
The purpose of the present work was to calculate the cloud coverage rate through
the sky image. For simplicity, the authors preprocessed the sky image to exclude the
unnecessary objects in the image including trees, buildings, and the outside black areas
of the image. Then, the information was fed into the proposed RGB sky image processing
method. This pretreatment converts each image into a numerical data type. It calculates
the cloud coverage rate as the weather characteristic value. In the meantime, the proposed
method also preprocesses the weather data from the local small weather station and
examines whether it contains missing data to ensure the photovoltaic data integrity.
2.3. Model Building and Training Processes
Establishing a deep learning model: The deep learning neural network model based
on the ANN, LSTM, and GRU model were applied in this study. In addition, determining
the model hyperparameters is essential. Thus, the parameter turning of hidden units,
training steps, and epochs are incorporated to ensure that there is no overfitting of these
models to maintain the suitability and predictive accuracy of hyperparameters.
2.4. Forecasting Processes
This part uses the model that has already trained hyperparameter and weights for
short-term PV power forecasting. On the other hand, the authors used four different
weather feature combinations and verified the prediction performance of cloud coverage
rate as the parameter for short-term PV power forecasting including Case 1: Six weather
values (wind speed and direction, UVI, temperature, relative humidity, coverage rate); Case
2: Five weather value (without coverage rate); Case 3: Five weather values (without UVI);
and Case 4: Only coverage rate and relative humidity. The predictive abilities of these three
models for short-term PV power forecasting were compared in the final result.
3. Methodology
A new sky image analysis method for short-term PV power generation forecasting
is proposed in this article. In this section, sky image coverage processing, the cloud
computing method, deep learning models (ANN, LSTM, GRU), and evaluation indices will
be introduced separately.
3.1. Sky Image Coverage Processing Method
The sky image processing of this research focuses on RGB pixel composition. In RGB,
the grayscale color refers to converting a color picture into a grayscale picture that is still
not far from the original color, as shown in Equation (1). Since each picture is composed
of different color components, the authors defined
∆
(pixel composition) to represent the
three combined color difference results, and is expressed in Equation (2):
R≈G≈B (1)
∆=(R−G)2+(R−B)2+(G−B)2(2)
Furthermore, to perform the RGB distribution analysis of each element image under
different weather patterns and to analyze the sunny, mostly clear, and cloudy images before
performing the sky image analysis, the authors adopted the Open-Source Computer Vision
(OpenCV) library in python to analyze the sky image and determine the RGB distribution
ratio. The analysis and statistical results are shown in Figure 3.

Energies 2022,15, 4779 5 of 17
Energies 2022, 15, x FOR PEER REVIEW 5 of 17
Figure 3. The results of the average distribution of each element.
After analyzing the color composition, the difference between the blue and green
pixels is about one δ variable. Moreover, the difference between blue and red pixels is
about two δ variables, and the mathematical relationship between the three combined col-
ors is listed in Equation (3). Substituting Equation (3) into Equation (2), subject to some
mathematical manipulation, yields Equation (4).
∀δ∈ℕ:(B−G)=δ⇒(B−R)≅2δ (3)
∆ = (δ)+(4δ)+(δ) (4)
The statistical analysis shows that the change in ∆ is mainly because of the color dif-
ference between the red and blue pixels. Therefore, the deviation in the green pixel in the
sky image analysis occurs. In order to maintain the same importance among the three
pixels (R, G, B), the authors employed the weighting coefficient method to adjust the re-
lationship of the three pixels. Finally, this study proposes the general mathematical for-
mula of the RGB analysis method as follows:
∆ = 4 × (R − G)+(R−B)+4×(G−B) (5)
Figure 3. The results of the average distribution of each element.
After analyzing the color composition, the difference between the blue and green
pixels is about one
δ
variable. Moreover, the difference between blue and red pixels is
about two
δ
variables, and the mathematical relationship between the three combined
colors is listed in Equation (3). Substituting Equation (3) into Equation (2), subject to some
mathematical manipulation, yields Equation (4).
∀δ∈N:(B−G)=δ⇒(B−R)∼
=2δ(3)
∆=(δ)2+(4δ)2+(δ)2(4)
The statistical analysis shows that the change in
∆
is mainly because of the color
difference between the red and blue pixels. Therefore, the deviation in the green pixel
in the sky image analysis occurs. In order to maintain the same importance among the
three pixels (R, G, B), the authors employed the weighting coefficient method to adjust
the relationship of the three pixels. Finally, this study proposes the general mathematical
formula of the RGB analysis method as follows:
∆=4×(R−G)2+(R−B)2+4×(G−B)2(5)
3.2. Cloud Covering Analysis Structure and Calculation Method
The cloud covering analysis structure is shown in Figure 4. This process can be divided
into the following four parts:
(1)
Calculate the Pixel Composition

Energies 2022,15, 4779 6 of 17
Energies 2022, 15, x FOR PEER REVIEW 6 of 17
3.2. Cloud Covering Analysis Structure and Calculation Method
The cloud covering analysis structure is shown in Figure 4. This process can be di-
vided into the following four parts:
(1) Calculate the Pixel Composition
Calculate the pixel composition and set up the initial threshold (TS) value: The pixel
composition is determined by Equation (5). The initial value of TS is random, and it will
increase by one after each cycle until the coverage condition is reached.
(2) Define the Sky Image Covering Limiting
Define the sky image covering limiting condition: By analyzing the pixel composi-
tion, the authors defined the limit condition of the sun, clouds, and sky as listed in Table
1. Sunny conditions prevail when the pixel composition exceeds TS, and the R, G, B ele-
ments are greater than 760. Conversely, it is regarded as cloudy if the value is less than
760. Furthermore, sky conditions occur if the pixel composition is less than or equal to TS
and the blue value is larger than the red and green values at the same time.
Figure 4. The cloud covering process.
Figure 4. The cloud covering process.
Calculate the pixel composition and set up the initial threshold (TS) value: The pixel
composition is determined by Equation (5). The initial value of TS is random, and it will
increase by one after each cycle until the coverage condition is reached.
(2)
Define the Sky Image Covering Limiting
Define the sky image covering limiting condition: By analyzing the pixel composition,
the authors defined the limit condition of the sun, clouds, and sky as listed in Table 1.
Sunny conditions prevail when the pixel composition exceeds TS, and the R, G, B elements
are greater than 760. Conversely, it is regarded as cloudy if the value is less than 760.
Furthermore, sky conditions occur if the pixel composition is less than or equal to TS and
the blue value is larger than the red and green values at the same time.
(3)
Cloud Covering Calculation
The authors defined the cloud coverage rate in Equation (6). In addition, there will be
a similar coverage rate in a specific range by analyzing the time-series sky images through
testing. Therefore, when the coverage rate increases and the slope decreases, the program
will record the coverage rate as the current value in this cycle. Table 2depicts the result of
the automatic convergence example. From this table, when TS is 16, the cloud coverage
rate is 49.95%, and when TS is 18, the cloud coverage rate changes by 15%. In addition,

Energies 2022,15, 4779 7 of 17
the sky image shows that it will be over-covered when the coverage rate = 61.53% and the
TS = 18, thereby revealing the non-cloud situations.
Cloud coverage rate(%) = [1−all pixels −mask0s pixels −cloud0s pixels
all pixels −cloud0s pixels](6)
Table 1. The sky image covering limiting condition.
Item Limit Condition Pixel Status Pixel Status
∆(pixel composition)
≥
Threshold (TS)
∑(R,G,B)>760 Sun Remain
∑(R,G,B)≤760 Cloud Cover
∆(pixel composition)
<
Threshold (TS)
B>R
∩
B>G
Sky Remain
Others Ex: Tree Cover
Table 2. The results of the automatic convergence.
TS Sky Image Cover Result Coverage Rate
12
Energies 2022, 15, x. https://doi.org/10.3390/xxxxx www.mdpi.com/journal/energies
TS Sky Image Cover Result Coverage Rate
12
32.52%
14
44.83%
16
45.95%
18
61.53%
(1) Update the Threshold Value
Update the threshold value: In this study, updating the threshold value for every 15
min was adopted to accelerate the cloud image analysis. Through the testing experience,
each coverage analysis was about 6 to 7 s, which was sufficient for short-term photovol-
taics forecasting.
