3282 IEEE TRANSACTIONS ON INDUSTRY APPLICATIONS, VOL. 57, NO.

4, JULY/AUGUST 2021

Frequency-Domain Decomposition and Deep

Learning Based Solar PV Power Ultra-Short-Term
Forecasting Model
Jichuan Yan, Student Member, IEEE, Lin Hu, Zhao Zhen, Member, IEEE, Fei Wang , Senior Member, IEEE,
Gang Qiu, Yu Li, Liangzhong Yao, Miadreza Shafie-khah , Senior Member, IEEE,
and João P. S. Catalão , Senior Member, IEEE

Abstract—Ultra-short-term photovoltaic (PV) power forecasting through frequency-domain analysis. Then, the PV power is de-
can support the real-time dispatching of the power grid. However, composed into the low-frequency and high-frequency components,
PV power has great fluctuations due to various meteorological which supports the rationality of decomposition results and solves
factors, which increase energy prices and cause difficulties in the problem that the current decomposition model only uses the
managing the grid. This article proposes an ultra-short-term PV direct decomposition method and the decomposition components
power forecasting model based on the optimal frequency-domain are not physical. Then, a convolutional neural network (CNN) is
decomposition and deep learning. First, the optimal frequency used to forecast the low-frequency and high-frequency components,
demarcation points for decomposition components are obtained and the final forecasting result is obtained by addition reconstruc-
tion. Based on the actual PV data in heavy rain days, the mean
absolute percentage error (MAPE) of the proposed forecasting
Manuscript received July 29, 2020; revised December 24, 2020 and February model is decreased by 52.97%, 64.07%, and 31.21%, compared
18, 2021; accepted April 7, 2021. Date of publication April 15, 2021; date of with discrete wavelet transform, variational mode decomposition,
current version July 16, 2021. Paper 2020-ESC-1025.R2, presented at the 2020 and direct prediction models. In addition, compared with recur-
Industry Applications Society Annual Meeting, Detroit, MI, USA, Oct. 11–15, rent neural network and long–short-term memory model, the
and approved for publication in the IEEE TRANSACTIONS ON INDUSTRY MAPE of the CNN forecasting model is decreased by 23.64% and
APPLICATIONS by the Energy Systems Committee of the IEEE Industry 46.22%, and the training efficiency of the CNN forecasting model is
Applications Society. This work was supported in part by the National Key improved by 85.63% and 87.68%. The results fully show that
R&D Program of China Technology and application of wind power/photovoltaic the proposed model in this article can improve both forecasting
power prediction for promoting renewable energy consumption under Grant
accuracy and time efficiency significantly.
2018YFB0904200 and in part by the Eponymous Complement S&T Program
of State Grid Corporation of China under Grant SGLNDKOOKJJS1800266. Index Terms—Decomposition, deep learning (DL), frequency
Paper 2020-ESC-1025.R2, presented at the 2020 Industry Applications Society
domain, photovoltaic (PV) power forecasting, ultra-short term.
Annual Meeting, Detroit, MI, USA, Oct. 11–15, and approved for publication in
the IEEE TRANSACTIONS ON INDUSTRY APPLICATIONS by the Energy Systems
Committee of the IEEE Industry Applications Society. (Corresponding author:
Fei Wang.) I. INTRODUCTION
Jichuan Yan and Lin Hu are with the Department of Electrical Engineer-
ing, North China Electric Power University, Baoding 071003, China (e-mail: A. Background and Motivation
[email protected]; [email protected]).
ACING the shortage of fossil energy and the deterioration
Zhao Zhen is with the State Key Lab of Power System, Department of
Electrical Engineering, Tsinghua University, Beijing 100084, China and also
with the Department of Electrical Engineering, North China Electric Power
University, Baoding 071003, China (e-mail: [email protected]).
F of climate, such as the greenhouse effect, ozone hole, and
melting of polar glaciers, sustainable development of energy
Fei Wang is with the Department of Electrical Engineering, North China Elec- and the environment has attracted worldwide attention. Due to
tric Power University, Baoding 071003, China, with the State Key Laboratory of the advantages of sustainability, clean and pollution-free, high
Alternate Electrical Power System, with Renewable Energy Source, Nothe China
Electric Power University, Beijing 102206, China, and also with the Hebei Key flexibility, etc., photovoltaic (PV) power generation technology
Laboratory of Distributed Energy Storage and Microgrid, North China Electric and related industries have experienced tremendous growth over
Power University, Baoding 071003, China (e-mail: [email protected]). the past few years [1]–[3]. PV power affected by various me-
Gang Qiu and Yu Li are with the Dispatch and Control Center, State
Grid Xinjiang Electric Power Co., Ltd., Urumqi 830018, China (e-mail: teorological factors is highly uncertain. These fluctuations will
[email protected]; [email protected]). cause energy prices, grid management difficulties, etc., which
Liangzhong Yao is with the School of Electrical Engineering will affect the stable operation of the power grid [4], [5]. PV
and Automation, Wuhan University, Wuhan 430072, China (e-mail:
[email protected]). power generation system is typically connected to the power grid
Miadreza Shafie-khah is with the School of Technology and Innovations, to compensate for losses in the conventional power generation
University of Vaasa, 65200 Vaasa, Finland (e-mail: [email protected]). systems. In a grid, some power generation systems that output
João P. S. Catalão is with the Faculty of Engineering of the University
of Porto and also with the Institute for Systems and Computer Engineering, constant power are called the conventional power plants. Others
Technology and Science (INESC TEC), 4200-465 Porto, Portugal (e-mail: convert their power in response to the changes of PV generation
[email protected]). and demand, which balances total power consumption with total
Color versions of one or more figures in this article are available at https:
//doi.org/10.1109/TIA.2021.3073652. power generation. These power plants that convert output power
Digital Object Identifier 10.1109/TIA.2021.3073652 are called load-tracking power plants. When the output power of
0093-9994 © 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: Universiti Malaysia Pahang. Downloaded on May 12,2023 at 07:16:53 UTC from IEEE Xplore. Restrictions apply.
YAN et al.: FDD AND DEEP LEARNING BASED SOLAR PV POWER ULTRA-SHORT-TERM FORECASTING MODEL 3283

TABLE I
DEVELOPMENT OF PV POWER PREDICTION TECHNOLOGY

the PV generation system is markedly increased or decreased, relying on the detailed power plant geographic information
the load-tracking power plants must respond promptly. PV and accurate meteorological data. The physical formula has
power forecast for 15 min in advance enables the load-tracking certain errors, and its model has poor anti-interference ability
power plant to react to sudden output changes timely. In order and weak robustness. Common statistical methods [16] include
to stabilize the operation of the power grid, ultra-short-term PV the time-series method, regression analysis method, gray theory,
power prediction is particularly significant [6]–[8]. fuzzy theory, and spatiotemporal correlation method. Compared
with physical methods, statistical methods are simpler, which
B. Literature Review
do not need to consider PV system parameters and complex
Due to various meteorological factors, the PV power has great photoelectric conversion models. However, statistical methods
fluctuations [9]–[11]. It is difficult to achieve satisfactory results often need to collect and process a large amount of historical
with the traditional prediction methods. In recent years, the data, which increases the difficulty of data acquisition and
method of first decomposing and then predicting PV power data processing. Common intelligent methods [17]–[19] include ar-
has become a research hotspot. Among them, the frequency- tificial neural network, support vector machine, Markov chain,
domain decomposition (FDD) method excavates and extracts and particle swarm algorithm. The intelligent method does not
PV power characteristics from the perspective of the frequency need to obtain a specific expression between input and output
domain, which has become a current research hotspot. Zhang et and obtains a prediction model of PV power generation output
al. [12] use the variational mode decomposition (VMD) to de- through training on the historical data, which is simple to operate
compose different frequency components from the historical PV and easy to implement, but has the defects of overlearning and
power time series. Kushwaha and Pindoriya [13] decomposed easy to fall into local optimal solutions. Compared with the
the time series of solar PV power generation by wavelet decom- above-mentioned traditional intelligent algorithms, deep learn-
position. Zhang et al. [14] proposed a new prediction model ing (DL) [20]–[22] algorithms are more promising. There are
based on the Hilbert–Huang transform and integrating improved various branches in the DL system, including recurrent neural
empirical mode decomposition (IEMD) with feature selection network (RNN), long- and short-term memory network (LSTM),
and prediction engine. IEMD is used to decompose the data. and convolutional neural network (CNN). Among them, CNN
In the above literature, the data are only decomposed using the is not only successfully applied to image processing but also
existing models, but few pieces of literature can support the ratio- used to solve one-dimensional (1-D) data classification and re-
nality of the decomposition results through an effective method. gression problems. CNN reduces the parameters to be estimated
Table I lists the various methods of solar PV power gener- due to the weight-sharing technology, thereby shortening the
ation forecasting, which can be summarized into the following training time for a prediction. In addition, CNN can perform
categories: physical methods, statistical methods, and intelligent feature extraction on the data, which has good robustness,
algorithms [15]. Among the physical methods [5], the three and there are few studies on the application of CNN to the
basic methods are numerical weather forecast model, total sky ultra-short-term prediction of PV power, so this article is ded-
image, and satellite image model based on the cloud movement, icated to the reasonable application of the CNN algorithm to
which can help estimate the PV power. However, physical meth- improve the accuracy of the ultra-short-term prediction of PV
ods require detailed historical data to train prediction models, power.

