Renewable and Sustainable Energy Reviews 196 (2024) 114349

Contents lists available at ScienceDirect

Renewable and Sustainable Energy Reviews

journal homepage: www.elsevier.com/locate/rser

Wind power forecasting system with data enhancement and

algorithm improvement
Yagang Zhang a, b, c, *, Xue Kong a, b, Jingchao Wang a, b, Hui Wang a, b, Xiaodan Cheng a, b
a
State Key Laboratory of Alternate Electrical Power System with Renewable Energy Sources, North China Electric Power University, Beijing, 102206, China
b
Hebei Key Laboratory of Physics and Energy Technology, North China Electric Power University, Baoding, 071003, China
c
Interdisciplinary Mathematics Institute, University of South Carolina, Columbia, SC, 29208, United States

A R T I C L E I N F O A B S T R A C T

Keywords: Wind power generation has strong volatility. Accurate wind speed forecasting can not only avoid the waste of
Data decomposition and denoising power resources, but also facilitate the development of clean energy and promote the energy transition world­
Chaotic system wide. However, previous research has predominantly focused on the accuracy of wind power prediction, while
Optimization algorithm
ignoring the reliability of wind speed prediction system. In this research, a hybrid forecasting system with both
Hybrid forecasting system
Data enhancement theory
accuracy and reliability of wind power forecasting is proposed. Firstly, a hybrid adaptive decomposition
Wind interval prediction denoising algorithm is proposed to solve the unreasonable decomposition and residual noise. To improve the
search performance, the seagull algorithm is optimized by chaotic system and Cauchy operator, and then the
parameters of long short-term memory model are adjusted. Finally, based on data enhancement theory, an in­
terval prediction model combined with kernel density estimation is proposed. The model is verified by the
historical data of Sotavento wind farm in Spain and Eman wind farm in China. The average absolute percentage
error values of wind speed point prediction are 2.87% and 8.01%, respectively. At the same confidence level, the
interval prediction model proposed has narrower widths compared to the comparative model, with higher
average interval scores. The results indicate that the point prediction model proposed in this research exhibits
higher accuracy, while the interval prediction model demonstrates greater stability and reliability. These findings
provide technical support for wind power forecasting.

reduce greenhouse gas emissions. Therefore, it is necessary to forecast

wind energy. However, the renewable energy generation represented by
1. Introduction
wind power has the characteristics of strong volatility and fast real-time
change. The randomness and intermittency of wind energy leads to
The greenhouse effect has been one of the key issues of global
frequent wind abandonment, which seriously affects the utilization of
concern. It arises due to the emission of greenhouse gases produced by
wind energy. Both the volatility and randomness of wind power can lead
fossil energy. To further reduce greenhouse gas emissions and better
to inaccurate wind power prediction, which affects the safe and stable
meet the challenges posed by climate change, promoting the develop­
power supply of the grid system and causes the waste of wind energy.
ment of renewable energy is a necessary means [1]. According to the
To help grid scheduling, researchers have conducted a lot of studies
latest report of the International Renewable Energy Agency, to achieve a
on the determinism and uncertainty of wind power. Ye et al. [3] pro­
successful energy transition and control the global temperature risk, the
posed a comprehensive method for short-term wind power prediction
installed capacity of renewable energy needs to increase from about
based on frequency analysis, fluctuation clustering and history matching
3000 GW (GW) now to more than 10,000 GW in 2030, so as to promote
to improve the accuracy of wind power prediction. Che et al. [4]
the energy transition dominated by renewable energy [2].
established a prediction model with quantitative uncertainty to guar­
Wind energy, as one of the renewable energy sources, is widely
antee grid scheduling and operation. Wang et al. [5] developed a pre­
developed and utilized all over the world, of which wind power is again
diction model for wind power prediction through the Integral fractional
the most popular application of wind energy. Advance forecasting of
quantile regression for interval prediction of wind power sequences to
wind energy is conducive to grid system integration, improving power
quantify the uncertainty of wind speed fluctuations. Mansoor et al. [6],
system consumption problems, and rationally utilizing wind energy to

* Corresponding author. Hebei Key Laboratory of Physics and Energy Technology, North China Electric Power University, Box 205, Baoding, Hebei, 071003, China.
E-mail address: yagangzhang@ncepu.edu.cn (Y. Zhang).

https://doi.org/10.1016/j.rser.2024.114349
Received 7 June 2023; Received in revised form 25 February 2024; Accepted 26 February 2024
Available online 1 March 2024
1364-0321/© 2024 Elsevier Ltd. All rights reserved.
Y. Zhang et al. Renewable and Sustainable Energy Reviews 196 (2024) 114349

Abbreviation LMD Local mean decomposition

LSTM Long short-term memory
AIS Average interval score MAE Mean absolute error
Bi-GRU Bidirectional gated recurrent unit MAPE Mean absolute percentage error
BP Back Propagation MODWT Maximum overlap discrete wavelet transform
CEEMDAN Complete ensemble empirical Mode decomposition with NWP Numerical weather prediction
adaptive noise OOA Osprey optimization algorithm
CWC Coverage width-based criterion PCA Principal component analysis
EEMD Ensemble empirical mode decomposition PICP PI coverage probability
EMD Empirical mode decomposition PINAW PI normalized averaged width
FA Factor analysis PSO Particle swarm optimization
GA Genetic optimization algorithm RMSE Root mean square error
GCDLA Graph convolutional deep learning architecture RNN Recurrent neural network
GRU Gated Recurrent Unit R2 Coefficient of determination
GW Gigawatt SARIMA Seasonal autoregressive integrated moving average
HADD Hybrid adaptive decomposition denoising SMOTE Synthetic minority over-sampling technique
ICA Independent component analysis SOA Seagull optimization algorithm
IMF Intrinsic Mode Functions SOM Self-organizing mapping
IRENA International Renewable Energy Agency SSD Singular spectrum decomposition
ISOA Inertia seagull optimization algorithm TA Time attention
KDE Kernel density estimation VMD Variational mode decomposition
KPCA Kernel principal component analysis WD Wavelet decomposition
KS Kolmogorov-Smirnov WOA Whale optimization algorithm
LDA Linear discriminant analysis WST Wavelet soft threshold
LLE Local linear embedding WT Wavelet transform

to address the unsatisfactory performance of the existing prediction decomposition was effectively solved. Zhang et al. combined principal
models in short-term power prediction, proposed a new hybrid model component analysis (PCA) with neural network to eliminate redundant
based on the intelligent model of hybrid stochastic algorithms, and at information and improve the accuracy of the prediction model [13]. In
the same time, enhanced the convergence through accelerated compu­ addition, common data screening methods include linear discriminant
tation, thus improving the degree of accuracy of the prediction. Jin et al. analysis (LDA) [14], factor analysis (FA) [15], local linear embedding
[7], for the inherent intermittency and stochasticity of wind energy that (LLE) [16] and kernel principal component analysis (KPCA) [17].
leads to the degradation of the performance of the prediction model, an The data decomposition algorithm is to decompose complex se­
adaptive method based on the integration of offline global and online quences into multiple relatively simple components to further reduce
local learning options is established, thus improving the accuracy and the difficulty of analysis modeling. Commonly decomposition algo­
reliability of wind prediction. Fan et al. [8] introduced fluctuation rithms are used for data feature extraction include singular spectrum
pattern recognition to quantify wind speed volatility by considering the decomposition (SSD) [18], wavelet decomposition (WD) [19] and
probabilistic prediction of wind power under the fluctuation pattern of empirical mode decomposition (EMD) [20]. Paper [21] proposed an
wind power. Farah et al. [9] performed wind power prediction in mul­ Ensemble Empirical Mode Decomposition (EEMD) algorithm and
tiple steps as well as at different elevations and developed a prediction feature selection methods to extract important variables and further
model with high prediction accuracy and faster learning speed on long improve the accuracy of predictions. Tian et al. implemented wind speed
series. Therefore, Improving the accuracy of wind speed prediction and data decomposition and feature extraction using EMD and local mean
quantifying the risk of wind speed fluctuation are still the focus and decomposition (LMD) algorithms, and selects different models for pre­
difficulty of current research. diction based on the different characteristics of odd and even sequences
Wind speed prediction accuracy is closely related to original data [22]. To solve the over-decomposition problem of VMD, paper [23]
features, prediction model selection and error correction methods. The introduced energy entropy theory to determine the number of decom­
research directions at home and abroad can be divided into the position modes and combines sample entropy (SE) theory to determine
following three aspects: (1) data processing algorithm based on data the complexity of different components. Zhang et al. proposed a fully
characteristics; (2) prediction model algorithm aiming at accurate pre­ integrated empirical modal decomposition algorithm based on wavelet
diction; (3) interval prediction method aiming at quantifying data transform (WT) and adaptive noise reduction (CEEMDAN) to effectively
fluctuation. remove noise during the original data acquisition process [24].
Data processing methods are mainly divided into two parts: data However, the traditional signal processing method has many defects.
screening and data decomposition. Data filtering algorithm can improve For example, WD needs to determine the wavelet basis function in
data quality by cleaning invalid data in original data set. Through data advance [25], mode aliasing exists in EMD [26], and the selection of the
dimensionality reduction, some features with the greatest impact on the number of VMD decomposition needs to be preset [27]. CEEMDAN, as
dependent variable are found to participate in the subsequent modeling an improved algorithm of EMD, can effectively solve mode aliasing and
process [10]. Reference [11] adopted the maximum overlap discrete end effects [28]. It has high computational efficiency and good robust­
wavelet transform (MODWT) for noise removal. Then the random forest ness. This research proposes a hybrid adaptive decomposition denoising
algorithm is used to calculate the importance of variables, reduce the algorithm. CEEMDAN was used for data decomposition, combined with
dimension, and reduce the number of variables involved in modeling. In self-organizing mapping (SOM) network and wavelet soft threshold
Ref. [12], independent component analysis (ICA) was introduced to (WST) method for signal denoising [29], so as to reduce the influence of
improve the variable screening effect. Combined with the adaptive noise on prediction results.
decomposition method, the boundary effect problem in the process of Wind power forecasting models are mainly divided into physical