32.52%
14
Energies 2022, 15, x. https://doi.org/10.3390/xxxxx www.mdpi.com/journal/energies
TS Sky Image Cover Result Coverage Rate
12
32.52%
14
44.83%
16
45.95%
18
61.53%
(1) Update the Threshold Value
Update the threshold value: In this study, updating the threshold value for every 15
min was adopted to accelerate the cloud image analysis. Through the testing experience,
each coverage analysis was about 6 to 7 s, which was sufficient for short-term photovol-
taics forecasting.
44.83%
16
Energies 2022, 15, x. https://doi.org/10.3390/xxxxx www.mdpi.com/journal/energies
TS Sky Image Cover Result Coverage Rate
12
32.52%
14
44.83%
16
45.95%
18
61.53%
(1) Update the Threshold Value
Update the threshold value: In this study, updating the threshold value for every 15
min was adopted to accelerate the cloud image analysis. Through the testing experience,
each coverage analysis was about 6 to 7 s, which was sufficient for short-term photovol-
taics forecasting.
45.95%
18
Energies 2022, 15, x. https://doi.org/10.3390/xxxxx www.mdpi.com/journal/energies
TS Sky Image Cover Result Coverage Rate
12
32.52%
14
44.83%
16
45.95%
18
61.53%
(1) Update the Threshold Value
Update the threshold value: In this study, updating the threshold value for every 15
min was adopted to accelerate the cloud image analysis. Through the testing experience,
each coverage analysis was about 6 to 7 s, which was sufficient for short-term photovol-
taics forecasting.
61.53%
(4)
Update the Threshold Value
Update the threshold value: In this study, updating the threshold value for every
15 min
was adopted to accelerate the cloud image analysis. Through the testing experience,
each coverage analysis was about 6 to 7 s, which was sufficient for short-term photovoltaics
forecasting.
3.3. Choice of The Deep Learning Model
The authors used artificial neural networks and recurrent neural network models
to train and predict the PV power, which are the ANN, LSTM, and GRU models. The
ANN model has a strong nonlinear fitting ability and strong adaptive ability. The LSTM
can control the transmission state through the gated state, remembering the unimportant

Energies 2022,15, 4779 8 of 17
information that needs to be remembered for a long time. The GRU has the same effect as
LSTM, but needs fewer parameters to build models with a large amount of training [
21
–
23
].
The models have their advantages in training and forecasting. During the training
process, all choices were performed by trying different hyperparameters on the same
training set including hidden units, training steps (Epochs), and input time intervals. Thus,
the authors will discuss which model is more suitable for this field based on the short-term
PV power forecast results in this paper.
3.4. Evaluation Indices
The forecasting performance of PV power prediction models is evaluated using three
statistical indicators, which are the mean absolute error (MAE), root mean squared error
(RMSE), and mean absolute percentage error (MAPE) [
24
]. Their corresponding formulas
are given by Equations (7)–(9):
MAE =1
N
N
∑
i=1
|ˆ
Xi−Xi|(7)
RMSE =v
u
u
t
1
N
N
∑
i=1
(ˆ
Xi−Xi)2(8)
MAPE =1
N
N
∑
i=1
|Xi−ˆ
Xi
Xi
| × 100% (9)
where
ˆ
Xi
and
Xi
represent the ith forecasted and actual value, respectively, and Nis the
size of the test dataset.
4. Numerical Results
This section evaluates the influence of the coverage rate under different weather
conditions on the performance of one-hour ahead PV power generation forecasting. The
numerical results were comprehensively compared and subsequently discussed.
4.1. Results of Sky Image Processing by RGB Formula
This study proposed two RGB formulas (Equations (2) and (5)) for the sky image
processing. In this section, the authors simulated and compared the threshold ratio of
Equations (2) and (5), which is (1 + 1 + 1):(4 + 1 + 4) = 1:3. Therefore, in the sky image
analysis, the threshold values will represent TS = n
×
30 and TS = n
×
90, where n is a
constant 1, 2, 3
. . .
. Note that it is quite challenging to analyze the difference between the
two RGB analysis methods on sunny days because there were fewer clouds under sufficient
sunlight. However, there was a noticeable difference in cloudy days. Specifically, when
n increased continuously, the difference in cloud covering will become more apparent, as
shown in Tables 3and 4. The result indicates that Equation (5) is superior to Equation (2),
especially on a cloudy day. From the result, Equation (5) is more sensitive and accurate,
and it is the main RGB formula used in this article.

Energies 2022,15, 4779 9 of 17
Table 3. The results of the RGB comparison on a sunny day.
Index Equations (2): TS = n ×30 Equations (5): TS = n ×90
n=1
Energies 2022, 15, x FOR PEER REVIEW 10 of 19
Table 3. The results of the RGB comparison on a sunny day.
Index Equations (2): TS = n × 30 Equations (5): TS = n × 90
n = 1
n = 2
n = 3
Table 4. The results of the RGB comparison on a cloudy day.
Index Equations (2): TS = n × 30 Equations (5): TS = n × 90
n = 1
n = 2
n = 3
4.2. Results of Adjust Hyperparameters
The adjustment of hyperparameters in deep learning has a specific impact on the
training and predictive results. Therefore, the authors first used 21 days of weather infor-
mation (starting from 24February) for training and testing including hidden units, train-
ing steps (Epochs), and input time intervals to find the most suitable hyperparameter com-
binations for the three models (ANN, LSTM, GRU). All data were sorted by the minute.
The fixed parameters were learning rate (learning rate: 1 × 10−3) and batch size (batch size:
16). Finally, three evaluation indices (MAE, RMSE, and MAPE) were used to evaluate the
model. The results showed that the ANN model used five hidden layers, the number of
Epochs was 3000, and data input interval was one hour; in the LSTM model, it used seven
hidden layers, the number of Epochs was 2500, and the data input interval was 2 h; in the
GRU model, it used six hidden layers, the number of Epochs was 1500, and data input
interval was 3 h. The results of the hyperparameter adjustment are shown in Table 5.
Energies 2022, 15, x FOR PEER REVIEW 10 of 19
Table 3. The results of the RGB comparison on a sunny day.
Index Equations (2): TS = n × 30 Equations (5): TS = n × 90
n = 1
n = 2
n = 3
Table 4. The results of the RGB comparison on a cloudy day.
Index Equations (2): TS = n × 30 Equations (5): TS = n × 90
n = 1
n = 2
n = 3
4.2. Results of Adjust Hyperparameters
The adjustment of hyperparameters in deep learning has a specific impact on the
training and predictive results. Therefore, the authors first used 21 days of weather infor-
mation (starting from 24February) for training and testing including hidden units, train-
ing steps (Epochs), and input time intervals to find the most suitable hyperparameter com-
binations for the three models (ANN, LSTM, GRU). All data were sorted by the minute.
The fixed parameters were learning rate (learning rate: 1 × 10−3) and batch size (batch size:
16). Finally, three evaluation indices (MAE, RMSE, and MAPE) were used to evaluate the
model. The results showed that the ANN model used five hidden layers, the number of
Epochs was 3000, and data input interval was one hour; in the LSTM model, it used seven
hidden layers, the number of Epochs was 2500, and the data input interval was 2 h; in the
GRU model, it used six hidden layers, the number of Epochs was 1500, and data input
interval was 3 h. The results of the hyperparameter adjustment are shown in Table 5.
n=2
Energies 2022, 15, x FOR PEER REVIEW 10 of 19
Table 3. The results of the RGB comparison on a sunny day.
Index Equations (2): TS = n × 30 Equations (5): TS = n × 90
n = 1
n = 2
n = 3
Table 4. The results of the RGB comparison on a cloudy day.