Authorized licensed use limited to: Universiti Malaysia Pahang. Downloaded on May 12,2023 at 07:16:53 UTC from IEEE Xplore. Restrictions apply.
3284 IEEE TRANSACTIONS ON INDUSTRY APPLICATIONS, VOL. 57, NO. 4, JULY/AUGUST 2021

C. Contribution
To overcome the shortcomings that there is currently no
effective method to support the rationality of the FDD results,
this article proposes an ultra-short-term prediction model of PV
power based on optimal FDD and DL. First, the amplitude and
phase of each frequency sine wave are obtained by fast Fourier
decomposition. As the frequency demarcation point is different,
the correlation between the decomposition component and the
original data is analyzed. By analyzing the prediction results
of the decomposition components at different frequency de-
marcation points, the optimal frequency demarcation points for
decomposing low-frequency components and high-frequency
components are obtained. Then, CNN is used to predict the
low-frequency component and high-frequency component, and Fig. 1. Relationship between the time-domain signals and sine wave signals
the final prediction result is obtained by addition reconstruction. of different frequencies.
The main contributions of this article include the following.
1) By minimizing the square of the difference that the
correlation between the low-frequency components and II. METHODOLOGY
raw data is subtracted from the correlation between the
high-frequency components and raw data, the optimal A. Fast Fourier Transform
frequency demarcation points for decomposition com- Discrete Fourier Transform (DFT) can discretize the fre-
ponents are obtained, and the PV power is decomposed quency domain of a finite-length sequence, but its computational
into the low-frequency components and high-frequency complexity is too large to process the problem in real time,
components. The low-frequency component represents thus leading to the FFT [23]–[25]. FFT uses periodicity and
the regular part of PV power generation, which shows PV symmetry to improve the DFT algorithm, which greatly reduces
power’s trend characteristics, while the high-frequency the amount of computation. The sine wave decomposed by FFT
component represents the randomness of PV power gener- is the most single frequency signal. Any complex signal can
ation, which shows PV power’s fluctuation characteristics be seen as the composite of many sine waves with various
affected by other factors, such as weather. This method frequencies and amplitudes. It can be considered that sine wave
supports the rationality of the decomposition results and is the basis of all waveforms. Therefore, this article uses an FFT
solves the problem that the current decomposition model to decompose PV power in the frequency domain. Fig. 1 shows
only uses the direct decomposition method, and the de- the relationship between the time-domain signal and the sine
composition components are not physical. wave signals of different frequencies. The black curve represents
2) CNN is not only successfully applied to image process- the time-domain signal of a PV power curve, and the color curves
ing but also used to solve 1-D data classification and represent the sine wave signals of different frequencies that make
regression problems. CNN reduces the parameters to be up this PV power curve.
estimated due to the weight-sharing technology, thereby According to (1), the discretized sequence is decomposed
shortening the training time for a prediction. CNN can into signals of multiple frequencies. and the sine wave and
perform feature extraction with good robustness. There- the cosine wave of the same frequency are superimposed by
fore, this article is dedicated to the reasonable application different coefficients to generate the cosine waves of various
of CNN to improve the accuracy and efficiency of ultra- phases of the same frequency. The modulus cn of a complex
short-term prediction of PV power. number in the frequency domain obtained by FFT represents
3) The method proposed in this article is verified by the data the energy of the cosine wave corresponding frequency, and
of PV power station in Ningxia and Jilin in China on sunny the angle of the complex number represents the phase θn of
days, cloudy days, light rainy days, and heavy rainy days. the cosine wave, thereby obtaining the amplitude spectrum and
The results show that the proposed model improves the the phase spectrum. The actual amplitude ak of the kth point of
prediction accuracy. the cosine is defined by (2). The physical frequency fk of the kth
The rest of this article is organized as follows. Section II point is calculated as follows:
introduces the theory of fast Fourier transform (FFT) and DL.
Section III introduces how to choose the optimal frequency

N
 
demarcation point of the FDD. In Section IV, the details of the f (n) = a0 + an cos (2π fn ) + bn cos (2π fn )
experimental simulation are introduced and the relevant analysis n=1
results are discussed. Finally, Section V presents the conclusions
in this article. 
N
= c0 + cn cos (2π fn + θn ) (1)
n=1

Authorized licensed use limited to: Universiti Malaysia Pahang. Downloaded on May 12,2023 at 07:16:53 UTC from IEEE Xplore. Restrictions apply.
YAN et al.: FDD AND DEEP LEARNING BASED SOLAR PV POWER ULTRA-SHORT-TERM FORECASTING MODEL 3285

Fig. 2. Structure of CNN.

2ck
ak = (2)
N Fig. 3. Illustration of 1-D convolution and 1-D pooling.

fs
fk = k × . (3)
N
of the input information, which is weights sharing. F (•) is the
In the formula, n is a positive integer. a0 is a constant compo- activation function applied to the convolution layer, which can
nent, an is the cosine component amplitude, bn is the sinusoidal effectively improve the fitting ability of the model.
component amplitude, c0 is the constant component, cn is the 2) Pooling Layer: The role of the pooling layer is to reduce
amplitude, fn = 2πnt is the frequency, θn is the phase, N is a the size of the feature map generated by the convolution layer
N
positive integer, and t is the sampling interval. and to effectively extract the feature information in the feature
map. The formula for the pooling layer is shown as follows:
B. Pearson Correlation Coefficient
Pi,l j,k = pool(ym,n,k
l
) (6)
Pearson’s correlation coefficient (R) represents the curve fit
of two sequences. Specifically, the critical criterion for R is a
where (m, n) ∈ Ri, j , and Ri, j is the information at location (i, j).
very strong correlation between 0.8 and 1.0, a strong correlation
3) Fully Connected Layer: The function of the fully con-
between 0.6 and 0.8, a medium correlation between 0.4 and
nected layer is to summarize the distributed feature represen-
0.6, a weak correlation between 0.2 and 0.4, and a very weak
tations learned by the previous layer into the same space for
correlation or no correlation between 0 and 0.2 [26]. R is defined
subsequent applications. All the neurons in the previous layer
as follows:
   are connected to each neuron in the current layer.
N XY − X Y 4) Description of 1-D CNN: 1-D CNN is a branch of CNN
RX,Y =    2 
  2 . (4)
N X2 − X N Y2 − Y [28]. The convolution kernel window of CNN1D slides in a
single direction (i.e., time step). Because PV power is the time-
series data, this article chooses CNN1D to build a prediction
C. Convolutional Neural Network
model. In the 1-D convolution layer, the size of the kernel is 3
CNN is a common DL model with powerful feature extraction and the corresponding stride is 1. In the 1-D pooling layer, the
capabilities. CNN generally consists of five types of cell layers, size of the kernel is 2 and the corresponding stride is 1. The
namely the input layer, the convolutional layer, the pooling layer, specific flowchart is shown in Fig. 3.
the fully connected layer, and the output layer [27]. The structure
of CNN is shown in Fig. 2.
1) Convolution Layer: The role of the convolutional layer is D. Combined Model Framework
to extract features from the input information. The convolutional The framework of the ultra-short-term prediction model of PV
layer is typically composed of multiple convolution kernels, each power based on optimal FDD and DL proposed in this article is
of which is used to compute a feature map. Each cell of the shown in Fig. 4. First, the amplitude and phase of each frequency
feature map is connected to the region of the adjacent cell in sine wave are obtained by fast Fourier decomposition. As the
the previous layer. Convolution of the input by the convolution frequency demarcation point is different, the correlation between
kernel and nonlinear processing of convolution results by acti- the decomposition component and the original data is analyzed.
vation function can acquire the new feature map. The formula By minimizing the square of the difference that the correlation
for the convolutional layer is given as follows: between the low-frequency components and raw data is sub-
yi,l j,k = F((wkl )T xi,l j + blk ) (5) tracted from the correlation between the high-frequency com-
ponents and raw data, the optimal frequency demarcation points
where wkl and blk are the weights and bias of the kth convolution for decomposition components are obtained. Then, CNN is used
kernel in the lth convolutional layer, respectively. xi,l j is the input to predict the low-frequency component and high-frequency
information of (i, j) region in the lth convolution layer. The component, and the final prediction result is obtained by addition
weight wkl in the lth convolution layer is shared by each region reconstruction.