2
Y. Zhang et al. Renewable and Sustainable Energy Reviews 196 (2024) 114349

Fig. 1. Schematic diagram of the Hybrid wind power forecasting system.

model, statistical model and neural network model. Numerical weather prediction model.
prediction (NWP) model is the most common physical model for wind The interval prediction method quantifies the fluctuation range of
speed prediction [30], and accurate wind speed prediction can be ach­ data by analyzing the error sequence. In studies [41], Zhang et al. pro­
ieved through detailed observation records of atmospheric conditions posed an uncertain prediction model based on Monte Carlo method that
and environmental variables. Traditional statistical methods are mainly can adapt to different error characteristics. Jiang et al. proposed a
used to predict stationary time series. The Seasonal Autoregressive In­ combined prediction model based on sample entropy and conditional
tegrated Moving Average (SARIMA) model is used to predict wind kernel density estimation to complete interval prediction [42]. How­
speeds for offshore and offshore wind farms in Scotland. Compared with ever, the error data has strong randomness [43], and the existing error
the Gated Recurrent Unit (GRU) and Long Short-Term Memory (LSTM) correction models have poor generalization ability and low operation
algorithms based on deep learning, SARIMA is more robust and accurate efficiency [44,45]. To improve the reliability of wind speed interval
[31]. Although on some data sets, statistical methods show good pre­ prediction model, data enhancement technology is introduced in this
dictive effect, for complex and variable wind speed data, the neural research. Put synthetic minority over-sampling technique (SMOTE)
network model with strong nonlinear processing ability can achieve sampling on the error sequence [46], get the extended error sequence,
more robust prediction. By combining the traditional Autoregressive then combine with the kernel density estimation method [47], establish
Integrated Moving Average (ARIMA) model with the BP model, the the range of wind speed prediction. This method effectively improves
correlation between wind and wave is analyzed, and the historical wind the running efficiency of the model and avoids the overfitting problem
speed data is fully utilized to further improve the accuracy of the wave caused by Bootstrap resampling.
prediction model [32]. In recent years, the deep learning model has been There are still some problems in the research of wind power pre­
developed rapidly, showing good performance in extracting the inherent diction: (1) the noise signal obtained by the decomposition algorithm is
rules of data and realizing complex classification prediction [33]. In improperly processed, and the effective signal information is lost, which
work [34], deep learning model was used to predict wind speed, feature greatly reduces the subsequent prediction effect; (2) Traditional opti­
extraction was carried out by LSTM, and wind speed at different loca­ mization algorithms are easy to fall into local optimum, weak search
tions was predicted by combining graph convolutional deep learning ability and slow convergence speed; (3) Due to the small amount of error
architecture (GCDLA). Research [35] used bidirectional gated recurrent data obtained in practical applications, it is easy to overfit when dealing
unit (Bi-GRU) to forecast ultra-short-term wind power. The feature with random error sequences during interval estimation. (4) The single
attention (FA) and time attention (TA) mechanisms were introduced to wind speed prediction model has low accuracy, poor performance, and
further improve the prediction effect of the model. The improper can not effectively quantify the uncertainty.
parameter initialization of neural network will lead to the prediction To solve those problems, this study proposes a new prediction model,
deviation. The use of optimization algorithm improves the prediction which is innovative and practical. Firstly, this research proposes a new
accuracy of the model. Common optimization algorithms include ge­ hybrid adaptive decomposition and denoising (HADD) algorithm, which
netic optimization algorithm (GA) [36], particle swarm optimization can adaptively decompose and denoise the original wind speed data.
algorithm (PSO) [37], whale optimization algorithm (WOA) [38] and Through this algorithm, can effectively eliminate the noise and inter­
osprey optimization algorithm (OOA) [39]. Seagull optimization algo­ ference in the data and improve the accuracy and stability of the pre­
rithm (SOA) is a swarm intelligent optimization algorithm proposed in diction model. Secondly, this work utilize the ISOA algorithm to
2018 [40], which imitates the migration and aggressive behavior of gull optimize the LSTM parameters. By continuously adjusting and opti­
and has strong search ability. In this research, the deep learning model mizing the LSTM parameters, the model can be better adapted to the
LSTM model is selected to predict wind speed data and achieve high characteristics of the data, thus improving the prediction performance.
precision prediction. The optimization search of seagull is improved According to the optimized model, preliminary prediction results is
using chaotic mapping and Cauchy distribution theory. Then, improved obtained. Finally, through the in-depth analysis of the error series,
seagull optimization algorithm (ISOA) is used to optimize the parame­ propose a kernel density wind speed interval prediction model based on
ters of the LSTM model to improve the accuracy of the high point data augmentation techniques. This model can quantify the fluctuation