Index Equations (2): TS = n × 30 Equations (5): TS = n × 90
n = 1
n = 2
n = 3
4.2. Results of Adjust Hyperparameters
The adjustment of hyperparameters in deep learning has a specific impact on the
training and predictive results. Therefore, the authors first used 21 days of weather infor-
mation (starting from 24February) for training and testing including hidden units, train-
ing steps (Epochs), and input time intervals to find the most suitable hyperparameter com-
binations for the three models (ANN, LSTM, GRU). All data were sorted by the minute.
The fixed parameters were learning rate (learning rate: 1 × 10−3) and batch size (batch size:
16). Finally, three evaluation indices (MAE, RMSE, and MAPE) were used to evaluate the
model. The results showed that the ANN model used five hidden layers, the number of
Epochs was 3000, and data input interval was one hour; in the LSTM model, it used seven
hidden layers, the number of Epochs was 2500, and the data input interval was 2 h; in the
GRU model, it used six hidden layers, the number of Epochs was 1500, and data input
interval was 3 h. The results of the hyperparameter adjustment are shown in Table 5.
Energies 2022, 15, x FOR PEER REVIEW 10 of 19
Table 3. The results of the RGB comparison on a sunny day.
Index Equations (2): TS = n × 30 Equations (5): TS = n × 90
n = 1
n = 2
n = 3
Table 4. The results of the RGB comparison on a cloudy day.
Index Equations (2): TS = n × 30 Equations (5): TS = n × 90
n = 1
n = 2
n = 3
4.2. Results of Adjust Hyperparameters
The adjustment of hyperparameters in deep learning has a specific impact on the
training and predictive results. Therefore, the authors first used 21 days of weather infor-
mation (starting from 24February) for training and testing including hidden units, train-
ing steps (Epochs), and input time intervals to find the most suitable hyperparameter com-
binations for the three models (ANN, LSTM, GRU). All data were sorted by the minute.
The fixed parameters were learning rate (learning rate: 1 × 10−3) and batch size (batch size:
16). Finally, three evaluation indices (MAE, RMSE, and MAPE) were used to evaluate the
model. The results showed that the ANN model used five hidden layers, the number of
Epochs was 3000, and data input interval was one hour; in the LSTM model, it used seven
hidden layers, the number of Epochs was 2500, and the data input interval was 2 h; in the
GRU model, it used six hidden layers, the number of Epochs was 1500, and data input
interval was 3 h. The results of the hyperparameter adjustment are shown in Table 5.
n=3
Energies 2022, 15, x FOR PEER REVIEW 10 of 19
Table 3. The results of the RGB comparison on a sunny day.
Index Equations (2): TS = n × 30 Equations (5): TS = n × 90
n = 1
n = 2
n = 3
Table 4. The results of the RGB comparison on a cloudy day.
Index Equations (2): TS = n × 30 Equations (5): TS = n × 90
n = 1
n = 2
n = 3
4.2. Results of Adjust Hyperparameters
The adjustment of hyperparameters in deep learning has a specific impact on the
training and predictive results. Therefore, the authors first used 21 days of weather infor-
mation (starting from 24February) for training and testing including hidden units, train-
ing steps (Epochs), and input time intervals to find the most suitable hyperparameter com-
binations for the three models (ANN, LSTM, GRU). All data were sorted by the minute.
The fixed parameters were learning rate (learning rate: 1 × 10−3) and batch size (batch size:
16). Finally, three evaluation indices (MAE, RMSE, and MAPE) were used to evaluate the
model. The results showed that the ANN model used five hidden layers, the number of
Epochs was 3000, and data input interval was one hour; in the LSTM model, it used seven
hidden layers, the number of Epochs was 2500, and the data input interval was 2 h; in the
GRU model, it used six hidden layers, the number of Epochs was 1500, and data input
interval was 3 h. The results of the hyperparameter adjustment are shown in Table 5.
Energies 2022, 15, x FOR PEER REVIEW 10 of 19
Table 3. The results of the RGB comparison on a sunny day.
Index Equations (2): TS = n × 30 Equations (5): TS = n × 90
n = 1
n = 2
n = 3
Table 4. The results of the RGB comparison on a cloudy day.
Index Equations (2): TS = n × 30 Equations (5): TS = n × 90
n = 1
n = 2
n = 3
4.2. Results of Adjust Hyperparameters
The adjustment of hyperparameters in deep learning has a specific impact on the
training and predictive results. Therefore, the authors first used 21 days of weather infor-
mation (starting from 24February) for training and testing including hidden units, train-
ing steps (Epochs), and input time intervals to find the most suitable hyperparameter com-
binations for the three models (ANN, LSTM, GRU). All data were sorted by the minute.
The fixed parameters were learning rate (learning rate: 1 × 10−3) and batch size (batch size:
16). Finally, three evaluation indices (MAE, RMSE, and MAPE) were used to evaluate the
model. The results showed that the ANN model used five hidden layers, the number of
Epochs was 3000, and data input interval was one hour; in the LSTM model, it used seven
hidden layers, the number of Epochs was 2500, and the data input interval was 2 h; in the
GRU model, it used six hidden layers, the number of Epochs was 1500, and data input
interval was 3 h. The results of the hyperparameter adjustment are shown in Table 5.
Table 4. The results of the RGB comparison on a cloudy day.
Index Equations (2): TS = n ×30 Equations (5): TS = n ×90
n=1
Energies 2022, 15, x FOR PEER REVIEW 10 of 19
Table 3. The results of the RGB comparison on a sunny day.
Index Equations (2): TS = n × 30 Equations (5): TS = n × 90
n = 1
n = 2
n = 3
Table 4. The results of the RGB comparison on a cloudy day.
Index Equations (2): TS = n × 30 Equations (5): TS = n × 90
n = 1
n = 2
n = 3
4.2. Results of Adjust Hyperparameters
The adjustment of hyperparameters in deep learning has a specific impact on the
training and predictive results. Therefore, the authors first used 21 days of weather infor-
mation (starting from 24February) for training and testing including hidden units, train-
ing steps (Epochs), and input time intervals to find the most suitable hyperparameter com-
binations for the three models (ANN, LSTM, GRU). All data were sorted by the minute.
The fixed parameters were learning rate (learning rate: 1 × 10−3) and batch size (batch size:
16). Finally, three evaluation indices (MAE, RMSE, and MAPE) were used to evaluate the
model. The results showed that the ANN model used five hidden layers, the number of
Epochs was 3000, and data input interval was one hour; in the LSTM model, it used seven
hidden layers, the number of Epochs was 2500, and the data input interval was 2 h; in the
GRU model, it used six hidden layers, the number of Epochs was 1500, and data input
interval was 3 h. The results of the hyperparameter adjustment are shown in Table 5.
Energies 2022, 15, x FOR PEER REVIEW 10 of 19
Table 3. The results of the RGB comparison on a sunny day.
Index Equations (2): TS = n × 30 Equations (5): TS = n × 90
n = 1
n = 2
n = 3
Table 4. The results of the RGB comparison on a cloudy day.
Index Equations (2): TS = n × 30 Equations (5): TS = n × 90
n = 1
n = 2
n = 3
4.2. Results of Adjust Hyperparameters
The adjustment of hyperparameters in deep learning has a specific impact on the
training and predictive results. Therefore, the authors first used 21 days of weather infor-
mation (starting from 24February) for training and testing including hidden units, train-
ing steps (Epochs), and input time intervals to find the most suitable hyperparameter com-
binations for the three models (ANN, LSTM, GRU). All data were sorted by the minute.
The fixed parameters were learning rate (learning rate: 1 × 10−3) and batch size (batch size:
16). Finally, three evaluation indices (MAE, RMSE, and MAPE) were used to evaluate the
model. The results showed that the ANN model used five hidden layers, the number of
Epochs was 3000, and data input interval was one hour; in the LSTM model, it used seven
hidden layers, the number of Epochs was 2500, and the data input interval was 2 h; in the
GRU model, it used six hidden layers, the number of Epochs was 1500, and data input
interval was 3 h. The results of the hyperparameter adjustment are shown in Table 5.
n=2
Energies 2022, 15, x FOR PEER REVIEW 10 of 19
Table 3. The results of the RGB comparison on a sunny day.