Authorized licensed use limited to: Universiti Malaysia Pahang. Downloaded on May 12,2023 at 07:16:53 UTC from IEEE Xplore. Restrictions apply.
3286 IEEE TRANSACTIONS ON INDUSTRY APPLICATIONS, VOL. 57, NO. 4, JULY/AUGUST 2021

Fig. 6. Phase spectrum of PV power.

Fig. 4. Ultra-short-term prediction model of PV power based on optimal FDD

and DL.
Fig. 7. Correlation between the decomposition components and raw data
before weighting.

TABLE II
COMPARISON OF R OF LOW-FREQUENCY COMPONENTS AT DIFFERENT
FREQUENCIES

Fig. 5. Amplitude spectrum of PV power. power, which indicates its fluctuation characteristics affected by
other factors, such as weather. How to accurately decompose the
low-frequency components and high-frequency components has
always been a difficult problem. In order to solve this problem,
III. SPECTRUM ANALYSIS
this article chooses the optimal frequency demarcation point
A. Frequency-Domain Decomposition from the perspective of the correlation between the decomposed
This article selects the PV data of Ningxia PV Power Station data and raw data. Fig. 7 shows the correlation between the
in China from 0:00 on January 1, 2017, to 23:45 on November decomposed components and raw data. When FDD is performed
30, 2017, to train the model, and perform FDD through FFT to on PV power data, the more the frequencies selected, the stronger
obtain the amplitude spectrum and phase spectrum of PV power, the correlation between the low-frequency components and raw
which are shown in Figs. 5 and 6. data, and the weaker the correlation between the high-frequency
components and raw data.
To illustrate the effect of the correlation between the decom-
B. Selection of Optimal Frequency Demarcation Point position component and raw data on the prediction results, this
Through FFT spectrum analysis, PV power can be decom- article uses the CNN model to predict the low-frequency com-
posed into low-frequency components and high-frequency com- ponents and high-frequency components from 0:00 on January
ponents. The low-frequency component represents the regular 1, 2017, to 23:45 on November 30, 2017. Table II compares the
part of PV power, which indicates its trend characteristics, while prediction results of low-frequency and high-frequency com-
the high-frequency component represents the randomness of PV ponents at different frequencies. The frequency demarcation

Authorized licensed use limited to: Universiti Malaysia Pahang. Downloaded on May 12,2023 at 07:16:53 UTC from IEEE Xplore. Restrictions apply.
YAN et al.: FDD AND DEEP LEARNING BASED SOLAR PV POWER ULTRA-SHORT-TERM FORECASTING MODEL 3287

point, as selected in Table II, is the frequency node with rel-

atively large amplitudes in the amplitude spectrum of Fig. 5.
As the correlation between the low-frequency component and
raw data increases, the curve fitting effect becomes better. As
the correlation between the high-frequency component and raw
data decreases, the curve fitting effect becomes worse. The
relationship between the selected frequencies and the correlation
of decomposition component and raw data shows that as the
frequency number increases, the better the prediction result of
the low-frequency component and the worse the prediction result
of the high-frequency component, so the two are contradictory.
Low-frequency components represent the regular part of PV Fig. 8. Graph of the squared difference before weighting.
power, which can be accurately predicted, while high-frequency
components are relatively random, which are difficult to pre-
dict. If the proportion of low-frequency components in raw
Fig. 8 is a graph of the obtained square of the difference
data can be increased, the accurate prediction of low-frequency
that the correlation between the low-frequency components
components to balance the prediction errors of high-frequency
and raw data is subtracted from the correlation between the
components will effectively improve the overall prediction ac-
high-frequency components and raw data. It can be seen from
curacy. Therefore, the idea of optimal frequency demarcation
Fig. 8 that when the frequency demarcation point f1 is 365, the
point selection proposed in this article is that the low-frequency
correlation between the low-frequency components and raw data
component considers to select the highest frequency demarca-
and the correlation between the high-frequency components and
tion point as possible and accounts for a large proportion, and
raw data reach a balance. However, due to the strong regularity
then the second level is the high-frequency component, which
and high prediction accuracy of low-frequency components,
hopes that the frequency selection is not too high. Otherwise,
this article first hopes that the proportion of low-frequency
prediction is too difficult to get a certain level of balance.
components is high. When the frequency value continues to
From Fig. 7, it can be seen that when the frequency reaches
increase, the correlation between the low-frequency components
a certain value, as the frequency band increases, the correlation
and raw data can be improved. However, when the correlation
between the low-frequency components and raw data does not
between the low-frequency components and raw data is basically
increase significantly. It can be considered that there is no signif-
unchanged, the prediction accuracy of low-frequency compo-
icant increase in the correlation between the low-frequency com-
nents basically reaches the maximum value. At this time, the
ponents and raw data, while the correlation between the high-
accuracy of the final prediction result mainly depends on the
frequency component and raw data continues to decline. This
high-frequency components. Therefore, in order to give priority
optimal frequency demarcation point that balances both low-
to low-frequency components, the frequency value is continu-
frequency components and high-frequency components not only
ously increased. This article adds weight values to the objective
ensures sufficient extraction of low-frequency components but
function in (7). By multiplying the correlation between the
also does not cause the difficulty of predicting high-frequency
low-frequency components and raw data by a low weight value
components to an unacceptable level due to the excessive extrac-
and multiplying the correlation between the high-frequency
tion of low-frequency components. In this article, the optimal
components and raw data by a high weight value. In this way,
frequency is solved by minimizing the square of the difference
the design idea of giving priority to low-frequency components
that the correlation between the low-frequency components and
is achieved. The new objective function is shown in (8). This
raw data is subtracted from the correlation between the high-
article chooses α = 0.173, β = 0.827, and obtains the frequency
frequency components and raw data, which can be regarded
demarcation point f2 = 1825
as an optimization problem. The specific formula is given as
follows:  2
min F ( f ) = αL ( f ) − βH ( f )
 2
min F ( f ) = L( f ) − H ( f ) N
s.t. f ∈ 0, (8)
N 2
s.t. f ∈ 0, . (7)
2
where α is the weight of the correlation between the low-
Among them, F ( f ) is an objective function. When the ob- frequency component and the original data, and β is the weight
jective function reaches the minimum value, the correlation of the correlation between the high-frequency component and
between the low-frequency components and raw data and the the original data.
correlation between the high-frequency components and raw Fig. 9 shows the correlation between the decomposition com-
data reach an optimal balance. L( f ) is the correlation between ponents and raw data after weighting. Fig. 10 shows the square
the low-frequency components and raw data, H ( f ) is the cor- of the difference that the correlation between the low-frequency
relation between the high-frequency components and raw data, components and raw data is subtracted from the correlation
and N is the total number of sampling frequencies. between the high-frequency components and raw data after

Authorized licensed use limited to: Universiti Malaysia Pahang. Downloaded on May 12,2023 at 07:16:53 UTC from IEEE Xplore. Restrictions apply.
3288 IEEE TRANSACTIONS ON INDUSTRY APPLICATIONS, VOL. 57, NO. 4, JULY/AUGUST 2021

the testing set. The time scale of input data is 24 h, and the time
scale of output data is 1 h.
We use Python 3.6.1 with Tensorflow and Scikit learn to
perform all the simulations.