3
Y. Zhang et al. Renewable and Sustainable Energy Reviews 196 (2024) 114349

range of wind speed, which provides a new idea and method for the field existing in wind farm data acquisition and decomposition. Secondly, an
of wind speed prediction. The Schematic diagram of the Hybrid wind improved Seagull algorithm (ISOA) based on mixed chaotic mapping
power forecasting system is shown in Fig. 1. The novelties and contri­ and Cauchy variation is proposed, which improves the population di­
butions of this study are as follows: versity and global search ability of the optimization algorithm. Finally,
on the basis of deterministic prediction, a data enhancement technique
(1) A new hybrid adaptive decomposition and denoising (HADD) is applied to the error sequence and kernel density estimation is com­
algorithm is proposed. By combining CEEMDAN, SOM and WST bined to construct wind speed interval prediction. The following is the
algorithms, the proposed algorithm can effectively avoid pattern specific content of the methodology.
aliasing, reduce residual noise and improve data quality.
(2) A hybrid improvement strategy is adopted to improve the effi­ 2.1. Hybrid adaptive decomposition denoising algorithm (HADD)
ciency of the genetic Optimization algorithm (SOA). Firstly, by
introducing chaotic mapping, the population diversity of the al­ Common decomposition-based denoising algorithms usually treat
gorithm was increased, so as to improve the local search ability of the component with the highest frequency as the noise signal and
the algorithm and avoid falling into the local optimal solution directly manipulate this component. This method can cause the devia­
prematurely. Secondly, to further improve the convergence tion of the predicted results from the actual values. Therefore, based on
ability, the cosine function is introduced to control the moving the shortcomings of existing algorithms, this study proposes a hybrid
direction of the position of the seagulls. The cosine function has adaptive decomposition denoising algorithm that combines CEEMDAN,
the characteristics of periodicity and volatility, which can make SOM, and wavelet soft threshold denoising methods adaptively [48].
the seagulls gradually converge to the vicinity of the optimal The decomposed signal components are obtained using CEEMDAN. The
solution during the search process. By adjusting the parameters of self-organizing mapping neural network is introduced to classify com­
the cosine function, the ability of exploration and exploitation ponents [49], and the components whose classification results are noise
can be balanced, so as to better find the global optimal solution. signals are denoised to further reduce the residual white noise. This
In addition, to avoid the optimization falling into the local algorithm makes the prediction result more robust.
optimal solution, Cauchy mutation perturbation operator is used
to improve the global search ability. 2.1.1. Complete Ensemble Empirical Mode Decomposition with adaptive
(3) A kernel density interval estimation model based on data noise
augmentation algorithm is proposed. The random error sequence CEEMDAN proposed by Torres in 2011 can better solve the problem
is first processed using the Smote algorithm to avoid the over­ of modal aliasing, reduce noise residue, and improve computational
fitting problem. The Smote algorithm balances the class distri­ efficiency, and has been widely used [50]. The advantage of CEEMDAN
bution in the dataset by generating synthetic samples, thereby method lies in its ability to adaptively add appropriate white noise,
reducing the impact of random errors. The model combines the effectively improving the signal-to-noise ratio [51]. It generates a new
kernel density estimation method to solve the problem of un­ decomposition sequence based on each residue to minimize noise
known error distribution. Kernel density estimation is a non- interference. As a result, the error between the reconstructed signal after
parametric estimation method, which can estimate the density decomposition and the original signal is extremely small, which accu­
function of data from a given set of data points. By applying rately reflects the characteristics of the original signal [52].
kernel density estimation to the processed data, a more accurate Ej ( ⋅) is defined as the component obtained by EMD decomposition,
and reliable interval estimate can be obtained. εi (i = 1, 2, …) is defined as the signal-to-noise ratio of the (i +1) − th
(4) A wind speed prediction system combining data decomposition, time, and the specific steps of CEEMDAN decomposition are as follows:
improved optimization algorithm, deep learning and interval
estimation is established. Firstly, the data decomposition method (1) Add Gaussian white noise N1 (t), N2 (t), …, Ni (t), …, Nk (t) with
was used to decompose the complex wind speed series into different signal-to-noise ratio with ε0 to the original sequence
simpler sub-series, so as to better analyze the wind speed change
x(t), decompose the sequence by EMD method to obtain IMFi1 ,
law. Secondly, the optimization algorithm is improved to
whose mean value is the first IMF1 , and the process is shown in
improve the accuracy and efficiency of prediction. Then, the
the following equation.
neural network model is optimized by the optimization algo­
rithm, which can deal with complex characteristics of wind speed 1∑k 1∑k
IMF1 = IMF1i = E1 (x(t) + ε0 Ni (t)) (1)
components, so as to improve the accuracy of prediction. Finally, k i=1 k i=1
the interval estimation method was used to quantify the fluctu­
ation range of wind speed, which could more accurately describe (2) Calculate the first order residual e1 (t).
the fluctuation range of wind speed and improve the stability and (3) The first-order residual e1 (t) is used as the new sequence to be
reliability of the prediction system. decomposed, and the Gaussian white noise with the signal-to-
noise ratio ε1 is added to the new sequence, which is expressed
The sections of this research are arranged as follows. Chapter 2 as: xnew1 = e1 (t) + ε1 E1 {Ni (t)}. After EMD decomposition, the
introduce HADD, ISOA based on chaotic mapping and Cauchy variation, second mode IMF2 is obtained, and the process is expressed as
LSTM and kernel density interval prediction (SKDE) theory based on follows:
data enhancement. Chapter 3 introduces the steps of the model estab­ (4) Calculate the j − th residual sequence.
lishment and the error index of this work. In chapter 4, the deterministic (5) Compute the (j +1) − th mode: Decompose sequence ej (t) +
prediction results on the model real wind farm data set are given. In Ej (Ni (t)) with EMD to obtain the IMFj+1 .
chapter 5, the effect of the proposed interval prediction model is (6) Return to step 4 and continue the calculation until the final re­
introduced. The sixth chapter is the summary of this research, and the sidual eN (t) satisfies the iteration termination condition or rea­
future development direction to make a prospect. ches the maximum number of iterations, and the iteration stops.
The original signal x(t) conforms to the formula:
2. Theoretical basis of methods
j+1
∑
x(t) = IMFi + eN (t) (2)
This chapter mainly introduces the methods used in the model and i=1
the improvement theory. Firstly, HADD is proposed to solve the noise

4
Y. Zhang et al. Renewable and Sustainable Energy Reviews 196 (2024) 114349

2.2. Improved seagull optimization algorithm

2.1.2. HADD algorithm implementation scheme
CEEMDAN does not need to set the number of modes, and the
2.2.1. Seagull optimization algorithm
decomposition speed is fast and the effect is good. However, there will
Seagulls are social seabirds that migrate in groups as the seasons
be noise residual due to the process of wind speed data acquisition and
change [58]. SOA is a new swarm intelligent optimization algorithm
decomposition. Therefore, components are classified by SOM method,
which simulates the migration and foraging behavior of seagulls in na­
which is an unsupervised and self-learning clustering network [53]. It
ture. Migration behavior refers to the process in which seagulls move
consists of the input layer and output layer and is linked by the weight
towards the optimal location from their current position, and during this
vector [54]. The output neuron is connected to other nodes in the
stage, individual seagulls have flight independence. Foraging behavior
domain, and the nodes are activated by competing with each other. The
refers to the predatory behavior of seagulls on prey on the water or land
component signals containing noise are identified by SOM and then
during flight, during which seagulls will spiral toward the prey to launch
processed with wavelet soft threshold denoising [55].
an attack [59]. The algorithm design of migration behavior affects the
Wavelet denoising is an algorithm to effectively separate the wavelet
global search ability of the seagull algorithm, while the foraging
transform of signal from the wavelet transform of noise in the wavelet
behavior determines the local optimization ability of the seagull
domain. The effective signal has a certain continuity in the time domain,
algorithm.
while the noise still has a strong randomness after wavelet transform,
In the migration stage, attention should be paid to avoid the colli­
which can be removed according to the mean square error value [56].
sion, and the position of seagulls keeps getting closer to the optimal
The wavelet threshold denoising method consists of three steps: wavelet
position. The formula for calculating the new position Cs of seagulls to
decomposition, threshold processing and signal reconstruction. The
avoid collisions is as follows.
process of threshold determination has great influence on the denoising
{
result. In the process of calculating the hard threshold function, there Cs (t) = A⋅Ps (t)
(5)
will be signal oscillation phenomenon and the denoising effect is poor. A = fc (1 − t/Maxiteration )
Therefore, using soft threshold function for denoising can avoid the
In the formula (5), Ps represents the current optimal position. A is the
generation of signal break and the signal smoothing effect is better [57].
motion behavior of a seagull in the search space. fc is a function that
Each operation process of HADD is as follows.
controls the frequency of change of A, decreasing linearly from 2 to 0.
Maxiteration is the maximum number of iterations. The seagull will move to
Step 1: White noise with different signal-to-noise ratios is added to
the optimal position, and the movement formula is shown as follows:
the original wind speed data x(t), and the corresponding components
⎧
IMF1 , IMF2 , …, IMFk are obtained by CEEMDAN decomposition. ⎨ Ms (t) = B⋅(Pbs (t) − Ps (t))
Step 2: Initialize the SOM network. The weight ωj of each neuron in B = 2⋅A2 ⋅rd (6)
⎩
the output layer is assigned a random initial value in [0,1], and the Ds = |Cs + Ms |
learning rate η is randomly initiated, 0 < η < 1.
Ms (t) is the direction of the optimal position. Pbs (t) is the current
Step 3: IMF1 is set as the standard component sample, and the input
optimal position. Ps (t) is the current position. B is a random number used
component vector is normalized. Each node of the output layer is
to balance global and local search, and rd ranges from 0 to 1. After
traversed. Here, Euclidean distance is selected as the discriminant
obtaining the convergence direction of the seagull, the seagull will move
rule. Calculate the value between the standard component sample
along this direction, and Ds will be the seagull’s new position.
and each output node, and select the minimum distance as the
Seagulls descend in a spiral pattern as they forage, their flight angle
winning node.
and radius constantly changing. The specific position of the seagull in
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
√ P
√∑ ( three-dimensional space is:
)2
dj = √ xp − wpj (3) ⎧
p=1 ⎪
⎪ x = r⋅cos(θ)
⎨
y = r⋅sin(θ)
(7)
⎪ z = r⋅θ
where, dj is the distance between the weight vector of neurons in the ⎪
⎩
r = u⋅eθv
output layer and the input vector. ωpj is the weight of the j-th neuron in
the output layer of the p-th input vector. xp is the value of the input In the formula, r is the spiral radius of seagull movement, θ is the
vector. angle value of the value range, u and v are spiral shape parameters. The
formula of the seagull’s update position is as follows:
Step 4: Update the weights in the domain of winning nodes and
Ps (t) = (Ds ⋅ x ⋅ y ⋅ z) + Pbs (t) (8)
adjust the formula is as follows:
[ ]
ωij (t + 1) = ωij (t) + η(t, N) xip − ωij (t) (4) 2.2.2. ISOA based on cosine function, cauchy mutation and perturbation
operator
In the formula (4), i is the number of iterations. ωij represents the To further improve the convergence speed of the SOA and enhance
weight of the j-th neuron during the i-th iteration. η is the learning rate. its search ability, this research proposes a hybrid strategy-improved
Seagull Optimization Algorithm. By combining cosine function, Cau­
Step 5: After multiple iterations, the positions of output layer nodes chy mutation and perturbation operator, the optimization algorithm is
are determined. The similarity is determined based on the Euclidean improved from the migration and position correction of seagulls [60].
distance between each component and the standard component. To In formula (5), A affects the migratory behavior of seagulls. It can be
avoid the occurrence of low probability events, the classification seen from the formula that A linearly decreases from fc to 0. To improve
process is iterated 1000 times, and the Euclidean distance after each the local search ability of seagulls in the early stage and the convergence
classification is recorded. Components with Euclidean distance less speed in the middle stage, a cosine function is introduced to define A as:
than 2 from the standard component and occurring more than 600 ( )/
times are considered as the same type of signal containing noise. t
A = fc ⋅ cos π ⋅ 2+1 (9)
Step 6: Wavelet soft threshold denoising is used to process the clas­ Maxiteration
sified noise signal. On the basis of data feature mining, the signal
Perturbation operators are added to the seagull position updating
adaptive noise reduction is realized.
process to improve the ability of SOA to jump out of local optimal. When