Index Equations (2): TS = n × 30 Equations (5): TS = n × 90
n = 1
n = 2
n = 3
Table 4. The results of the RGB comparison on a cloudy day.
Index Equations (2): TS = n × 30 Equations (5): TS = n × 90
n = 1
n = 2
n = 3
4.2. Results of Adjust Hyperparameters
The adjustment of hyperparameters in deep learning has a specific impact on the
training and predictive results. Therefore, the authors first used 21 days of weather infor-
mation (starting from 24February) for training and testing including hidden units, train-
ing steps (Epochs), and input time intervals to find the most suitable hyperparameter com-
binations for the three models (ANN, LSTM, GRU). All data were sorted by the minute.
The fixed parameters were learning rate (learning rate: 1 × 10−3) and batch size (batch size:
16). Finally, three evaluation indices (MAE, RMSE, and MAPE) were used to evaluate the
model. The results showed that the ANN model used five hidden layers, the number of
Epochs was 3000, and data input interval was one hour; in the LSTM model, it used seven
hidden layers, the number of Epochs was 2500, and the data input interval was 2 h; in the
GRU model, it used six hidden layers, the number of Epochs was 1500, and data input
interval was 3 h. The results of the hyperparameter adjustment are shown in Table 5.
Energies 2022, 15, x FOR PEER REVIEW 10 of 19
Table 3. The results of the RGB comparison on a sunny day.
Index Equations (2): TS = n × 30 Equations (5): TS = n × 90
n = 1
n = 2
n = 3
Table 4. The results of the RGB comparison on a cloudy day.
Index Equations (2): TS = n × 30 Equations (5): TS = n × 90
n = 1
n = 2
n = 3
4.2. Results of Adjust Hyperparameters
The adjustment of hyperparameters in deep learning has a specific impact on the
training and predictive results. Therefore, the authors first used 21 days of weather infor-
mation (starting from 24February) for training and testing including hidden units, train-
ing steps (Epochs), and input time intervals to find the most suitable hyperparameter com-
binations for the three models (ANN, LSTM, GRU). All data were sorted by the minute.
The fixed parameters were learning rate (learning rate: 1 × 10−3) and batch size (batch size:
16). Finally, three evaluation indices (MAE, RMSE, and MAPE) were used to evaluate the
model. The results showed that the ANN model used five hidden layers, the number of
Epochs was 3000, and data input interval was one hour; in the LSTM model, it used seven
hidden layers, the number of Epochs was 2500, and the data input interval was 2 h; in the
GRU model, it used six hidden layers, the number of Epochs was 1500, and data input
interval was 3 h. The results of the hyperparameter adjustment are shown in Table 5.
n=3
Energies 2022, 15, x FOR PEER REVIEW 10 of 19
Table 3. The results of the RGB comparison on a sunny day.
Index Equations (2): TS = n × 30 Equations (5): TS = n × 90
n = 1
n = 2
n = 3
Table 4. The results of the RGB comparison on a cloudy day.
Index Equations (2): TS = n × 30 Equations (5): TS = n × 90
n = 1
n = 2
n = 3
4.2. Results of Adjust Hyperparameters
The adjustment of hyperparameters in deep learning has a specific impact on the
training and predictive results. Therefore, the authors first used 21 days of weather infor-
mation (starting from 24February) for training and testing including hidden units, train-
ing steps (Epochs), and input time intervals to find the most suitable hyperparameter com-
binations for the three models (ANN, LSTM, GRU). All data were sorted by the minute.
The fixed parameters were learning rate (learning rate: 1 × 10−3) and batch size (batch size:
16). Finally, three evaluation indices (MAE, RMSE, and MAPE) were used to evaluate the
model. The results showed that the ANN model used five hidden layers, the number of
Epochs was 3000, and data input interval was one hour; in the LSTM model, it used seven
hidden layers, the number of Epochs was 2500, and the data input interval was 2 h; in the
GRU model, it used six hidden layers, the number of Epochs was 1500, and data input
interval was 3 h. The results of the hyperparameter adjustment are shown in Table 5.
Energies 2022, 15, x FOR PEER REVIEW 10 of 19
Table 3. The results of the RGB comparison on a sunny day.
Index Equations (2): TS = n × 30 Equations (5): TS = n × 90
n = 1
n = 2
n = 3
Table 4. The results of the RGB comparison on a cloudy day.
Index Equations (2): TS = n × 30 Equations (5): TS = n × 90
n = 1
n = 2
n = 3
4.2. Results of Adjust Hyperparameters
The adjustment of hyperparameters in deep learning has a specific impact on the
training and predictive results. Therefore, the authors first used 21 days of weather infor-
mation (starting from 24February) for training and testing including hidden units, train-
ing steps (Epochs), and input time intervals to find the most suitable hyperparameter com-
binations for the three models (ANN, LSTM, GRU). All data were sorted by the minute.
The fixed parameters were learning rate (learning rate: 1 × 10−3) and batch size (batch size:
16). Finally, three evaluation indices (MAE, RMSE, and MAPE) were used to evaluate the
model. The results showed that the ANN model used five hidden layers, the number of
Epochs was 3000, and data input interval was one hour; in the LSTM model, it used seven
hidden layers, the number of Epochs was 2500, and the data input interval was 2 h; in the
GRU model, it used six hidden layers, the number of Epochs was 1500, and data input
interval was 3 h. The results of the hyperparameter adjustment are shown in Table 5.
4.2. Results of Adjust Hyperparameters
The adjustment of hyperparameters in deep learning has a specific impact on the train-
ing and predictive results. Therefore, the authors first used 21 days of weather information
(starting from 24 February) for training and testing including hidden units, training steps
(Epochs), and input time intervals to find the most suitable hyperparameter combinations
for the three models (ANN, LSTM, GRU). All data were sorted by the minute. The fixed
parameters were learning rate (learning rate: 1
×
10
−3
) and batch size (batch size: 16).
Finally, three evaluation indices (MAE, RMSE, and MAPE) were used to evaluate the model.
The results showed that the ANN model used five hidden layers, the number of Epochs
was 3000, and data input interval was one hour; in the LSTM model, it used seven hidden
layers, the number of Epochs was 2500, and the data input interval was 2 h; in the GRU
model, it used six hidden layers, the number of Epochs was 1500, and data input interval
was 3 h. The results of the hyperparameter adjustment are shown in Table 5.
Table 5. The results of the adjustment of the hyperparameters.
Index Model
Adjust Parameters Fixed Parameters Evaluation
Input Time Layers Epochs Learning Rate Batch Size MAE RMSE MAPE
1 ANN 1 h 5 3000 1×10−316 0.026 0.046 18.94%
2 LSTM 2 h 7 1500 1×10−316 0.027 0.049 14.33%
3 GRU 3 h 6 2500 1×10−316 0.015 0.030 10.54%

Energies 2022,15, 4779 10 of 17
4.3. Performance Comparison with Different Weather Features on Sunny and Cloudy Days
The four different weather feature combinations for short-term solar photovoltaic
forecast analysis are as follows. Case 1: Six weather features (wind speed and direction,
UVI, temperature, relative humidity, coverage rate); Case 2: Five weather features (without
coverage rate); Case 3: Five weather features (without UVI), and Case 4: Only coverage rate
and relatively humidity. These cases aim to study whether the cloud coverage proposed
in this paper can reduce the use of features and promote the accuracy of the PV power
generation forecasts.
4.3.1. Case 1: Results of Six Weather Features
In this case, there were six weather features of the dataset: wind speed and direction,
UVI, temperature, relative humidity, and coverage rate. Figure 5intuitively shows the PV
forecast results of the three models on sunny and mostly cloudy days. The best forecasting
model was ANN on sunny days, which is more consistent with the actual tracked solar
photovoltaic power generation. In addition, the GRU and LSTM yielded large predictive
deviations between 3 and 4 PM, respectively. The preliminary assessment may be the
interaction between the weather parameters. On the other hand, it can be found that the
predictive result of ANN varied widely on mostly cloudy days. In contrast, the prediction
trend of LSTM and GRU was better. There are two possible reasons for this difference. First,
the time series forecasting model was more adaptable in the large ramp up and ramp down
situation. Second, there were less data of cloudy days and poor weather. As a result, the
ANN model showed insufficient experience upon prediction.