B. Data Processing
During model training, to eliminate the difference in magni-
tude between each dimension data, the samples of input data will
be normalized. The range of values of all samples is converted
to [0, 1], which avoids large differences in sample magnitude,
leading to large network prediction errors. Its formula is shown
Fig. 9. Correlation between the decomposition components and raw data after
weighting. as follows:
M − Mmin
M∗ = . (9)
Mmax − Mmin

C. Performance Criterion
In order to evaluate the performance of the prediction model,
we employ two effective error indices that are the mean absolute
percentage error (MAPE) and root-mean-squared error (RMSE).
Under the same set of training data, the smaller the MAPE and
RMSE, together with the higher R, the better the prediction
model effect. The mathematical formulas for the two error
indices are shown as follows:
1  yt − ŷt
N
Fig. 10. Graph of the squared difference after weighting. MAPE = (10)
N t=1 yt

N
(yt − ŷt )2
RMSE = t=1
. (11)
N

D. Calculation of Forecast Uncertainty

Uncertainty is a major problem in all PV power prediction
models. Various uncertain factors affecting the grid operation
Fig. 11. Optimal frequency decomposition result.
can be divided into three factors as follows:
1) due to input, that is measurement error related to weather
variables;
2) due to the inherent randomness of the basic physical
weighting. It can be seen that after f2 , as the frequency band in- process;
creases, the correlation between the low-frequency components 3) inherent to the model structure, namely the selected pa-
and raw data does not increase significantly, while the corre- rameters and the training dataset used for the purpose of
lation between the high-frequency components still maintains a model building.
downward trend in Fig. 9. Therefore, f2 is the optimal frequency This article refers to the pieces of literature [29], [30] and
demarcation point fo. Fig. 11 shows the decomposition results adopts the method of quantifying solar PV power generation
at the frequency demarcation point fo. forecast in the form of prediction interval (PI).
The PI with a predefined confidence level α% is the interval
IV. CASE STUDY between the upper and lower limits of the predicted power gener-
upper
ation at time t, t ∈ [1, 24] of day d [P̂test lower
j
(t, d ), P̂testj (t, d )] so
A. Simulation Data and Simulation Platform that the actual “unknown” value Ptestj (t, d ), of the jth test pattern
The data used for the simulation in this section are historical at time t of day d, falls within the interval, and its probability is
weather data and historical power with 15 min time resolution equal to α %.
from 0:00 on January 1, 2017, to 23:45 on December 31, 2017, in 
PIα = P̂test
upper
Ningxia PV Power Station and Jilin PV Power Station in China.
lower
j
(t, d ) , P̂testj (t, d )
This article uses the data from 0:00 on January 1, 2017, to 23:45  
upper
on November 30, 2017, as the training set, and the data from
lower
Prob P̂test j
(t, d ) ≤ P testj (t, d ) ≤ P̂testj (t, d ) = α%.
0:00 on December 1, 2017, to 23:45 on December 31, 2017, as (12)

Authorized licensed use limited to: Universiti Malaysia Pahang. Downloaded on May 12,2023 at 07:16:53 UTC from IEEE Xplore. Restrictions apply.
YAN et al.: FDD AND DEEP LEARNING BASED SOLAR PV POWER ULTRA-SHORT-TERM FORECASTING MODEL 3289

TABLE III
PIS’ PERFORMANCES OBTAINED BY FDD PREDICTION MODEL ON THE TEST
DATASET FOR α= 80% TARGET CONFIDENCE LEVEL

Referring to the pieces of literature [29], [30], this article uses

two performance indicators to evaluate the forecast uncertainty,
namely PI coverage probability (PICP) and PI width (PIW). The
goal of a PI with a confidence level of α% is to cover at least
α% with as small a width as possible. This section uses the
data from 0:00 on January 1, 2017, to 23:45 on November 30,
2017, as the training set, and the data from 0:00 on December Fig. 12. Actual and forecasted PV power on four different weather types using
1, 2017, to 23:45 on December 31, 2017, as the testing set in the dataset of Ningxia Power Station. (a) Sunny. (b) Cloudy. (c) Light rainy.
(d) Heavy rainy.
Ningxia PV Power Station and Jilin PV Power Station in China.
The time scale of the input data is 24 h, and the time scale of
the output data is 1 h. Table III reports the PIs’ performances
obtained by the proposed approach in terms of PICP and PIW.
From the results in Table III, it can be seen that the coverage of
PI is higher than the target coverage level of 0.8, and the width
of PI is also smaller. To a certain extent, it can be seen that the
fast fourier decomposition (FFD) prediction model proposed in
this article can reduce the forecast uncertainty.

E. Performance Analysis of FDD Forecasting Model on

Different Weather Types
This article establishes the FDD forecasting model of PV
power on a sunny day, cloudy day, light rainy day, and heavy
rainy day. In this part, two different datasets, Ningxia Power
Station and Jilin Power Station are used to compare the per- Fig. 13. Error distribution curves of different decomposition prediction mod-
formance of the FDD forecasting model of different weather els on four different weather types using the dataset of Ningxia Power Station.
types. In this section, the FDD forecasting model with predicting (a) Sunny. (b) Cloudy. (c) Light rainy. (d) Heavy rainy.
1 h time scale is compared with other VMD, discrete wavelet
transformation (DWT), and direct models, and the prediction
model is unified using the CNN model.
1) Comparison Analysis of Sunny Days: Fig. 12(a) shows
the actual and forecasted PV power on a sunny day using the
dataset of Ningxia Power Station. Fig. 13(a) shows the error
distribution curves of the direct prediction model and frequency
decomposition prediction model on a sunny day using the dataset
of Ningxia Power Station. Fig. 14(a) shows linear regression
graph of the direct prediction model and frequency decomposi-
tion prediction model on a sunny day using the dataset of Ningxia
Power Station. Fig. 15(a) shows the generation of forecast error
graph of different sunny days’ decomposition prediction models
using the dataset of Ningxia Power Station.
Tables IV and V present the performance comparison of the
sunny weather prediction models for different datasets. Taking
Ningxia Power Station as an example, in sunny days, compared
Fig. 14. Linear regression graph of different decomposition prediction models
with DWT, VMD, and direct prediction models, the MAPE ac- on four different weather types using the dataset of Ningxia Power Station.
curacy of the FDD model proposed in this article is increased by (a) Sunny. (b) Cloudy. (c) Light rainy. (d) Heavy rainy.
42.05%, 32.94%, and 44.59%, respectively, the RMSE accuracy

Authorized licensed use limited to: Universiti Malaysia Pahang. Downloaded on May 12,2023 at 07:16:53 UTC from IEEE Xplore. Restrictions apply.
3290 IEEE TRANSACTIONS ON INDUSTRY APPLICATIONS, VOL. 57, NO. 4, JULY/AUGUST 2021

TABLE VI
PERFORMANCE COMPARISON OF DIFFERENT CLOUDY DAYS’ DECOMPOSITION
PREDICTION MODELS USING THE DATASET OF NINGXIA POWER STATION

TABLE VII
PERFORMANCE COMPARISON OF DIFFERENT CLOUDY DAYS’ DECOMPOSITION
PREDICTION MODELS USING THE DATASET OF JILIN POWER STATION

Fig. 15. Generation of forecast error graph of different decomposition predic-

tion models on four different weather types using the dataset of Ningxia Power
Station. (a) Sunny. (b) Cloudy. (c) Light rainy. (d) Heavy rainy.