5
Y. Zhang et al. Renewable and Sustainable Energy Reviews 196 (2024) 114349

Fig. 2. Flow chart of ISOA based on hybrid strategy and chaotic mapping.

the perturbation operator satisfies β < (1 − t /Maxiteration ), the position algorithm.

updating formula is formula (10); otherwise, Cauchy variation is added To better illustrate the improvement process of ISOA, this work
in the process of position updating, then: combines Section 2.2 and draws the overall framework diagram of ISOA
as shown in Fig. 2.
Ps (t) = (Ds ⋅ x ⋅ y ⋅ z) ⋅ (1 + cauchy) + Pbs (t) (10)
The Cauchy distribution is similar to the standard normal distribu­ 2.3. Long short-term memory network (LSTM)
tion, but has smaller values at the origin and is flatter at the tails,
allowing for greater disturbance [61]. When combined with perturba­ Long Short-Term Memory (LSTM) is a deep learning-based Recurrent
tion operators and Cauchy mutation, it can enhance the search capa­ Neural Network (RNN) that inherits the advantages of RNN models.
bility of the SOA. The standard Cauchy distribution formula is shown in Compared to traditional RNNs, LSTM does not suffer from the vanishing
formula (11). gradient problem and can better store and access information. The basic
1 form of an LSTM structure includes input gates, output gates, forget
cauchy(x; 0, 1) = (11) gates, and memory cell states [65]. The three gates can control the
π(1 + x2 )
memory cell state [66]. The input gate determines the information that
2.2.3. ISOA based on hybrid chaotic mapping will be stored in the memory cell state, the output gate determines the
Although SOA has good optimization capabilities, the random information that will enter the next hidden state, and the forget gate
initialization process of the population increases the possibility of get­ determines whether to forget the information from the previous state
ting stuck in local optima. To enhance the diversity of the population, [67]. The specific steps are as follows:
this work introduces the Logistics-Sine chaotic mapping [62] into the
algorithm. Step 1: The amnesia gate determines how many neurons Ut− 1 from
The logistics chaotic system has a simple structure and can generate the previous moment can be retained for the current moment Ut . The
complex chaotic characteristics. However, the Logistics chaotic range is forgetting gate inputs include the output Yt− 1 of the previous
limited, and when the input parameters are not within the range, the moment and the input Lt of the current time, and ft is obtained by
generated chaotic sequence becomes uneven [63]. Therefore, by inte­ activating the function Sigmoid. The formula is as follows:
{ ( )
grating the one-dimensional Logistics chaotic system with the Sine ft = σ ωf ⋅[Yt− 1 , Lt ] + cf
chaotic system, the resulting compound chaotic system has complex − x − 1 (12)
σ(x) = (1 + e )
chaotic dynamics, faster iteration speed, and is suitable for a large
number of sequences [64]. Chaotic mapping optimization uses chaotic where, ωf is the weight matrix of the forgetting gate, cf is the bias vector,
sequences to generate random numbers. It can effectively increase the and σ is activation function Sigmoid. ft is the probability that the last
coverage of the initial solution space, make the population closer to the layer of neurons will be forgotten, and the value range is [0,1]. ft = 0
optimal solution faster, and accelerate the convergence rate of the means completely discarded, ft = 1 means completely retained.

6
Y. Zhang et al. Renewable and Sustainable Energy Reviews 196 (2024) 114349

Fig. 3. The overall flow chart from point prediction to interval prediction.

7
Y. Zhang et al. Renewable and Sustainable Energy Reviews 196 (2024) 114349

Step 2: Input gates determine the new input in Lt that can be stored in and eij,new is the new error value obtained by interpolation.
neurons, which are mainly divided into pt and qt . The formula is as
follows: Step 3: Interval estimation of wind speed.
⎧ ( )
⎨ pt = σ ω ( p ⋅[Yt− 1 , Lt ] + cp ) The wind speed interval is established according to the expanded
q = tanh ωq ⋅[Yt− 1 , Lt ] + cq (13)
⎩ t error data. The KDE method is selected to fit the error distribution. The
Ut = Ut− 1 ⋅ft + pt qt
formula for KDE is as follows:
In the formulas (13), ωp and ωq are the weight matrices of the input (x − x )
∑
n
1 ∑n
gates. cp and cq are the bias vectors. pt and qt indicate the proportion of ̂f m (x) = 1 Km (x − xi ) = Km
i
(16)
load information that needs to be retained at the current time. The new n i=1
nm i=1 m
neuron state information Ut consists of the neuron state retained at the
previous time and pt and qt . where, x1 , x2 , …, xn is a random variable distributed independently and
identically according to the probability density function f(x). m is a
Step 3: The output gate is used to control the output state of neurons smoothing parameter, and K(•) is a kernel function. In this research, the
and transmit the state to the next neuron. The final output value can Gaussian kernel function is selected as the kernel function, and it is
be obtained by the last layer of calculation. shown in the formula (17).
{ ( )
Ot = σ (ωO [Yt− 1 , Lt ] + cO ) 1 (x − xi )2
(14) K(x, m) = √̅̅̅̅̅ exp − (17)
Yt = Ot ⋅tanh(Ut ) 2π 2m2

where, Ot is the output threshold. ωo is the weight matrix of the output Fit the error sequence and calculate the probability density function
gate, co is the bias vector and Yt is the output value at the current ̂f m (x). According to the confidence level, different sub-sites were ob­
moment. tained to construct the wind speed prediction interval.

2.4. Interval estimation of wind speed based on SKDE method 3. Establishment of prediction model

Due to the limitations of actual wind speed prediction and engi­ To address the intermittency and variability of wind speeds, this
neering applications, only the wind speed prediction results for a certain work proposes a novel hybrid wind speed prediction system that com­
period of time in the future can be obtained, resulting in a relatively bines data decomposition, algorithm optimization, and deep learning
small amount of error data. The simple replication of error samples in models. Using data augmentation techniques to generate wind speed
the Bootstrap method can not introduce new effective information, and prediction intervals, the fluctuation range of wind speed can be effec­
may lead to overfitting of errors [68]. Therefore, the error processing tively quantified. First, the original data are decomposed by CEEMDAN,
method based on data augmentation technology has attracted the and then the SOM is applied to classify the components, and combined
attention. with WST denoising to remove the residual white noise to improve the
SMOTE algorithm is an improved solution of a random oversampling prediction robustness. Secondly, the LSTM model is adopted to predict
algorithm. It can generate new samples by linear interpolation between the wind speed using the denoised component signals. The ISOA is
two existing samples, based on their relationship to expand the dataset employed to enhance the search and convergence ability of the algo­
[69]. Kernel density estimation (KDE) is a non-parametric estimation rithm. Finally, the error data is processed using data augmentation
method for inferring the probability density function of an unknown techniques. KDE method is used to process the error sequence, and wind
variable [70]. This method does not require any assumptions about the speed interval prediction is established. The flow chart of wind speed
distribution function and relies entirely on the distribution character­ prediction model established is shown in Fig. 3.
istics of the data sample itself. Therefore, based on the distribution The specific steps of the hybrid wind speed prediction system
characteristics of the error data, the KDE method can provide interval established in this research are as follows:
estimates for wind speed prediction [8]. The steps of wind speed interval
estimation based on SKDE method are shown as follows: Step 1: Wind speed data were selected from the datasets collected
from the Eman wind farm in China and the Sotavento wind farm in
Step 1: Construct imbalanced classes based on point prediction Spain. The proposed HADD method was used to decompose and filter
results. the data, generating component sub-sequences through iterations.
Step 2: Chaotic mapping and Cauchy mutation algorithm are used to
According to the division of the original data set, the ratio of the improve the global search capability and convergence speed of SOA.
number of sampling points between the training set and the test set is M/ The parameters of deep learning model LSTM are optimized by ISOA.
N (M > N). The wind speed point prediction model is obtained according Using the optimized LSTM model forecast component, the wind
to the training set. Then, the error sequence e1 , e2 , …, eM of the training speed point prediction results are obtained.
set and e1 , e2 , …, eN of the test set are obtained according to the model. Step 3: Four kinds of prediction evaluation indicators are used to
estimate the point prediction results of the proposed model, which
Step 2: Oversampling of error based on SMOTE method. are the mean absolute error (MAE), root mean square error (RMSE),
mean absolute percentage error (MAPE), and coefficient of deter­
Firstly, each e in the test set error is selected as the starting point for mination (R2 ). The prediction evaluation metrics formula is as fol­
oversampling in turn. Then, k nearest neighbor samples is found. A lows:
random one is chosen from the nearby samples each time to generate
new error samples using linear interpolation. This process is repeated n
times.
( )
eij,new = ei + η⋅ eij − ei (15)