Energies 2022, 15, x FOR PEER REVIEW 10 of 17
4.3. Performance Comparison with Different Weather Features on Sunny and Cloudy Days
The four different weather feature combinations for short-term solar photovoltaic
forecast analysis are as follows. Case 1: Six weather features (wind speed and direction,
UVI, temperature, relative humidity, coverage rate); Case 2: Five weather features (with-
out coverage rate); Case 3: Five weather features (without UVI), and Case 4: Only coverage
rate and relatively humidity. These cases aim to study whether the cloud coverage pro-
posed in this paper can reduce the use of features and promote the accuracy of the PV
power generation forecasts.
4.3.1. Case 1: Results of Six Weather Features
In this case, there were six weather features of the dataset: wind speed and direction,
UVI, temperature, relative humidity, and coverage rate. Figure 5 intuitively shows the PV
forecast results of the three models on sunny and mostly cloudy days. The best forecasting
model was ANN on sunny days, which is more consistent with the actual tracked solar
photovoltaic power generation. In addition, the GRU and LSTM yielded large predictive
deviations between 3 and 4 PM, respectively. The preliminary assessment may be the in-
teraction between the weather parameters. On the other hand, it can be found that the
predictive result of ANN varied widely on mostly cloudy days. In contrast, the prediction
trend of LSTM and GRU was better. There are two possible reasons for this difference.
First, the time series forecasting model was more adaptable in the large ramp up and ramp
down situation. Second, there were less data of cloudy days and poor weather. As a result,
the ANN model showed insufficient experience upon prediction.
(a) (b)
Figure 5. Case 1—(a) Forecasting results of 16 March 2021(sunny). (b) Forecasting results of 21 May
2021(mostly cloudy). The frequency of data collection was once per minute.
Further evaluation and analysis results are shown in Table 6. The result indicates that
there was a better prediction effect of ANN under sunny conditions, with a MAPE of
6.54%, MAE of 0.017, and RMSE of 0.029. In contrast, the prediction effect of GRU was
better under cloudy conditions with a MAPE of 10.25%, MAE of 0.039, and RMSE of 0.078.
Based on the average of the two-day forecast results, it was found that GRU performed
better, with a MAPE of 9.02%, MAE of 0.029, and RMSE of 0.055.
Figure 5.
Case 1—(
a
) Forecasting results of 16 March 2021 (sunny). (
b
) Forecasting results of
21 May 2021
(mostly cloudy). The frequency of data collection was once per minute.
Further evaluation and analysis results are shown in Table 6. The result indicates that
there was a better prediction effect of ANN under sunny conditions, with a MAPE of 6.54%,
MAE of 0.017, and RMSE of 0.029. In contrast, the prediction effect of GRU was better
under cloudy conditions with a MAPE of 10.25%, MAE of 0.039, and RMSE of 0.078. Based
on the average of the two-day forecast results, it was found that GRU performed better,
with a MAPE of 9.02%, MAE of 0.029, and RMSE of 0.055.

Energies 2022,15, 4779 11 of 17
Table 6. Case 1. The results of the six weather features.
Index Model
Six Weather Values
Sunny Cloudy Average
MAPE% MAE RMSE MAPE% MAE RMSE MAPE% MAE RMSE
1 ANN 6.544 0.017 0.029 11.763 0.052 0.115 9.154 0.035 0.072
2 LSTM 8.869 0.024 0.046 12.439 0.046 0.088 10.654 0.035 0.067
3 GRU 7.779 0.019 0.032 10.253 0.039 0.078 9.016 0.029 0.055
4.3.2. Case 2: The Results of Five Weather Features (without Coverage Rate)
In this case, there were five weather features of the dataset: wind speed and direction,
UVI, temperature, relative humidity. The predictive results are shown in Figure 6. In the
sunny condition, it can be found that the solar photovoltaic power curve predicted by ANN
significantly deviated from the actual one due to a lack of the cloud coverage rate. The
deviation occurred in the interval of 9 to 10 AM and 2 to 3 PM. Furthermore, the GRU
model also revealed a significant forecast departure at 1 PM. However, the prediction trend
of LSTM was close to the actual PV power curve. Under mostly cloudy conditions, the
forecast trends of the ANN and GRU models deviated significantly from the actual curve,
especially between 1:30 to 2:30 PM. In contrast, the predictive results of the LSTM model
were only slightly different from Case 1.
Energies 2022, 15, x FOR PEER REVIEW 11 of 17
Table 6. Case 1. The results of the six weather features.
Index Model
Six Weather Values
Sunny Cloudy Average
MAPE% MAE RMSE MAPE% MAE RMSE MAPE% MAE RMSE
1 ANN 6.544 0.017 0.029 11.763 0.052 0.115 9.154 0.035 0.072
2 LSTM 8.869 0.024 0.046 12.439 0.046 0.088 10.654 0.035 0.067
3 GRU 7.779 0.019 0.032 10.253 0.039 0.078 9.016 0.029 0.055
4.3.2. Case 2: The Results of Five Weather Features (without Coverage Rate)
In this case, there were five weather features of the dataset: wind speed and direction,
UVI, temperature, relative humidity. The predictive results are shown in Figure 6. In the
sunny condition, it can be found that the solar photovoltaic power curve predicted by
ANN significantly deviated from the actual one due to a lack of the cloud coverage rate.
The deviation occurred in the interval of 9 to 10 AM and 2 to 3 PM. Furthermore, the GRU
model also revealed a significant forecast departure at 1 PM. However, the prediction
trend of LSTM was close to the actual PV power curve. Under mostly cloudy conditions,
the forecast trends of the ANN and GRU models deviated significantly from the actual
curve, especially between 1:30 to 2:30 PM. In contrast, the predictive results of the LSTM
model were only slightly different from Case 1.
(a) (b)
Figure 6. Case 2—(a) Forecasting results of 16 March 2021 (sunny). (b) Forecasting results of 21 May
2021 (mostly cloudy). The frequency of data collection was once per minute.
Table 7 describes the forecasting results of the three models on sunny and mostly
cloudy days. In the sunny day, the ANN still outperformed other models where MAPE
was about 7.85%, MAE was 0.019, and RMSE was 0.032. Under the mostly cloudy condi-
tion, the accuracy of the LSTM model was about 2 to 3% higher than the others.
Table 7. Case 2. The results of five weather values (without coverage rate).
Index Model
Five Weather Value (without Coverage Rate)
Sunny Cloudy Average
MAPE% MAE RMSE MAPE% MAE RMSE MAPE% MAE RMSE
1 ANN 7.857 0.019 0.032 16.251 0.071 0.135 12.054 0.045 0.084
2 LSTM 7.233 0.020 0.035 12.861 0.049 0.101 10.047 0.035 0.068
Figure 6.
Case 2—(
a
) Forecasting results of 16 March 2021 (sunny). (
b
) Forecasting results of
21 May 2021
(mostly cloudy). The frequency of data collection was once per minute.
Table 7describes the forecasting results of the three models on sunny and mostly
cloudy days. In the sunny day, the ANN still outperformed other models where MAPE was
about 7.85%, MAE was 0.019, and RMSE was 0.032. Under the mostly cloudy condition,
the accuracy of the LSTM model was about 2 to 3% higher than the others.
Table 7. Case 2. The results of five weather values (without coverage rate).