TABLE IV
PERFORMANCE COMPARISON OF DIFFERENT SUNNY DAYS’ DECOMPOSITION
PREDICTION MODELS USING THE DATASET OF NINGXIA POWER STATION

using the dataset of Ningxia Power Station. Fig. 14(b) shows

the linear regression graph of the direct prediction model and
frequency decomposition prediction model on a cloudy day
using the dataset of Ningxia Power Station. Fig. 15(b) shows
the generation of forecast error graph of different cloudy days’
decomposition prediction models using the dataset of Ningxia
Power Station. Tables Ⅵ and Ⅶ present the performance com-
TABLE V parison of the cloudy weather prediction models for different
PERFORMANCE COMPARISON OF DIFFERENT SUNNY DAYS’ DECOMPOSITION
PREDICTION MODELS USING THE DATASET OF JILIN POWER STATION
datasets. Taking Ningxia Power Station as an example, in cloudy
days, compared with DWT, VMD, and direct prediction models,
the MAPE accuracy of the FDD model proposed in this article
is increased by 57.42%, 29.64%, and 53.86%, respectively, the
RMSE accuracy is increased by 52.16%, 28.22%, and 50.10%,
respectively, and the R is increased by 14.77%, 4.16%, and
12.94%, respectively. The performance comparison result of the
Jilin Power Station, as given in Table VII, is similar to Ningxia
Power Station. These changes have slightly reduced the accuracy
of PV power generation forecasts. The FDD prediction model
proposed in this article can improve the prediction performance
is increased by 38.91%, 25.72%, and 36.20%, respectively, and of the model under cloudy conditions.
the R is increased by 2.24%, 1.41%, and 2.21%, respectively. 3) Comparison Analysis of Light Rainy Days: Fig. 12(c)
The performance comparison result of the Jilin Power Station, shows the actual and forecasted PV power on a light rainy day
as given in Table V, is similar to Ningxia Power Station. This using the dataset of Ningxia Power Station. Fig. 13(c) shows
is mainly because on a sunny day, since the sky has almost no the error distribution curves of the direct prediction model and
shadow and the physical state of the atmosphere is relatively frequency decomposition prediction model on a light rainy day
stable, the PV power curve is smooth with small fluctuations. using the dataset of Ningxia Power Station. Fig. 14(c) shows the
Therefore, the FDD prediction model, as proposed in this article, linear regression graph of direct prediction model and frequency
will not bring about a very significant improvement in prediction decomposition prediction model on a light rainy day using
performance. the dataset of Ningxia Power Station. Fig. 15(c) shows the
2) Comparison Analysis of Cloudy Days: Fig. 12(b) shows generation of forecast error graph of different light rainy days’
the actual and forecasted PV power on a cloudy day using decomposition prediction models using the dataset of Ningxia
the dataset of Ningxia Power Station. Fig. 13(b) shows the Power Station. Tables VIII and IX present the performance com-
error distribution curves of the direct prediction model and parison of the light rainy weather prediction models for different
frequency decomposition prediction model on a cloudy day datasets. Taking Ningxia Power Station as an example, in light

Authorized licensed use limited to: Universiti Malaysia Pahang. Downloaded on May 12,2023 at 07:16:53 UTC from IEEE Xplore. Restrictions apply.
YAN et al.: FDD AND DEEP LEARNING BASED SOLAR PV POWER ULTRA-SHORT-TERM FORECASTING MODEL 3291

TABLE VIII TABLE X

PERFORMANCE COMPARISON OF DIFFERENT LIGHT RAINY DAYS’ PERFORMANCE COMPARISON OF DIFFERENT HEAVY RAINY DAYS’
DECOMPOSITION PREDICTION MODELS USING THE DATASET OF DECOMPOSITION PREDICTION MODELS USING THE DATASET OF
NINGXIA POWER STATION NINGXIA POWER STATION

TABLE IX TABLE XI
PERFORMANCE COMPARISON OF DIFFERENT LIGHT RAINY DAYS’ PERFORMANCE COMPARISON OF DIFFERENT HEAVY RAINY DAYS’
DECOMPOSITION PREDICTION MODELS USING THE DATASET OF DECOMPOSITION PREDICTION MODELS USING THE DATASET OF
JILIN POWER STATION JILIN POWER STATION

rainy days, compared with DWT, VMD, and direct prediction

models, the MAPE accuracy of the FDD model proposed in this
article is increased by 70.11%, 36.79%, and 67.77%, respec-
tively, the RMSE accuracy is increased by 59.05%, 33.22%, and
57.79%, respectively, and the R is increased by 77.84%, 40.25%,
and 45.64%, respectively. The performance comparison result
of the Jilin Power Station, as given in Table IX, is similar to
Ningxia Power Station. Due to the changes in the physical state
of the atmosphere during light rainy days, the PV power curves
of light rainy days may include spikes and nonlinear fluctuation
components. The accuracy of PV power generation forecasts is
reduced. The FDD prediction model proposed in this article can
improve the prediction performance of the model under light
rainy days. Fig. 16. Performance of different prediction models on four different weather
types using the dataset of Ningxia Power Station. (a) Sunny. (b) Cloudy.
4) Comparison Analysis of Heavy Rainy Days: Fig. 12(d) (c) Light rainy. (d) Heavy rainy.
shows the actual and forecasted PV power on a heavy rainy day
using the dataset of Ningxia Power Station. Fig. 13(d) shows
the error distribution curves of the direct prediction model and
and the R increased by 49.02%, 6.43%, and 37.69%, respec-
frequency decomposition prediction model on a heavy rainy day
tively. The performance comparison result of the Jilin Power
using the dataset of Ningxia Power Station. Fig. 14(d) shows the
Station, as given in Table XI, is similar to Ningxia Power Station.
linear regression graph of direct prediction model and frequency
In heavy rainy days, the irradiance attenuation is serious, and the
decomposition prediction model on a heavy rainy day using the
PV power curve of heavy rainy days includes more spikes and
dataset of Ningxia Power Station. Fig. 15(d) shows the genera-
nonlinear fluctuation components, which makes the accuracy
tion of forecast error graph of different heavy rainy days’ decom-
of PV power generation prediction poor. The FDD prediction
position prediction models using the dataset of Ningxia Power
model proposed in this article can improve the prediction per-
Station. Tables X and XI present the performance comparison of
formance of the model under heavy rainy days.
the heavy rainy weather prediction models for different datasets.
Taking Ningxia Power Station as an example, in heavy rainy
days, compared with DWT, VMD, and direct prediction models, F. Comparison and Analysis of the Performance of Different
PV Forecasting Models
the MAPE accuracy of the FDD model proposed in this article
is increased by 64.07%, 31.21%, and 52.97%, respectively, the This article uses the unified FDD model to compare the CNN
RMSE accuracy is 63.30%, 23.19%, and 57.56%, respectively, prediction model with other prediction models. Figs. 16 and 17

Authorized licensed use limited to: Universiti Malaysia Pahang. Downloaded on May 12,2023 at 07:16:53 UTC from IEEE Xplore. Restrictions apply.
3292 IEEE TRANSACTIONS ON INDUSTRY APPLICATIONS, VOL. 57, NO. 4, JULY/AUGUST 2021

In addition, the results also show that the training time for
CNN prediction is reduced. This is because the weight-sharing
technology enables CNN to reduce the parameters to be es-
timated, thereby shortening the training time, and CNN can
perform feature extraction on data, which has good robust-
ness. Therefore, the CNN model can improve the accuracy
and efficiency of the ultra-short-term prediction of PV power
generation.