where, ei is the i-th sample in the error sequence of the test set and eij is
the j-th adjacent sample of ei . η is the random number between [0,1],

8
Y. Zhang et al. Renewable and Sustainable Energy Reviews 196 (2024) 114349

Table 1 Step 4: Based on the point prediction results, a data enhancement

Statistical index of wind speed distribution. algorithm is introduced. The prediction interval of wind speed is
Wind farm Max Min Mean Std. Lower Upper obtained by oversampling and expanding error data combined with
(m/s) (m/s) (m/s) (m/s) quartile(m/ quartile kernel density estimation algorithm.
s) (m/s) Step 5: The interval prediction effect is evaluated by interval eval­
Sotavento 12.26 0.02 6.58 2.36 5.02 8.19 uation index. PICP represents the coverage probability of the pre­
Eman 10.05 0 4.07 2 2.6 5.25 diction interval. PINAW is the normalized average width of the
prediction interval, which represents the width of the interval at
⎧ different confidence levels. To better evaluate the interval prediction
1 ∑ m
results, taking into account the accuracy and reliability of the pre­
⎪
⎪ MAE = (yi − ̂y i )
⎪
⎪
⎪
⎪
⎪
m i=1 diction, the coverage standard width CWC and the average interval
⎪
⎪
⎪
⎪ ( )12 score AIS are introduced by considering both interval coverage and
⎪ 1 ∑ m
⎪
⎪
⎪
⎪ RMSE = (yi − ̂yi) 2 interval width. Under the same confidence level, the smaller the
⎪
⎪
⎪
⎪
m i=1 CWC value represents the better interval prediction; the larger the
⎨
(18) AIS value represents the better interval prediction.
⎪ 1 ∑ m
|yi − ̂ yi|
⎪
⎪ MAPE = ∑
m
⎪
⎪ m yi
⎪
⎪ i=1 Ai
⎪
⎪
⎪
⎪
⎪ ∑m PICP = i=1 (19)
⎪
⎪
⎪ (yi − ̂y i )2 m
⎪
⎪ 2 i=1
⎪
⎪ R =1− ∑
⎪
⎩ m ∑
m
(yi − y)2 ξi
(20)
i=1
PINAW = i=1
m
where yi represents the true series data, ̂
y i represents the predicted data,
y is the series mean, and m is the number of series samples.

Fig. 4. Data sets of Sotavento and Eman farm, (a) Wind9speed of Sotavento farm, (b) Wind speed of Eman farm.
Y. Zhang et al. Renewable and Sustainable Energy Reviews 196 (2024) 114349

Table 2 last day are used as the prediction set, represented by a red curve, while
Distribution of the components. the training set is represented by a blue curve.
Sotavento Eman Sotavento Eman

IMF1 1000 1000 IMF6 553 2 4.2. Hybrid adaptive decomposition denoising
IMF2 776 909 IMF7 126 5
IMF3 522 870 IMF8 17 23
The original wind speed of wind farm data is processed by HADD.
IMF4 678 385 IMF9 3 45
IMF5 748 5 IMF10 0 34 First, the CEEMDAN method was used to decompose the wind speed and
obtain the corresponding IMF sequence. High-frequency components
( ) may retain white noise, thus IMF components containing noise need to
CWC = PINAW 1 + λe− η(PICP− α)
(21) be identified and denoised.
Input the components obtained by decomposition into the SOM
∑
m
Si neural network model. Set IMF1 component as the standard component
sample and normalize the input components. Calculate the Euclidean
AIS = i=1 (22)
m distance between the remaining components and the IMF1 component,
and determine the similarity of the components by calculating the dis­
⎧ − 2αξi − 4(Li − ̂y i ) ̂ y i < Li
⎨ tance. To reduce experimental randomness, the classification is repeated
Si = − 2αξi Li ≤ ̂y i ≤ Ui (23) 1000 times. Record the IMF components with Euclidean distance to the
⎩
− 2αξi − 4(̂y i − Ui ) ̂ y i > Ui standard component less than 2 units, and consider the components with
repetition frequency greater than 600 to contain noise. The cumulative
In the formula (23), Ai = 1 when the true value is within the pre­ frequency of components is shown in Table 2, which shows that some
diction interval, otherwise, Ai = 0. ξi is the width of the interval. The components have repetition frequency greater than 600. This indicates
value of λ is related to the confidence level. When the PICP is greater that these components are very similar to the IMF1 component and have
than the confidence level, λ = 1, otherwise, λ = 0. Si is the interval noise residues as judged by the SOM neural network. In the Sotavento
fraction of the prediction interval. Li and Ui are the lower and upper wind farm, IMF1, IMF2, IMF4, and IMF5 are considered as noise signal
bounds of the prediction interval, respectively. components. In the wind speed data components of the Eman wind farm,
IMF1, IMF2, and IMF3 are considered as signal components with noise.
4. Wind speed point prediction results and error analysis The signal components containing noise were subjected to wavelet soft-
threshold denoising separately to reduce the residual noise effect. HADD
4.1. Data introduction not only implements data decomposition, reducing the complexity of
prediction, but also achieves adaptive signal denoising, further
This work selects wind speed data from Sotavento wind farm in Spain improving prediction accuracy. The decomposed components of the
and Eman wind farm in China as data sets(see Table 1). The interval original wind speed data obtained through mixed adaptive decomposi­
between both wind farms’ data is 10 min. Fig. 4(a) displays wind speed tion denoising are shown in Figs. 5 and 6.
data from Sotavento wind farm, where the wind speeds are mainly
distributed in the range of [5.015, 8.19]. Fig. 4(b) displays wind speed
4.3. Point prediction model validation
data from Eman wind farm, where the wind speeds are mainly distrib­
uted in the range of [2.6, 5.245]. It can be observed that both wind farms
In this research, the feasibility and validity of the prediction model
have significant wind speed fluctuations, with different peak values at
are verified using a variable control method. First, the validity of the
different times, and there are large differences in peak values before and
selected LSTM model is verified by comparing the model with the BP
after, indicating strong randomness. This research selects one week’s
model, Elman model, LSSVM model, and RBF model. The specific pre­
wind speed data from both wind farms, and 144 sampling points on the
diction errors are shown in Fig. 7. The model parameter settings here are

Fig. 5. HADD components of Sotavento.

10
Y. Zhang et al. Renewable and Sustainable Energy Reviews 196 (2024) 114349

Fig. 6. HADD components of Eman.

Fig. 7. Prediction error of single models, (a) Prediction error of Sotavento, (b) Prediction error of Eman.