Index Model
Five Weather Value (without Coverage Rate)
Sunny Cloudy Average
MAPE% MAE RMSE MAPE% MAE RMSE MAPE% MAE RMSE
1 ANN 7.857 0.019 0.032 16.251 0.071 0.135 12.054 0.045 0.084
2 LSTM 7.233 0.020 0.035 12.861 0.049 0.101 10.047 0.035 0.068
3 GRU 7.628 0.022 0.039 15.953 0.0614 0.122 11.791 0.042 0.081

Energies 2022,15, 4779 12 of 17
4.3.3. Case 3: Results of Five Weather Feature (without UVI)
Five weather features except for UVI were used in this case. The results of the predic-
tion are shown in Figure 7. Compared with the actual curve, the prediction trend of the
three models was roughly the same when the UVI was removed under sunny conditions,
which was quite different from those of removing the coverage rate. Therefore, the coverage
rate was the dominant feature that influences the predictive performance in the sunny
condition. On the other hand, under a mostly cloudy condition, the performance of the
three models was not significantly different from that in Case 2.
Energies 2022, 15, x FOR PEER REVIEW 12 of 17
3 GRU 7.628 0.022 0.039 15.953 0.0614 0.122 11.791 0.042 0.081
4.3.3. Case 3: Results of Five Weather Feature (without UVI)
Five weather features except for UVI were used in this case. The results of the pre-
diction are shown in Figure 7. Compared with the actual curve, the prediction trend of the
three models was roughly the same when the UVI was removed under sunny conditions,
which was quite different from those of removing the coverage rate. Therefore, the cover-
age rate was the dominant feature that influences the predictive performance in the sunny
condition. On the other hand, under a mostly cloudy condition, the performance of the
three models was not significantly different from that in Case 2.
The further evaluation and analysis results are summarized in Table 8. The forecast
evaluation indices were significantly improved when the coverage rate was included, and
this is applicable even without the UVI feature. Based on the average forecasting results
of the different weather conditions, the overall forecasting effect of ANN was better than
the other models, with a MAPE of 7.7%, a MAE of 0.025, and an RMSE of 0.058, respec-
tively.
(a) (b)
Figure 7. Case 3—(a) Forecasting results of 16 March 2021 (sunny). (b) Forecasting results of 21 May
2021 (mostly cloudy). The frequency of data collection was once per minute.
Table 8. Case 3. The results of five weather values (without UVI).
Index Model
Five Weather Values (without UVI)
Sunny Cloudy Average
MAPE% MAE RMSE MAPE% MAE RMSE MAPE%% MAE RMSE
1 ANN 4.001 0.009 0.019 11.492 0.041 0.096 7.747 0.025 0.058
2 LSTM 6.023 0.019 0.033 15.749 0.057 0.114 10.886 0.038 0.074
3 GRU 8.799 0.021 0.036 15.902 0.587 0.112 12.351 0.304 0.074
4.3.4. Case 4: The Results of Only Coverage Rate and Relative Humidity
In this case, only the relative humidity and coverage rate features of the dataset were
used in the prediction. Figure 8 shows the forecast result of the three models subjected to
sunny and mostly cloudy weather. In the sunny condition, either ANN, LSTM, or GRU
all showed good predictive ability, in particular, the prediction curve of ANN almost co-
incided with the trajectory of the real PV power curve. The prediction curves of ANN and
GRU were close to the actual curve between 1:30 PM to 2:10 PM.
Figure 7.
Case 3—(
a
) Forecasting results of 16 March 2021 (sunny). (
b
) Forecasting results of
21 May 2021
(mostly cloudy). The frequency of data collection was once per minute.
The further evaluation and analysis results are summarized in Table 8. The forecast
evaluation indices were significantly improved when the coverage rate was included, and
this is applicable even without the UVI feature. Based on the average forecasting results of
the different weather conditions, the overall forecasting effect of ANN was better than the
other models, with a MAPE of 7.7%, a MAE of 0.025, and an RMSE of 0.058, respectively.
Table 8. Case 3. The results of five weather values (without UVI).
Index Model
Five Weather Values (without UVI)
Sunny Cloudy Average
MAPE% MAE RMSE MAPE% MAE RMSE MAPE%% MAE RMSE
1 ANN 4.001 0.009 0.019 11.492 0.041 0.096 7.747 0.025 0.058
2 LSTM 6.023 0.019 0.033 15.749 0.057 0.114 10.886 0.038 0.074
3 GRU 8.799 0.021 0.036 15.902 0.587 0.112 12.351 0.304 0.074
4.3.4. Case 4: The Results of Only Coverage Rate and Relative Humidity
In this case, only the relative humidity and coverage rate features of the dataset were
used in the prediction. Figure 8shows the forecast result of the three models subjected to
sunny and mostly cloudy weather. In the sunny condition, either ANN, LSTM, or GRU all
showed good predictive ability, in particular, the prediction curve of ANN almost coincided
with the trajectory of the real PV power curve. The prediction curves of ANN and GRU
were close to the actual curve between 1:30 PM to 2:10 PM.

Energies 2022,15, 4779 13 of 17
Energies 2022, 15, x FOR PEER REVIEW 13 of 17
Table 9 summarizes the further analysis results of three models in the sunny and
mostly cloudy conditions. In the sunny condition, the forecast result of ANN still pre-
vailed over the LSTM and GRU model. The MAPE, MAE, and RMSE were 3.93%, 0.008,
and 0.015, respectively.
(a) (b)
Figure 8. Case 4—(a) Forecasting results of 16 March 2021 (sunny). (b) Forecasting results of 21 May
2021 (mostly cloudy). The frequency of data collection was once per minute.
Table 9. Case 4. The results of the only coverage rate and relatively humidity.
Index Model
Only Coverage Rate and Relatively Humidity
Sunny Cloudy Average
MAPE% MAE RMSE MAPE% MAE RMSE MAPE% MAE RMSE
1 ANN 3.933 0.008 0.015 10.342 0.036 0.069 7.138 0.022 0.042
2 LSTM 7.195 0.021 0.035 15.377 0.069 0.129 11.286 0.045 0.082
3 GRU 7.130 0.021 0.036 13.939 0.046 0.087 10.535 0.034 0.062
In summary of the aforementioned four forecasting results, it can be concluded that
the ANN model is more accurate in terms of the short-term PV forecasting in the sunny
condition. However, the advantage of ANN gradually decreased in the mostly cloudy
condition. From another point of view, to balance the predictive ability of the sunny and
mostly cloudy, the average MAPE change in different cases is shown in Figure 9. The
statistical results show that in contrast to other cases, the average change in the MAPE of
case 1 was smaller. Therefore, Case 1 (using six weather features) can adapt better when
subjected to the different weather conditions upon PV prediction.
Furthermore, Case 1 showed that the average degree of change in the MAPE of ANN
was about 6%, which was significantly higher than that of LSTM and GRU, while the av-
erage degree of change in MAPE in the GRU model was about 2%. Thus, when the pre-
diction time is lengthened, more weather conditions will accumulate during photovoltaic
power forecasting, so the prediction effect of the GRU model may have better adaptability
than the LSTM and ANN models. This can be subsequently seen from the analysis results
of one-week forecasting.
Figure 8.
Case 4—(
a
) Forecasting results of
16 March 2021
(sunny). (
b
) Forecasting results of
21 May 2021
(mostly cloudy). The frequency of data collection was once per minute.
Table 9summarizes the further analysis results of three models in the sunny and
mostly cloudy conditions. In the sunny condition, the forecast result of ANN still prevailed
over the LSTM and GRU model. The MAPE, MAE, and RMSE were 3.93%, 0.008, and 0.015,
respectively.
Table 9. Case 4. The results of the only coverage rate and relatively humidity.
Index Model
Only Coverage Rate and Relatively Humidity
Sunny Cloudy Average
MAPE% MAE RMSE MAPE% MAE RMSE MAPE% MAE RMSE
1 ANN 3.933 0.008 0.015 10.342 0.036 0.069 7.138 0.022 0.042
2 LSTM 7.195 0.021 0.035 15.377 0.069 0.129 11.286 0.045 0.082
3 GRU 7.130 0.021 0.036 13.939 0.046 0.087 10.535 0.034 0.062
In summary of the aforementioned four forecasting results, it can be concluded that
the ANN model is more accurate in terms of the short-term PV forecasting in the sunny
condition. However, the advantage of ANN gradually decreased in the mostly cloudy
condition. From another point of view, to balance the predictive ability of the sunny and
mostly cloudy, the average MAPE change in different cases is shown in Figure 9. The
statistical results show that in contrast to other cases, the average change in the MAPE of
case 1 was smaller. Therefore, Case 1 (using six weather features) can adapt better when
subjected to the different weather conditions upon PV prediction.