V. CONCLUSION
To overcome the shortcomings that there is currently no
effective method to support the rationality of the FDD results,
this article proposes an ultra-short-term prediction model of PV
Fig. 17. Performance of different prediction models on four different weather power based on the optimal FDD and DL. First, the amplitude
types using the dataset of Jilin Power Station. (a) Sunny. (b) Cloudy. (c) Light and phase of each frequency sine wave are obtained by fast
rainy. (d) Heavy rainy. Fourier decomposition. As the frequency demarcation point is
different, the correlation between the decomposition component
and the original data is analyzed. By minimizing the square of
the difference that the correlation between the low-frequency
show the MAPE, RMSE, R, and time of different prediction components and raw data is subtracted from the correlation be-
models for sunny, cloudy, light rainy, and heavy rainy days tween the high-frequency components and raw data, the optimal
using the data of Ningxia and Jilin power stations. Compared frequency demarcation points for decomposition components
with other prediction models, such as back propagation (BP) are obtained. Then, CNN is used to predict the low-frequency
neural network, RNN, and LSTM, the prediction accuracy of component and high-frequency component, and the final pre-
the CNN prediction model is higher, and due to weight sharing, diction result is obtained by addition reconstruction.
the model training time is greatly shortened. Taking Ningxia For the decomposition model proposed in this article, by min-
power station data on a heavy rainy day as an example, after imizing the square of the difference that the correlation between
decomposing uniformly using the FDD model proposed in this the low-frequency components and raw data is subtracted from
article, compared with BP neural network, RNN, and LSTM the correlation between the high-frequency components and raw
models, the MAPE accuracy of the CNN prediction model data, the optimal frequency demarcation points for decomposi-
is improved by 14.02%, 23.64%, and 46.22%, and compared tion components are obtained, and the PV power is decomposed
with RNN and LSTM models, the training time efficiency of into the low-frequency components and high-frequency compo-
CNN prediction model is improved by 85.63% and 87.68%. nents. The low-frequency component represents the regular part
Although the training time of the BP neural network is a little of PV power generation, which shows its trend characteristics,
less than CNN, CNN has better prediction accuracy. Generally while the high-frequency component represents the randomness
speaking, compared with other prediction models, such as BP of PV power generation, which shows its fluctuation character-
neural network, RNN, and LSTM, the prediction accuracy of istics affected by other factors, such as weather. This method
the CNN prediction model is higher, and due to weight sharing, supports the rationality of the decomposition results and solves
the model training time is greatly shortened. The results fully the problem that the current decomposition model only uses the
show that the CNN model proposed in this article improves the direct decomposition method, and the decomposition compo-
prediction accuracy and time efficiency. nents have no physical meaning. Taking Ningxia power station
data as an example, compared with discrete wavelet transform,
VMD, and direct prediction models, the MAPE accuracy of
G. Simulation Discussion the proposed FDD prediction model is improved by 42.05%,
It can be seen from the above results that in the four weather 32.94%, and 44.59%, respectively, on a sunny day, by 57.42%,
types, the proposed FDD method improves the prediction accu- 29.64%, and 53.86% on a cloudy day, by 70.11%, 36.79%, and
racy. Although the high-frequency components are more volatile 67.77% on a light rainy day, and 64.07%, 31.21%, and 52.97%
after the lack of low-frequency components, this may cause on a heavy rainy day. The RMSE accuracy of the proposed FDD
the prediction accuracy of the high-frequency components to prediction model is improved by 38.91%, 25.72%, and 36.20%,
decrease. However, the method of selecting the best frequency respectively, on a sunny day, by 52.16%, 28.22%, and 50.10% on
division point proposed in this article not only increases the a cloudy day, by 59.05%, 33.22%, and 57.79% on a light rainy
proportion of low-frequency components in the original data day, and by 63.30%, 23.19%, and 57.56% on a heavy rainy
but also balances the difficulty of predicting high-frequency day. The R accuracy of the proposed FDD prediction model
components to a certain extent. Therefore, after adding and is improved by 2.24%, 1.41%, and 2.21%, respectively, on a
reconstructing low-frequency and high-frequency components, sunny day, by 14.77%, 4.16%, and 12.94% on a cloudy day,
the accuracy of the final result is usually improved. by 77.84%, 40.25%, and 45.64% on a light rainy day, and by

Authorized licensed use limited to: Universiti Malaysia Pahang. Downloaded on May 12,2023 at 07:16:53 UTC from IEEE Xplore. Restrictions apply.
YAN et al.: FDD AND DEEP LEARNING BASED SOLAR PV POWER ULTRA-SHORT-TERM FORECASTING MODEL 3293

49.02%, 6.43%, and 37.69% on a heavy rainy day. The results recognition model for short-term photovoltaic power forecasting,” Energy
fully show that the FDD method in this article improves the Buildings, vol. 86, pp. 427–438, Jan. 2015.
[8] F. Wang, Z. Mi, S. Su, and H. Zhao, “Short-term solar irradiance fore-
consistency of experimentally observed data and forecast data. casting model based on artificial neural network using statistical feature
For the prediction model proposed, CNN is not only success- parameters,” Energies, vol. 5, no. 5, pp. 1355–1370, May 2012.
fully applied to image processing but also used to solve 1-D [9] H. Yang, C. Huang, Y. Huang, and Y. Pai, “A weather-based hybrid method
for 1-day ahead hourly forecasting of PV power output,” IEEE Trans.
data classification and regression problems. CNN reduces the Sustain. Energy, vol. 5, no. 3, pp. 917–926, Jul. 2014.
parameters to be estimated due to the weight-sharing technology, [10] J. Shi, W. Lee, Y. Liu, Y. Yang, and P. Wang, “Forecasting power out-
thereby shortening the training time for prediction, and CNN put of photovoltaic systems based on weather classification and support
vector machines,” IEEE Trans. Ind. Appl., vol. 48, no. 3, pp. 1064–1069,
can perform feature extraction on the data, which has good May/Jun. 2012.
robustness. Therefore, this article is dedicated to the reasonable [11] P. Lin, Z. Peng, Y. Lai, S. Cheng, Z. Chen, and L. Wu, “Short-term
application of the CNN model to improve the accuracy and power prediction for photovoltaic power plants using a hybrid improved
k means-GRA-Elman model based on multivariate meteorological fac-
efficiency of ultra-short-term prediction of PV power. Taking tors and historical power datasets,” Energy Convers. Manage., vol. 177,
Ningxia power station data on a heavy rainy day as an example, pp. 704–717, Dec. 2018.
after decomposing uniformly using the FDD model proposed in [12] H. Zhang et al., “Hybrid method for short-term photovoltaic power
forecasting based on deep convolutional neural network,” IET Gener.,
this article, compared with BP neural network, RNN, and LSTM Transmiss. Distrib., vol. 12, no. 20, pp. 4557–4567, Nov. 2018.
models, the MAPE accuracy of the CNN prediction model [13] V. Kushwaha and N. Pindoriya, “A SARIMA-RVFL hybrid model assisted
is improved by 14.02%, 23.64%, and 46.22%, and compared by wavelet decomposition for very short-term solar PV power generation
forecast,” Renewable Energy, vol. 140, pp. 124–139, Sep. 2019.
with RNN and LSTM models, the training time efficiency of [14] W. Zhang, H. Dang, and R. Simoes, “A new solar power output prediction
CNN prediction model is improved by 85.63% and 87.68%. based on hybrid forecast engine and decomposition model,” ISA Trans.,
Although the training time of the BP neural network is a little vol. 81, pp. 105–120, Oct. 2018.
[15] M. Lima, P. Carvalho, L. Fernandez-Ramírez, and A. Braga, “Improving
less than CNN, CNN has better prediction accuracy. Generally solar forecasting using deep learning and portfolio theory integration,”
speaking, compared with other prediction models, such as BP Energy, vol. 195, Mar. 2020, Art. no. 117016.
neural network, RNN, and LSTM, the prediction accuracy of [16] M. Madhiarasan, “Accurate prediction of different forecast horizons wind
speed using a recursive radial basis function neural network,” Protection
the CNN prediction model is higher, and due to weight sharing, Control Mod. Power Syst., vol. 5, no. 1, 2020, Art. no. 22.
the model training time is greatly shortened. The results fully [17] J. Liu, W. Fang, X. Zhang, and C. Yang, “An improved photovoltaic power
show that the CNN model proposed in this article improves the forecasting model with the assistance of aerosol index data,” IEEE Trans.
Sustain. Energy, vol. 6, no. 2, pp. 434–442, Apr. 2015.
prediction accuracy and time efficiency. [18] M. Bozorg, A. Bracale, P. Caramia, G. Carpinelli, M. Carpita, and P. De
The above results show that the FDD prediction model Falco, “Bayesian bootstrap quantile regression for probabilistic photo-
has high superiority in the accuracy and time efficiency of voltaic power forecasting,” Protection Control Mod. Power Syst., vol. 5,
no. 1, 2020, Art. no. 21.
ultra-short-term prediction of PV power, especially in extreme [19] Z. Zhen et al., “SVM based cloud classification model using total sky
weather conditions [31]. The ultra-short-term prediction of PV images for PV power forecasting,” in Proc. IEEE Power Energy Soc. Innov.
power based on DL proposed in this article has very high Smart Grid Technol. Conf., Washington, DC, USA, Feb. 2015, pp. 1–5.
[20] X. Qing and Y. Niu, “Hourly day-ahead solar irradiance prediction
application prospects in the area of demand response aggregators using weather forecasts by LSTM,” Energy, vol. 148, pp. 461–468,
[32], such as aggregated capacity forecasting [33], automatic Apr. 2018.
demand response strategy [34], and day-ahead optimal bidding [21] F. Wang, Z. Xuan, Z. Zhen, K. Li, T. Wang, and M. Shi, “A day-ahead
PV power forecasting method based on LSTM-RNN model and time
and scheduling [35]. correlation modification under partial daily pattern prediction framework,”
Energy Convers. Manage., vol. 212, May 2020, Art. no. 112766.
[22] Q. Huang and S. Wei, “Improved quantile convolutional neural net-
REFERENCES work with two-stage training for daily-ahead probabilistic forecasting
of photovoltaic power,” Energy Convers. Manage., vol. 220, Sep. 2020,
[1] K. Li, F. Wang, Z. Mi, M. Fotuhi-Firuzabad, N. Dui, and T. Wang, “Capac- Art. no. 113085.
ity and output power estimation approach of individual behind-the-meter [23] M. Alam, S. Rehman, L. Al-Hadhrami, and J. Meyer, “Extraction of the
distributed photovoltaic system for demand response baseline estimation,” inherent nature of wind speed using wavelets and FFT,” Energy Sustain.
Appl. Energy, vol. 253, Nov. 2019, Art. no. 113595. Develop., vol. 22, pp. 34–47, Oct. 2014.
[2] F. Wang, Z. Zhen, B. Wang, and Z. Mi, “Comparative study on KNN [24] M. Yang and X. Huang, “Ultra-short-term prediction of photovoltaic power
and SVM based weather classification models for day ahead short term based on periodic extraction of PV energy and LSH algorithm,” IEEE
solar PV power forecasting,” Appl. Sci.-Basel, vol. 8, no. 1, Jan. 2018, Access, vol. 6, pp. 51200–51205, 2018.
Art. no. 28. [25] M. Yang and X. Huang, “An evaluation method of the photovoltaic
[3] F. Wang, Z. Zhen, C. Liu, Z. Mi, M. Shafie-Khah, and J. P. S. Catalao, power prediction quality,” Math. Problems Eng., vol. 2018, 2018,
“Time-section fusion pattern classification based day-ahead solar irradi- Art. no. 9049215.
ance ensemble forecasting model using mutual iterative optimization,” [26] T. Kim and S. Cho, “Predicting residential energy consumption using
Energies, vol. 11, no. 1, Jan. 2018, Art. no. 184. CNN-LSTM neural networks,” Energy, vol. 182, pp. 72–81, Sep. 2019.
[4] Z. Zhen et al., “Pattern classification and PSO optimal weights based [27] F. Wang, Y. Yu, Z. Zhang, J. Li, Z. Zhen, and K. Li, “Wavelet decomposition
sky images cloud motion speed calculation method for solar PV power and convolutional LSTM networks based improved deep learning model
forecasting,” IEEE Trans. Ind. Appl., vol. 55, no. 4, pp. 3331–3342, for solar irradiance forecasting,” Appl. Sci.-Basel, vol. 8, no. 8, Aug. 2018,
Jul./Aug. 2019. Art. no. 1286.
[5] F. Wang et al., “Image phase shift invariance based cloud motion dis- [28] F. Wang et al., “Generative adversarial networks and convolutional neural
placement vector calculation method for ultra-short-term solar PV power networks based weather classification model for day ahead short-term
forecasting,” Energy Convers. Manage., vol. 157, pp. 123–135, Feb. 2018. photovoltaic power forecasting,” Energy Convers. Manage., vol. 181,
[6] J. Kamadinata, T. Ken, and T. Suwa, “Sky image-based solar irradiance pp. 443–462, Feb. 2019.
prediction methodologies using artificial neural networks,” Renewable [29] S. Al-Dahidi, O. Ayadi, M. Alrbai, and J. Adeeb, “Ensemble approach of
Energy, vol. 134, pp. 837–845, Apr. 2019. optimized artificial neural networks for solar photovoltaic power predic-
[7] F. Wang, Z. Zhen, Z. Mi, H. Sun, S. Su, and G. Yang, “Solar irradiance fea- tion,” IEEE Access, vol. 7, pp. 81741–81758, 2019.
ture extraction and support vector machines based weather status pattern