11
Y. Zhang et al. Renewable and Sustainable Energy Reviews 196 (2024) 114349

Fig. 8. Prediction results after using HADD, (a) Prediction results of Sotavento, (b) Prediction results of Eman.

based on extensive literature and experimental testing. Similarly, the HADD-denoised Eman wind farm data components were
From Fig. 7, it can be seen that the purple dashed line represents the input into different models for prediction, and the obtained prediction
prediction error of the RBF model, the yellow curve represents the curves were displayed in Fig. 8(b).
prediction error of the Elman model. While the blue curve represents the From Fig. 8, it can be observed that the model prediction accuracy is
prediction error of the BP model, and the orange solid line represents the significantly improved after being processed by the HADD method. All
prediction error of the LSTM model. Although the prediction effect of a of the prediction curves effectively reflect the changing trend of the
single model is poor and has obvious IMFtime lag. A single model can original wind speed data curve, and the prediction accuracy is greatly
achieve preliminary prediction of wind speed, and the LSTM and BP improved for wind speed abrupt changes compared to the single model.
models have better prediction performance. In the Eman wind farm, the Combining Figs. 8 and 9, and Table 3, it can be seen that the prediction
data fluctuates significantly and the LSTM demonstrates better predic­ error is greatly improved after the HADD decomposition. Taking Sota­
tion performance than the BP model. It shows the practical significance vento wind farm as an example, after adding HADD, the prediction re­
of choosing deep learning model as prediction model. sults demonstrate a significant decrease in MAE and RMSE. The RMSE
To verify the impact of mixed adaptive decomposition denoising value of HADD-BP decreased by 61.69% compared to the BP model, and
method on prediction accuracy, the original data of two wind farms are the MAE value of HADD-LSTM decreased by 61.08% compared to the
first decomposed by the mixed adaptive decomposition denoising LSTM model, with a decrease of 59.26% in MAPE and an increase of
method before prediction. Some components had complex fluctuations, 11.5% in regression coefficients. In Eman wind farm, the prediction set
but some reflected the main trend of the original wind speed. The partial fluctuates more obviously, resulting in larger errors in some peak-to-
components of Sotavento wind farm decomposition fluctuate relatively valley prediction values. However, the prediction results after adding
stable, the amplitude is small, but the signal frequency domain is high HADD show a significant improvement in error indicators. After
and the bandwidth is long. In combination with Tables 2 and it can be decomposition and denoising, the prediction results improved by over
seen that residual white noise exists in IMF1, IMF2, IMF4 and IMF5 of 30% in both MAE and RMSE for the RBF model. The RMSE value
Sotavento wind farm identified by HADD method, and Fig. 5 is obtained decreased from 0.8185 to 0.4238 and the R2 value increased from
after de-noising. The processed component sequences are input into the 0.6074 to 0.8948. Compared with the LSTM model, the RMSE value of
model for prediction, and the prediction results are shown in Fig. 8(a). HADD-LSTM decreased by 42.77% and the MAPE value decreased by

12
Y. Zhang et al. Renewable and Sustainable Energy Reviews 196 (2024) 114349

Fig. 9. Prediction error after using HADD, (a) Prediction error of Sotavento, (b) Prediction error of Eman.

Table 3
Prediction evaluation indicators of models.
Sotavento Eman
2
MAE (m/s) RMSE (m/s) MAPE R MAE (m/s) RMSE (m/s) MAPE R2

RBF 0.8448 1.2436 17.55% 0.5636 0.5973 0.8185 21.37% 0.6074

Elman 0.4869 0.6586 8.58% 0.8701 0.5014 0.6959 16.99% 0.7162
LSSVM 0.4825 0.672 8.34% 0.8647 0.5116 0.7035 16.93% 0.7099
BP 0.4649 0.6468 7.84% 0.8747 0.5039 0.6999 16.94% 0.7131
LSTM 0.4687 0.6617 7.8% 0.8693 0.4999 0.6935 16.69% 0.718
HADD-RBF 0.1869 0.2465 3.33% 0.9818 0.3016 0.4238 9.68% 0.8948
HADD-Elman 0.2093 0.2875 3.94% 0.9753 0.3315 0.4464 10.22% 0.8834
HADD-LSSVM 0.196 0.2644 3.55% 0.9791 0.3121 0.4149 10.44% 0.8998
HADD-BP 0.194 0.2478 3.31% 0.9816 0.2954 0.4068 9.77% 0.9033
HADD-LSTM 0.1824 0.2414 3.19% 0.9825 0.2888 0.3969 9.42% 0.9077
SOA-LSTM 0.4608 0.6504 7.77% 0.8733 0.4867 0.6839 15.65% 0.7273
ISOA-LSTM 0.4449 0.6238 7.2% 0.8836 0.4205 0.5538 14.07% 0.8203
HADD-SOA-LSTM 0.181 0.2343 3.13% 0.9836 0.2757 0.3693 9.42% 0.9204
HADD-ISOA-LSTM 0.1682 0.2199 2.87% 0.9855 0.2402 0.3052 8.01% 0.9454

43.56%. Based on the error evaluation indicators, further optimization 4.4. Verify the effectiveness of the ISOA
and improvement based on HADD-LSTM can have a positive effect on
improving wind speed prediction accuracy. The SOA is an intelligent swarm optimization search algorithm that
exhibits excellent performance in model parameter optimization. To
further improve the search ability of the SOA, accelerate the conver­
gence speed, and increase the population diversity of the algorithm, this

13
Y. Zhang et al. Renewable and Sustainable Energy Reviews 196 (2024) 114349

Fig. 10. Fitness curve of CEC2022 test functions.

work proposes an ISOA that uses chaotic mapping to increase the initial predictive performance of both SOA-LSTM and ISOA-LSTM models
population diversity. The cosine function, Cauchy mutation, and improved compared to the LSTM model, with RMSE values decreased by
perturbation operator are introduced to enhance the global search 1.71% and 5.73%, respectively. Compared to the SOA-LSTM model, the
ability. ISOA-LSTM model reduced the predictive MAE by 0.0159 m/s, or
The initialization process of the SOA is improved by logistics-Sine 3.45%. The MAPE value of the HADD-ISOA-LSTM model was 2.87%,
hybrid chaotic mapping system. Using chaotic sequences to generate which decreased by 8.31% compared to the HADD-SOA-LSTM model. In
random numbers, the random initialization in the solution space is more the Eman wind farm, the RMSE value of ISOA-LSTM decreased by
evenly distributed, making the population reach the relatively optimal 19.02% compared to SOA-LSTM. The MAPE value of HADD-ISOA-LSTM
solution more quickly. was 8.01%, which is a 14.97% decrease compared to HADD-SOA-LSTM.
The cosine function is used to transform linear transformations into In the prediction results of the two wind farms, the HADD-ISOA-LSTM
non-linear transformations to change the movement behavior of the model incorporating the HADD algorithm and ISOA achieved an R2
seagull. Cauchy mutation and perturbation operators are added. When value of over 94%, indicating good fitting and high prediction accuracy.
the algorithm iteration falls into a local optimum, these operators can Finally, this research compared and calculated the errors of single
increase the possibility of the algorithm jumping out of the local opti­ models and combined prediction models, and presented the error in­
mum and improve the convergence ability of the algorithm. dicators in Table 3. The proposed HADD algorithm and ISOA have
To illustrate the superiority of the improved algorithm, this work improved the accuracy of wind speed prediction, indicating that the
introduces two optimization algorithms with superior performance, improvement strategy proposed has a positive effect on improving wind
WOA and OOA, which were proposed in 2016 and 2023 respectively. speed prediction.
The CEC2022 test function is used to compare traditional SOA and ISAO,
and the effectiveness of ISOA in improving search efficiency and 5. Wind speed interval prediction
avoiding local optimality is verified. Compared with WOA and OOA
algorithms, ISOA’s performance is superior. The red curve shows the The innovation and optimization of point prediction models for wind
change in the fitness value of the ISOA function. speed can improve the accuracy of prediction to a certain extent.
From Fig. 10, it can be observed that compared to SOA, in the early However, it is difficult to evaluate the fluctuation of wind speed based
iterations, ISOA with the inclusion of a hybrid chaotic mapping exhibits on point predictions alone. To enhance the reliability of wind speed
lower fitness values and a faster declining trend. In the later iterations, prediction, this work proposes a wind speed interval prediction method
ISOA considering the Cauchy mutation and perturbation operator based on data augmentation techniques and kernel density estimation.
demonstrates a better ability to escape local optima and tends towards
the global optimum, showcasing stronger convergence capability.
Compared to WOA and OOA, ISOA does not show a significant advan­ 5.1. Wind speed range forecast based on SMOTE and KDE
tage in the early iterations. However, in the later iterations, ISOA dis­
plays a higher ability to escape local optima. This advantage is more Firstly, the error sequence of training set and test set is obtained
pronounced in the relatively complex F3 and F4. The performance of using the wind speed point prediction model proposed and wind farm
ISOA indicates that the inclusion of a hybrid improvement strategy leads data. Then, use SMOTE to build most classes, the train-set error
to a notable enhancement in both convergence capability and the ability sequence, and a few classes, the test-set error sequence.
to escape local optima. In the test set error, new error samples were generated based on
Using ISOA to optimize the parameters of the LSTM model and neighboring samples. The frequency histogram is drawn based on the
combining with the HADD algorithm, the predicted results are shown in empirical cumulative distribution function. The horizontal axis of the
Fig. 11. From Fig. 11 and Tables 3 and it can be seen that for Sotavento graph represents the value of the prediction error from small to large,
wind farm as an example, after adding optimization algorithm, the and the vertical axis represents the corresponding empirical distribution
function value. Fig. 12 shows the original error distribution of the test