Furthermore, Case 1 showed that the average degree of change in the MAPE of ANN
was about 6%, which was significantly higher than that of LSTM and GRU, while the aver-
age degree of change in MAPE in the GRU model was about 2%. Thus, when the prediction
time is lengthened, more weather conditions will accumulate during photovoltaic power
forecasting, so the prediction effect of the GRU model may have better adaptability than
the LSTM and ANN models. This can be subsequently seen from the analysis results of
one-week forecasting.

Energies 2022,15, 4779 14 of 17
Energies 2022, 15, x FOR PEER REVIEW 14 of 17
Figure 9. The comprehensive analysis and average degree of the MAPE change with different
weather feature combinations.
4.4. Performance Comparison with Different Weather Features in One Week
The forecast performance evaluation of the three deep learning models with four
cases for one week (from 7 June to 13 June 2021 including sunny, mostly clear, cloudy,
and short showers weather) is shown in Figure 10 and Table 10. The results showed that
using the six weather features yielded better forecasting accuracy. Among them, 7 June
and 8 June were days of short showers. In contrast with the sunny day, adding the cloud
coverage rate to the forecast for the PV power showed that the effect was not significant.
The main reason for the poor prediction results on rainy days is that the raindrop attaches
to the sky imager, which will cause misjudgment when the equation calculates each pixel,
resulting in more errors when calculating the coverage rate. This is a shortcoming of the
RGB method in calculating the coverage rate. As addressed in the previous section, upon
longer prediction period and under more weather conditions, the prediction effect of the
GRU model showed better adaptability than that of the LSTM and ANN models. How-
ever, the corresponding MAPE was about 16.9%, the MAE was 0.048, and the RMSE was
0.098. On the other hand, from the comparison results, the proposed coverage rate as a
weather feature can improve the accuracy by about 2% when compared with the case
without coverage rate, meaning that the coverage rate is an imperative feature for short-
term PV forecasting.
(a)
Figure 9.
The comprehensive analysis and average degree of the MAPE change with different weather
feature combinations.
4.4. Performance Comparison with Different Weather Features in One Week
The forecast performance evaluation of the three deep learning models with four cases
for one week (from 7 June to 13 June 2021 including sunny, mostly clear, cloudy, and short
showers weather) is shown in Figure 10 and Table 10. The results showed that using the six
weather features yielded better forecasting accuracy. Among them, 7 June and 8 June were
days of short showers. In contrast with the sunny day, adding the cloud coverage rate to
the forecast for the PV power showed that the effect was not significant. The main reason
for the poor prediction results on rainy days is that the raindrop attaches to the sky imager,
which will cause misjudgment when the equation calculates each pixel, resulting in more
errors when calculating the coverage rate. This is a shortcoming of the RGB method in
calculating the coverage rate. As addressed in the previous section, upon longer prediction
period and under more weather conditions, the prediction effect of the GRU model showed
better adaptability than that of the LSTM and ANN models. However, the corresponding
MAPE was about 16.9%, the MAE was 0.048, and the RMSE was 0.098. On the other hand,
from the comparison results, the proposed coverage rate as a weather feature can improve
the accuracy by about 2% when compared with the case without coverage rate, meaning
that the coverage rate is an imperative feature for short-term PV forecasting.
Energies 2022, 15, x FOR PEER REVIEW 14 of 17
Figure 9. The comprehensive analysis and average degree of the MAPE change with different
weather feature combinations.
4.4. Performance Comparison with Different Weather Features in One Week
The forecast performance evaluation of the three deep learning models with four
cases for one week (from 7 June to 13 June 2021 including sunny, mostly clear, cloudy,
and short showers weather) is shown in Figure 10 and Table 10. The results showed that
using the six weather features yielded better forecasting accuracy. Among them, 7 June
and 8 June were days of short showers. In contrast with the sunny day, adding the cloud
coverage rate to the forecast for the PV power showed that the effect was not significant.
The main reason for the poor prediction results on rainy days is that the raindrop attaches
to the sky imager, which will cause misjudgment when the equation calculates each pixel,
resulting in more errors when calculating the coverage rate. This is a shortcoming of the
RGB method in calculating the coverage rate. As addressed in the previous section, upon
longer prediction period and under more weather conditions, the prediction effect of the
GRU model showed better adaptability than that of the LSTM and ANN models. How-
ever, the corresponding MAPE was about 16.9%, the MAE was 0.048, and the RMSE was
0.098. On the other hand, from the comparison results, the proposed coverage rate as a
weather feature can improve the accuracy by about 2% when compared with the case
without coverage rate, meaning that the coverage rate is an imperative feature for short-
term PV forecasting.
(a)
Figure 10. Cont.

Energies 2022,15, 4779 15 of 17
Energies 2022, 15, x FOR PEER REVIEW 15 of 17
(b)
(c)
(d)
Figure 10. The forecasting results of the four cases from 7 June to 13 June 2021. The frequency of
data collection was once per minute. (a) One week forecasting results of Case 1 (six weather values).
Figure 10.
The forecasting results of the four cases from 7 June to 13 June 2021. The frequency of
data collection was once per minute. (
a
) One week forecasting results of Case 1 (six weather values).
(b) One
week forecasting results of Case 2 (five weather values (without coverage rate)). (
c
) One
week forecasting results of Case 3 (five weather values (without UVI)). (
d
) One week forecasting
results of Case 4 (only coverage rate and relatively humidity).

Energies 2022,15, 4779 16 of 17
Table 10. The results of the one week PV prediction evaluation.
Index Model
Case1 1Case2 2Case3 3Case4 4
MAPE
(%) MAE RMSE MAPE
(%) MAE RMSE MAPE
(%) MAE RMSE MAPE
(%) MAE RMSE
1 ANN 19.635 0.048 0.095 22.295 0.054 0.106 20.928 0.055 0.102 21.517 0.056 0.114
2 LSTM 16.969 0.048 0.098 18.856 0.054 0.109 24.536 0.058 0.107 20.745 0.061 0.119
3 GRU 16.679 0.049 0.098 18.565 0.529 0.105 17.962 0.057 0.106 18.731 0.056 0.106
1
Case 1: Six weather values;
2
Case 2: Five weather values (without coverage rate);
3
Case 3: Five weather values
(without UVI); 4Case 4: Only coverage rate and relatively humidity.
5. Conclusions
The accuracy of grasping cloud distribution has greatly helped short-term PV power
forecasting. The RGB calculation formula and limitation conditions as sky image analysis
are proposed in this paper. Furthermore, the authors defined the threshold for obtaining
the best coverage rate per minute and as the deep learning model training feature. The
novelty of this method is that the calculation is fast and simple. By comparing the one-week
forecast results of the four cases, it can be seen that considering all six weather features
yielded a complementary effect to improve the forecast accuracy. In particular, adding
the cloud coverage rate as a key feature for the PV power forecast increased up to a 2%
improvement in accuracy. On the other hand, the result shows that the application of
the cloud coverage rate with the deep learning models LSTM and GRU is beneficial to
the short-term PV power forecasting. In this regard, adding the cloud coverage rate as a
feature parameter to forecast the renewable energy power generation of the microgrid is
reliable and robust. In addition, the forecasting result of PV power can also be provided as
a reference for the EMS of the regional microgrid to power dispatch.
However, the forecasting result shows that this method is ineffective on rainy days,
mainly because the raindrop attaches to the sky imager; the RGB pixels cannot be distin-
guished. Therefore, the authors will introduce a hybrid model into a micro-grid energy
management system for further research. If the weather forecast is specifically a rainy day,
a hybrid model will be utilized instead of using the model run by the cloud coverage rate.