Authorized licensed use limited to: Universiti Malaysia Pahang. Downloaded on May 12,2023 at 07:16:53 UTC from IEEE Xplore. Restrictions apply.
3294 IEEE TRANSACTIONS ON INDUSTRY APPLICATIONS, VOL. 57, NO. 4, JULY/AUGUST 2021

[30] R. Ahmed, V. Sreeram, Y. Mishra, and M. D. Arif, “A review and eval- Fei Wang (Senior Member, IEEE) received the B.S.
uation of the state-of-the-art in PV solar power forecasting: Techniques degree from Hebei University, Baoding, China in
and optimization,” Renewable Sustain. Energy Rev., vol. 124, May 2020, 1993, the M.S. and Ph.D. degree from North China
Art. no. 109792. Electric Power University (NCEPU), Baoding, China,
[31] L. Hu et al., “Ultra-short-term solar PV power forecasting method based in 2005 and 2013, respectively, all in electrical engi-
on frequency-domain decomposition and deep learning,” in Proc. IEEE neering.
Ind. Appl. Soc. Annu. Meeting, Detroit, MI, USA, 2020, pp. 1–8, doi: He is currently a Professor with the Department
10.1109/IAS44978.2020.9334889. of Electrical Engineering, NCEPU and the State Key
[32] X. Lu, K. Li, H. Xu, F. Wang, Z. Zhou, and Y. Zhang, “Fundamentals and Laboratory of Alternate Electrical Power System with
business model for resource aggregator of demand response in electricity Renewable Energy Sources, Baoding and Beijing,
markets,” Energy, vol. 204, May 2020, Art. no. 117885. China. He is the Director of Smart Energy Network
[33] F. Wang et al., “Smart households’ aggregated capacity forecasting for Integrated Operation Research Center (SENIOR) and the leader of “Double
load aggregators under incentive-based demand response programs,” IEEE First-Class” research team project at NCEPU. He was a Visiting Professor with
Trans. Ind. Appl., vol. 56, no. 2, pp. 1086–1097, Mar./Apr. 2020. the Department of Electrical and Computer Engineering, University of Illinois at
[34] Q. Chen et al., “Dynamic price vector formation model-based automatic Urbana-Champaign, Urbana, IL, USA, from 2016 to 2017. He was a Researcher
demand response strategy for PV-assisted EV charging stations,” IEEE with the Department of Electrical Engineering, Tsinghua University, Beijing,
Trans. Smart Grid, vol. 8, no. 6, pp. 2903–2915, Nov. 2017. China, from 2014 to 2016.
[35] F. Wang et al., “Day-ahead optimal bidding and scheduling strategies Prof. Wang is an Editor of the IEEE Transactions on Sustainable Energy,
for DER aggregator considering responsive uncertainty under real-time an Editor of the IEEE Power Engineering Letters, an Associate Editor of the
pricing,” Energy, vol. 213, Dec. 2020, Art. no. 118765. IET Renewable Power Generation, an Editor of the IEEE Open Access Journal
of Power and Energy and Protection and Control of Modern Power Systems.
He was the Guest Editor for the Special Issue on “Demand Side Management
Jichuan Yan (Student Member, IEEE) received the and Market Design for Renewable Energy Support and Integration” of IET
B.S. degree in electrical engineering from the Zhe- Renewable Power Generation. He is an IEEE Senior Member and the Expert
jiang University of Technology, Hangzhou, China, in Member of IEC SC8A/WG2. He supervised more than 60 Post-docs, Ph.D.
2020. He is currently working toward the M.S. de- and M.Sc. students. He has authored or coauthored more than 220 publications,
gree in electrical engineering with the Department of including 80 journal papers. His research interests include renewable energy
Electrical Engineering, North China Electric Power power, electricity price and electricity load forecasting; demand response and
University, Baoding, China. electricity market; smart grid; microgrid; and integrated energy system. He was
His research interests include demand response, the recipient of the 2020 Science and Technology Progress First Award of Hebei
power forecasting, and electricity market. Province, 2018 Technical Invention First Award of Hebei Province, the 2018
Patent Third Award of Hebei Province, the 2014 Natural Sciences Academic
Innovation Achievement Award of Hebei Province, the 2018 China Electric
Power Science and Technology Progress Third Award, and the 2014 Outstanding
Doctoral Dissertation Award of NCEPU. He was the General Chair of the 2017
International Seminar of Renewable Energy Power Forecasting and Absorption
Technology and 2018 International Seminar of Integrated Energy and Smart
Microgrid Technology.
Lin Hu received the B.S. degree in electrical engi-
neering and automation from Hebei Normal Univer-
sity, Shijiazhuang, China, in 2018. She is currently Gang Qiu received the B.S. degree from Tianjin Uni-
working toward the M.S. degree in electrical engi- versity, Tianjin, China, in 2000, and the M.S. degree
neering with the Department of Electrical Engineer- from Beijing Jiaotong University, Beijing, China, in
ing, North China Electric Power University, Baoding, 2007, both in electrical engineering.
China. He is currently a Senior Engineer with the Dis-
Her research interests include solar photovoltaic patch and Control Center, State Grid Xinjiang Electric
power and net load forecasting. Power Company, Ltd., Urumqi, China. He has long
been engaged in new energy operation control, power
forecasting mid- and long-term power safety verifi-
cation, and large-scale power grid operation analysis
and control.