14
Y. Zhang et al. Renewable and Sustainable Energy Reviews 196 (2024) 114349

Fig. 11. Prediction results after using ISOA, (a) Prediction results of Sotavento, (b) Prediction results of Eman.

set for Sotavento and Eman wind farms and the error distribution ob­ 5.2. Verification of wind speed interval prediction effect
tained after SMOTE data augmentation. After augmentation, the number
of error samples increased fourfold, but the distribution remained From Fig. 12, it can be seen that fitting two groups of errors with
similar. Then, the probability density function of the error sequence was normal distribution has poor results and the Kolmogorov-Smirnov (KS)
obtained using KDE. The distribution curve estimated by kernel density test result for the distribution of the two error sequences is 1, so the
estimation is shown as the yellow line in Fig. 12(a) and (b), and it can be normal distribution method is not used for fitting. Using the SMOTE
observed that KDE is effective in fitting the error distribution. method to process the error data can avoid overfitting that may be
Different wind speed prediction intervals were generated based on caused by the Bootstrap method, generate effective new error data, and
different confidence levels. The results are shown in Fig. 13, indicating increase data volume without changing the data structure. Intuitively,
that the wind speed intervals can effectively provide the fluctuation the SMOTE-generated error sequence is similar in distribution to the
range of the wind speed, and most of the actual data can be covered by original error sequence, and the fit of the normal distribution is poor.
the intervals. To better compare the interval prediction performance, Therefore, this work uses KDE without prior knowledge of the sequence
this study also used different comparison models and error evaluation distribution characteristics and uses Gaussian function as the kernel
indicators for validation. function. The yellow curve representing the kernel density probability
distribution can fit the error distribution well.
This research uses the quantile interval method to establish wind

15
Y. Zhang et al. Renewable and Sustainable Energy Reviews 196 (2024) 114349

Fig. 12. Original error distribution and SMOTE error distribution, (a) Error distribution of Sotavento, (b) Error distribution of Eman.

speed intervals as a comparison model, and uses the CWC and AIS as comprehensively considering coverage rate and average interval width,
error evaluation indicators for interval prediction. The results are shown the proposed model has better predictive performance. Similarly, ac­
in the Table 4. cording to AIS values, at the same confidence level, the AIS value of the
Contrast the range given based on SMOTE and KDE with the quartile proposed interval prediction model is larger than that of the quantile
range. It can be seen that, due to strong wind speed fluctuation and high interval prediction, which also proves that the proposed interval pre­
randomness, the prediction effect of wind speed interval is poor when diction model can balance accuracy and reliability, effectively quantify
the wind speed changes greatly and the peak-peak value and peak-valley wind speed fluctuations.
part. However, under the 98% confidence level, the interval coverage of
both wind farms can reach more than 93%, and the average interval 6. Conclusion
width is small. Under the same confidence level, the average width of
wind speed interval proposed is smaller than the quartile interval. To Accurate wind speed prediction is the basis of the development of
comprehensively consider the coverage rate and interval width of in­ wind power industry, and it is the necessary technical guarantee to
terval prediction, in terms of CWC, the proposed interval prediction reduce power generation cost and avoid resource waste. To improve the
model in this research has a smaller CWC value compared to the interval safety and reliability of the wind power forecasting system effectively,
prediction results of two wind farms. This means that on the basis of this study proposes a hybrid prediction system that combines data

16
Y. Zhang et al. Renewable and Sustainable Energy Reviews 196 (2024) 114349

Fig. 13. Wind speed prediction intervals under different confidence levels, (a) Prediction intervals of Sotavento, (b) Prediction intervals of Eman.

(1) By proposing the HADD method, data decomposition can be

Table 4 adaptively performed to identify residual noise sequences and
Evaluation indicators for interval prediction. denoise noise components. This effectively improves data quality
Data set Confidence Interval prediction CWC AIS and avoids the occurrence of over-decomposition phenomena.
(2) The ISOA algorithm obtained through the hybrid improvement
Sotavento 0.9 Quantile 1.4602 − 1.3622
SMOTE-Kernel 1.1894 − 1.1286 strategy has stronger global search ability and faster convergence
0.98 Quantile 2.3324 − 2.2867 speed. The ISOA is used to optimize the parameters of the LSTM
SMOTE-Kernel 1.8996 − 1.8451 model, which greatly improves the accuracy and stability of the
Eman 0.9 Quantile 1.9905 − 1.8375 prediction model.
SMOTE-Kernel 1.6013 1.491
(3) Using the SMOTE data augmentation algorithm to process error
−
0.98 Quantile 2.9548 − 2.9017
SMOTE-Kernel 2.4405 − 2.3686 data can avoid overfitting problems. Combined with KDE
method, interval estimation of wind speed sequences is achieved.
The interval prediction model proposed takes into account both
processing, model optimization, deep learning, and interval prediction. the accuracy and reliability of the prediction, making it an
Firstly, the HADD algorithm is used to adaptively decompose and effective interval prediction method.
denoise the original wind speed data, improving data quality and (4) The point prediction model established in this work can accu­
reducing prediction complexity. Secondly, the search capability of SOA rately predict the change of wind speed in the next period of time,
is enhanced using methods such as chaotic mapping, Cauchy mutation, provide technical support for power system dispatching and
and perturbation operator, and the parameters of the LSTM are searched improve the safety. The interval prediction model established in
and optimized using ISOA. Finally, wind speed interval prediction is this research takes into account the accuracy and reliability of
constructed by combining data augmentation techniques and the KDE wind power fluctuations, and can improve the proportion of clean
method. The accuracy and reliability of the system are verified using energy.
historical data from the Sotavento wind farm in Spain and the Eman
wind farm in China. The main conclusions are as follows: The hybrid prediction model proposed in this study can be applied