In contrast, other weather forecast results will be estimated with the model run by the
cloud coverage rate.
Author Contributions:
This paper is a collaborative work of all authors. Conceptualization, C.-H.C.
and C.-C.W.; methodology, W.-C.K. and S.-Y.C.; software, W.-C.K. and S.-Y.C.; validation, W.-C.K.,
C.-H.C. and S.-Y.C.; formal analysis, W.-C.K.; investigation, W.-C.K. and C.-H.C.; resources, C.-H.C.;
data curation, W.-C.K.; writing—original draft preparation, W.-C.K.; writing—review and editing,
C.-C.W.; supervision, C.-H.C.; project administration, C.-H.C., C.-C.W. and W.-C.K. All authors have
read and agreed to the published version of the manuscript.
Funding:
This research was funded by the Ministry of Science and Technology, Taiwan, R.O.C., grant
number MOST 110-2221-E-035-091 and MOST 110-2221-E-A49-079.
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.
Shah, R.; Mithulananthan, N.; Bansal, R.C.; Ramachandaramurthy, V.K. A review of key power system stability challenges for
large-scale PV integration. Renew. Sustain. Energy Rev. 2015,41, 1423–1436. [CrossRef]
2.
Zhang, Y.; Beaudin, M.; Taheri, R.; Zareipour, H.; Wood, D. Day-ahead power output forecasting for small-scale solar photovoltaic
electricity generators. IEEE Trans. Smart Grid 2015,6, 2253–2262. [CrossRef]

Energies 2022,15, 4779 17 of 17
3.
Verma, T.; Tiwana, A.; Reddy, C.; Arora, V.; Devanand, P. Data Analysis to Generate Models Based on Neural Network and
Regression for Solar Power Generation Forecasting. In Proceedings of the 2016 7th International Conference on Intelligent
Systems, Modelling and Simulation (ISMS), Bangkok, Thailand, 25–27 January 2016; IEEE: New York, NY, USA, 2016; pp. 97–100.
4.
Alluhaidah, B.M.; Shehadeh, S.; El-Hawary, M. Most Influential Variables for Solar Radiation Forecasting Using Artificial Neural
Networks. In Proceedings of the 2014 IEEE Electrical Power and Energy Conference, Toronto, ON, Canada, 10–11 October 2014;
IEEE: New York, NY, USA, 2014; pp. 71–75.
5.
Nitisanon, S.; Hoonchareon, N. Solar Power Forecast with Weather Classification Using Self-Organized Map. In Proceedings of the
2017 IEEE Power & Energy Society General Meeting, Chicago, IL, USA, 16–20 July 2017; IEEE: New York, NY, USA, 2017; pp. 1–5.
6.
Agoua, X.G.; Girard, R.; Kariniotakis, G. Short-term spatio-temporal forecasting of photovoltaic power production. IEEE Trans.
Sustain. Energy 2017,9, 538–546. [CrossRef]
7.
Suksamosorn, S.; Hoonchareon, N.; Songsiri, J. Influential Variable Selection for Improving Solar Forecasts from Numerical
Weather Prediction. In Proceedings of the 2018 15th International Conference on Electrical Engineering/Electronics, Computer,
Telecommunications and Information Technology (ECTI-CON), Chiang Rai, Thailand, 18–21 July 2018; IEEE: New York, NY, USA,
2018; pp. 333–336.
8.
Ghonima, M.; Urquhart, B.; Chow, C.; Shields, J.; Cazorla, A.; Kleissl, J. A method for cloud detection and opacity classification
based on ground based sky imagery. Atmos. Meas. Tech. 2012,5, 2881–2892. [CrossRef]
9.
Shields, J.E.; Karr, M.E.; Burden, A.R.; Johnson, R.W.; Mikuls, V.W.; Streeter, J.R.; Hodgkiss, W.S. Research toward Multi-Site
Characterization of Sky Obscuration by Clouds; Scripps Institution of Oceanography La Jolla Ca Marine Physical Lab: La Jolla, CA,
USA, 2009.
10.
Ghanbarzadeh, A.; Noghrehabadi, A.; Assareh, E.; Behrang, M. Solar Radiation Forecasting Based on Meteorological Data Using
Artificial Neural Networks. In Proceedings of the 2009 7th IEEE International Conference on Industrial Informatics, Cardiff, UK,
23–36 June 2009; IEEE: New York, NY, USA, 2009; pp. 227–231.
11. Pawar, P.; Cortés, C.; Murray, K.; Kleissl, J. Detecting clear sky images. Sol. Energy 2019,183, 50–56. [CrossRef]
12.
Zhao, X.; Wei, H.; Wang, H.; Zhu, T.; Zhang, K. 3D-CNN-based feature extraction of ground-based cloud images for direct normal
irradiance prediction. Sol. Energy 2019,181, 510–518. [CrossRef]
13.
Mammoli, A.; Terren-Serrano, G.; Menicucci, A.; Caudell, T.P.; Martínez-Ramón, M. An experimental method to merge far-field
images from multiple longwave infrared sensors for short-term solar forecasting. Sol. Energy 2019,187, 254–260. [CrossRef]
14.
Kamadinata, J.O.; Ken, T.L.; Suwa, T. Sky image-based solar irradiance prediction methodologies using artificial neural networks.
Renew. Energy 2019,134, 837–845. [CrossRef]
15.
Quesada-Ruiz, S.; Chu, Y.; Tovar-Pescador, J.; Pedro, H.; Coimbra, C. Cloud-tracking methodology for intra-hour DNI forecasting.
Sol. Energy 2014,102, 267–275. [CrossRef]
16. Zhang, J.; Verschae, R.; Nobuhara, S.; Lalonde, J.-F. Deep photovoltaic nowcasting. Sol. Energy 2018,176, 267–276. [CrossRef]
17.
Kong, W.; Jia, Y.; Dong, Z.Y.; Meng, K.; Chai, S. Hybrid approaches based on deep whole-sky-image learning to photovoltaic
generation forecasting. Appl. Energy 2020,280, 115875. [CrossRef]
18.
Wang, F.; Zhang, Z.; Chai, H.; Yu, Y.; Lu, X.; Wang, T.; Lin, Y. Deep Learning Based Irradiance Mapping Model for Solar PV Power
Forecasting Using Sky Image. In Proceedings of the 2019 IEEE Industry Applications Society Annual Meeting, Baltimore, MD,
USA, 29 September–3 October 2019; IEEE: New York, NY, USA, 2019; pp. 1–9.
19.
Zhen, Z.; Liu, J.; Zhang, Z.; Wang, F.; Chai, H.; Yu, Y.; Lu, X.; Wang, T.; Lin, Y. Deep learning based surface irradiance mapping
model for solar PV power forecasting using sky image. IEEE Trans. Ind. Appl. 2020,56, 3385–3396. [CrossRef]
20.
Wen, H.; Du, Y.; Chen, X.; Lim, E.; Wen, H.; Jiang, L.; Xiang, W. Deep learning based multistep solar forecasting for PV ramp-rate
control using sky images. IEEE Trans. Ind. Inform. 2020,17, 1397–1406. [CrossRef]
21.
Srivastava, N.; Hinton, G.; Krizhevsky, A.; Sutskever, I.; Salakhutdinov, R. Dropout: A simple way to prevent neural networks
from overfitting. J. Mach. Learn. Res. 2014,15, 1929–1958.
22.
Greff, K.; Srivastava, R.K.; Koutník, J.; Steunebrink, B.R.; Schmidhuber, J. LSTM: A search space odyssey. IEEE Trans. Neural Netw.
Learn. Syst. 2016,28, 2222–2232. [CrossRef] [PubMed]
23.
Cho, K.; Van Merriënboer, B.; Gulcehre, C.; Bahdanau, D.; Bougares, F.; Schwenk, H.; Bengio, Y. Learning phrase representations
using RNN encoder-decoder for statistical machine translation. arXiv 2014, arXiv:1406.1078.
24. Chen, Z.; Yang, Y. Assessing forecast accuracy measures. Prepr. Ser. 2004,2010, 2004–2010.