Zhao Zhen (Member, IEEE) received the B.S. and

Ph.D. degrees in electrical engineering from North Yu Li received the B.S. degree in electrical engi-
China Electric Power University (NCEPU), Baoding, neering from North China Electric Power University
China, in 2012 and 2018, respectively. (NCEPU), Baoding, China, in 1992.
He is currently an Assistant Professor with the De- He is currently a Professor of Engineering with
partment of Electrical Engineering, NCEPU, and also Dispatch and Control Center, State Grid Xinjiang
a Postdoctoral Researcher with the Department of Electric Power Company, Ltd., Urumqi, China. He
Electrical Engineering, Tsinghua University, Beijing, is the Standing Director of IEEE PES Energy Storage
China. His research interests include solar irradiance Technical Committee, China, and Energy Storage
and photovoltaic power forecasting, wind speed, and Market and Planning Subcommittee, and the member
power forecasting. of the Energy Industry Wind Power Grid-Connected
Standards Committee. He has long been engaged in
large-scale power grid operation analysis and control, energy storage planning,
and network source load storage.

Authorized licensed use limited to: Universiti Malaysia Pahang. Downloaded on May 12,2023 at 07:16:53 UTC from IEEE Xplore. Restrictions apply.
YAN et al.: FDD AND DEEP LEARNING BASED SOLAR PV POWER ULTRA-SHORT-TERM FORECASTING MODEL 3295

Liangzhong Yao received the M.Sc. and Ph.D. de- João P. S. Catalão (Senior Member, IEEE) received
grees in electrical engineering from Tsinghua Univer- the M.Sc. degree from the Instituto Superior Técnico
sity, Beijing, China, in 1989 and 1993, respectively. (IST), Lisbon, Portugal, in 2003, and the Ph.D. degree
He is currently a Professor with the School and Habilitation for Full Professor (“Agregação”)
of Electrical Engineering and Automation, Wuhan from the University of Beira Interior (UBI), Covilha,
University, Wuhan, China. Before joining the Portugal, in 2007 and 2013, respectively, all in elec-
Wuhan University in 2019, he had worked with trical engineering.
Tsinghua University, University of Manchester Currently, he is a Professor at the Faculty of En-
(former UMIST), ABB, U.K., Ltd., Alstom Grid, gineering of the University of Porto (FEUP), Porto,
U.K., and State Grid Corporation of China, from 1993 Portugal, and Research Coordinator at INESC TEC.
to 2019. His research fields cover renewable energy He was also appointed as Visiting Professor by North
grid integration and control, energy storage, and HVdc and dc grid technologies. China Electric Power University (NCEPU), Beijing, China. He was the Primary
Coordinator of the EU-funded FP7 project SiNGULAR (“Smart and Sustain-
able Insular Electricity Grids Under Large-Scale Renewable Integration”), a
5.2-million-euro project involving 11 industry partners. He has authored or
coauthored more than 850 publications, including 400 journal papers (125 IEEE
Transactions/Journal papers), 400 conference proceedings papers (vast majority
co-sponsored by IEEE), 5 books, 41 book chapters, and 14 technical reports,
with an h-index of 67, an i10-index of 320, and over 17,500 citations (according
Miadreza Shafie-khah (Senior Member, IEEE) re- to Google Scholar), having supervised more than 85 post-docs, Ph.D. and M.Sc.
ceived the M.Sc. and Ph.D. degrees in electrical en- students. He was the General Chair of SEST 2019, technically sponsored by
gineering from Tarbiat Modares University, Tehran, IEEE PES and IEEE IES, the General Co-Chair of SEST 2020, technically
Iran, in 2008 and 2012, respectively, first postdoctoral sponsored by IEEE PES, IEEE IES and IEEE IAS, and also the Honorary Chair
degree in electrical engineering from the University of SEST 2021, technically sponsored by IEEE PES, IEEE IES, IEEE IAS and
of Beira Interior, Covilha, Portugal, in 2015, and the IEEE PELS. He is the Editor of the books entitled “Electric Power Systems:
second postdoctoral degree in electrical engineering Advanced Forecasting Techniques and Optimal Generation Scheduling” and
from the University of Salerno, Salerno, Italy, in 2016. “Smart and Sustainable Power Systems: Operations, Planning and Economics
He is currently an Associate Professor with the of Insular Electricity Grids” (Boca Raton, FL, USA: CRC Press, 2012 and
University of Vaasa, Vaasa, Finland. He was consid- 2015, respectively). His research interests include power system operations
ered as one of the Outstanding Reviewers of the IEEE and planning, distributed renewable generation, power system economics and
TRANSACTIONS ON SUSTAINABLE ENERGY, in 2014 and 2017, one of the Best electricity markets, demand response and smart grid.
Reviewers of the IEEE TRANSACTIONS ON SMART GRID, in 2016 and 2017, Prof. Catalão is the Senior Editor of the IEEE TRANSACTIONS ON SMART
and one of the Outstanding Reviewers of the IEEE TRANSACTIONS ON POWER GRID, the Promotion and Outreach (Senior) Editor of the IEEE OPEN AC-
SYSTEMS, in 2017 and 2018. His research interests include power market simu- CESS JOURNAL OF POWER AND ENERGY, an Associate Editor of the IEEE
lation, market power monitoring, power system optimization, demand response, TRANSACTIONS ON POWER SYSTEMS, an Associate Editor of the IEEE POWER
electric vehicles, price forecasting, and smart grids. ENGINEERING LETTERS, an Associate Editor of the IEEE TRANSACTIONS ON
INDUSTRIAL INFORMATICS, an Associate Editor of the IEEE TRANSACTIONS ON
NEURAL NETWORKS AND LEARNING SYSTEMS, an Associate Editor of the IEEE
TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, an Associate Ed-
itor of the IEEE TRANSACTIONS ON CLOUD COMPUTING, an Associate Editor
of the IEEE TRANSACTIONS ON INDUSTRY APPLICATIONS, an Associate Editor
of the IEEE SYSTEMS JOURNAL, and an Associate Editor of IEEE ACCESS.
From 2011 till 2018 he was an Associate Editor of the IEEE TRANSACTIONS ON
SUSTAINABLE ENERGY, and from 2013 till 2020 he was an Associate Editor of the
IEEE TRANSACTIONS ON SMART GRID. He was the Guest Editor-in-Chief for the
Special Section on “Real-Time Demand Response” of the IEEE TRANSACTIONS
ON SMART GRID, published in December 2012, the Guest Editor-in-Chief for the
Special Section on “Reserve and Flexibility for Handling Variability and Uncer-
tainty of Renewable Generation” of the IEEE TRANSACTIONS ON SUSTAINABLE
ENERGY, published in April 2016, and the Corresponding/Lead Guest Editor
(Guest Editor-in-Chief) for the Special Section on “Industrial and Commercial
Demand Response” of the IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS,
published in November 2018. Currently, he is the Guest Editor-in-Chief for
the Special Section on “Demand Response Applications of Cloud Computing
Technologies” of the IEEE TRANSACTIONS ON CLOUD COMPUTING, and also
the Guest Lead Editor (Guest Editor-in-Chief) for the Special Section on “Real-
World Challenges of TSO-DSO Coordination” of the IEEE TRANSACTIONS ON
POWER SYSTEMS. He was the recipient of the 2011 Scientific Merit Award UBI-
FE/Santander Universities, the 2012 Scientific Award UTL/Santander Totta, the
2016-2019 (four years in a row) FEUP Diplomas of Scientific Recognition,
the 2017 Best INESC-ID Researcher Award, and the 2018 Scientific Award
ULisboa/Santander Universities. He is a Top Scientist in the Guide2Research
Ranking (number one in Portugal), which lists only scientists having h-index
equal or greater than 40. Moreover, he has won 5 Best Paper Awards at IEEE
Conferences.

Authorized licensed use limited to: Universiti Malaysia Pahang. Downloaded on May 12,2023 at 07:16:53 UTC from IEEE Xplore. Restrictions apply.