17
Y. Zhang et al. Renewable and Sustainable Energy Reviews 196 (2024) 114349

not only to wind farms, but also to photovoltaic power plants. The [10] Jiang Y, Liu S, Zhao N, Xin J, Wu B. Short-term wind speed prediction using time
varying filter-based empirical mode decomposition and group method of data
research improves the existing algorithm to some extent, but it still
handling-based hybrid model. Energy Convers Manag 2020;220:113076. https://
needs further improvement. This research assumes that the future wind doi.org/10.1016/j.enconman.2020.113076.
speed is the same as when the model is train, and the possibility of [11] Li Y, Peng T, Zhang C, Sun W, Hua L, Ji C, Shahzad N. Multi-step ahead wind speed
sudden changes is excluded, so the prediction has limitations. In the forecasting approach coupling maximal overlap discrete wavelet transform,
improved grey wolf optimization algorithm and long short-term memory. Renew
future, it is necessary to consider the possibility of sudden change of Energy 2022;196:1115–26. https://doi.org/10.1016/j.renene.2022.07.016.
wind speed and analyze the relevant influencing factors combined with [12] Xiao Y, Wang X, Wang J, Zhang H. An adaptive decomposition and ensemble
statistical methods. Interval prediction needs to explore new prediction model for short-term air pollutant concentration forecast using ICEEMDAN-ICA.
Technol Forecast Soc Change 2021;166:120655. https://doi.org/10.1016/j.
models to further expand the coverage range of wind speed intervals on techfore.2021.120655.
the premise of ensuring accuracy. [13] Zhang Y, Chen B, Pan G, Zhao Y. A novel hybrid model based on VMD-WT and
PCA-BP-RBF neural network for short-term wind speed forecasting. Energy Convers
Manag 2019;195:180–97. https://doi.org/10.1016/j.enconman.2019.05.005.
CRediT authorship contribution statement [14] Zhou Y, Yan S, Ren Y, Liu S. Rolling bearing fault diagnosis using transient-
extracting transform and linear discriminant analysis. Measurement 2021;178:
109298. https://doi.org/10.1016/j.measurement.2021.109298.
Yagang Zhang: All authors participated in the research discussion
[15] Kaya Y, Kuncan F. A hybrid model for classification of medical data set based on
and provided their views on the results and writing. The division of labor factor analysis and extreme learning machine: FA + ELM. Biomed Signal Process
is as follows. Xue Kong: Conceptualization, Methodology, Supervision, Control 2022;78:104023. https://doi.org/10.1016/j.bspc.2022.104023.
and, Writing – original draft, preparation. Jingchao Wang: Visualiza­ [16] Ma H, Wang T, Li Y, Meng Y. A time Picking method for microseismic data based
on LLE and improved PSO clustering algorithm. Geosci Rem Sens Lett IEEE 2018;
tion, Software, and, Methodology. Hui Wang: Formal analysis, Inves­ 15:1677–81. https://doi.org/10.1109/LGRS.2018.2854834.
tigation. Xiaodan Cheng: Data Curation, Validation. [17] Sun S, Qiao H, Wei Y, Wang S. A new dynamic integrated approach for wind speed
forecasting. Appl Energy 2017;197:151–62. https://doi.org/10.1016/j.
apenergy.2017.04.008.
[18] Lv S-X, Wang L. Multivariate wind speed forecasting based on multi-objective
Declaration of competing interest feature selection approach and hybrid deep learning model. Energy 2023;263:
126100. https://doi.org/10.1016/j.energy.2022.126100.
The authors declare that they have no known competing financial [19] Ren Z, Wu L, Pang Y, Zhang W, Yang R. Strategies for effectively harvesting wind
energy based on triboelectric nanogenerators. Nano Energy 2022;100:107522.
interests or personal relationships that could have appeared to influence https://doi.org/10.1016/j.nanoen.2022.107522.
the work reported in this paper. [20] Abedinia O, Lotfi M, Bagheri M, Sobhani B, Shafie-khah M, Catalão JPS. Improved
EMD-based complex prediction model for wind power forecasting. IEEE Trans
Sustain Energy 2020;11:2790–802. https://doi.org/10.1109/TSTE.2020.2976038.
Data availability [21] He Y, Wang Y. Short-term wind power prediction based on EEMD–LASSO–QRNN
model. Appl Soft Comput 2021;105:107288. https://doi.org/10.1016/j.
Data will be made available on request. asoc.2021.107288.
[22] Tian Z, Chen H. Multi-step short-term wind speed prediction based on integrated
multi-model fusion. Appl Energy 2021;298:117248. https://doi.org/10.1016/j.
Acknowledgements apenergy.2021.117248.
[23] Zhang Y, Chen Y, Qi Z, Wang S, Zhang J, Wang F. A hybrid forecasting system with
complexity identification and improved optimization for short-term wind speed
The authors thank Prof. Paul Leahy and the anonymous referees for prediction. Energy Convers Manag 2022;270:116221. https://doi.org/10.1016/j.
the thoughtful and constructive suggestions that led to a considerable enconman.2022.116221.
improvement of the research. This research was supported partly by the [24] Zhang C, Hua L, Ji C, Shahzad Nazir M, Peng T. An evolutionary robust solar
radiation prediction model based on WT-CEEMDAN and IASO-optimized outlier
National Natural Science Foundation of China (U22B6006), the Natural robust extreme learning machine. Appl Energy 2022;322:119518. https://doi.org/
Science Foundation of Hebei Province (G2024502006) and the Funda­ 10.1016/j.apenergy.2022.119518.
mental Research Funds for the Central Universities(2023JC006). [25] Jurasz J. Modeling and forecasting energy flow between national power grid and a
solar–wind–pumped-hydroelectricity (PV–WT–PSH) energy source. Energy
Convers Manag 2017;136:382–94. https://doi.org/10.1016/j.
References enconman.2017.01.032.
[26] Zhang W, Qu Z, Zhang K, Mao W, Ma Y, Fan X. A combined model based on
CEEMDAN and modified flower pollination algorithm for wind speed forecasting.
[1] Al-Nimr M, Khashan SA, Al-Oqla H. Novel techniques to enhance the performance
Energy Convers Manag 2017;136:439–51. https://doi.org/10.1016/j.
of Stirling engines integrated with solar systems. Renew Energy 2023;202:
enconman.2017.01.022.
894–906. https://doi.org/10.1016/j.renene.2022.11.086.
[27] Liu Z, Liu H. A novel hybrid model based on GA-VMD, sample entropy
[2] Investment needs of USD 35 trillion by 2030 for successful energy transition. 2023.
reconstruction and BiLSTM for wind speed prediction. Measurement 2023;222:
https://www.irena.org/News/pressreleases/2023/Mar/Investment-Needs-of-USD-
113643. https://doi.org/10.1016/j.measurement.2023.113643.
35-trillion-by-2030-for-Successful-Energy-Transition. [Accessed 31 March 2023].
[28] Li M, Li Y, Choi SS. Dispatch planning of a wide-area wind power-energy storage
[3] Ye L, Li Y, Pei M, Zhao Y, Li Z, Lu P. A novel integrated method for short-term wind
scheme based on ensemble empirical mode decomposition technique. IEEE Trans
power forecasting based on fluctuation clustering and history matching. Appl
Sustain Energy 2021;12:1275–88. https://doi.org/10.1109/TSTE.2020.3042385.
Energy 2022;327:120131. https://doi.org/10.1016/j.apenergy.2022.120131.
[29] Liu Z, Zhang L, Carrasco J. Vibration analysis for large-scale wind turbine blade
[4] Che J, Yuan F, Deng D, Jiang Z. Ultra-short-term probabilistic wind power
bearing fault detection with an empirical wavelet thresholding method. Renew
forecasting with spatial-temporal multi-scale features and K-FSDW based weight.
Energy 2020;146:99–110. https://doi.org/10.1016/j.renene.2019.06.094.
Appl Energy 2023;331:120479. https://doi.org/10.1016/j.apenergy.2022.120479.
[30] Yakoub G, Mathew S, Leal J. Intelligent estimation of wind farm performance with
[5] Wang J, Qian Y, Zhang L, Wang K, Zhang H. A novel wind power forecasting
direct and indirect ‘point’ forecasting approaches integrating several NWP models.
system integrating time series refining, nonlinear multi-objective optimized deep
Energy 2023;263:125893. https://doi.org/10.1016/j.energy.2022.125893.
learning and linear error correction. Energy Convers Manag 2024;299:117818.
[31] Liu X, Lin Z, Feng Z. Short-term offshore wind speed forecast by seasonal ARIMA - a
https://doi.org/10.1016/j.enconman.2023.117818.
comparison against GRU and LSTM. Energy 2021;227:120492. https://doi.org/
[6] Majad M, Adeel M, Muhammad U, Ling Q. Hybrid forecasting models for wind-PV
10.1016/j.energy.2021.120492.
systems in diverse geographical locations: performance and power potential
[32] Wu F, Jing R, Zhang X-P, Wang F, Bao Y. A combined method of improved grey BP
analysis. Energy Convers Manag 2023;287:117080. https://doi.org/10.1016/j.
neural network and MEEMD-ARIMA for day-ahead wave energy forecast. IEEE
enconman.2023.117080.
Trans Sustain Energy 2021;12:2404–12. https://doi.org/10.1109/
[7] Jin H, Li Y, Wang B, Yang B, Jin H, Cao Y. Adaptive forecasting of wind power
TSTE.2021.3096554.
based on selective ensemble of offline global and online local learning. Energy
[33] Hua L, Zhang C, Peng T, Ji C, Shahzad Nazir M. Integrated framework of extreme
Convers Manag 2022;271:116296. https://doi.org/10.1016/j.
learning machine (ELM) based on improved atom search optimization for short-
enconman.2022.116296.
term wind speed prediction. Energy Convers Manag 2022;252:115102. https://doi.
[8] Fan H, Zhen Z, Liu N, Sun Y, Chang X, Li Y, Wang F, Mi Z. Fluctuation pattern
org/10.1016/j.enconman.2021.115102.
recognition based ultra-short-term wind power probabilistic forecasting method.
[34] Khodayar M, Wang J. Spatio-Temporal graph deep neural network for short-term
Energy 2023;266:126420. https://doi.org/10.1016/j.energy.2022.126420.
wind speed forecasting. IEEE Trans Sustain Energy 2019;10:670–81. https://doi.
[9] Shahid F, Wood A, Nisar H, Zameer A, Eger S. Short-term multi-hour ahead
org/10.1109/TSTE.2018.2844102.
country-wide wind power prediction for Germany using gated recurrent unit deep
learning. Renew Sustain Energy Rev 2022;167:112700. https://doi.org/10.1016/j.
rser.2022.112700.

18
Y. Zhang et al. Renewable and Sustainable Energy Reviews 196 (2024) 114349

[35] Liu L, Liu J, Ye Y, Liu H, Chen K, Li D, Dong X, Sun M. Ultra-short-term wind power [53] Pan Y, Hong R, Chen J, Wu W. A hybrid DBN-SOM-PF-based prognostic approach
forecasting based on deep Bayesian model with uncertainty. Renew Energy 2023; of remaining useful life for wind turbine gearbox. Renew Energy 2020;152:138–54.
205:598–607. https://doi.org/10.1016/j.renene.2023.01.038. https://doi.org/10.1016/j.renene.2020.01.042.
[36] Nguyen THT, Phan QB. Hourly day ahead wind speed forecasting based on a hybrid [54] Coulson M, Ferles C, Winberg S, Naidoo KJ. Growing hierarchical self-organising
model of EEMD, CNN-Bi-LSTM embedded with GA optimization. Energy Rep 2022; representation map (GHSORM). Inf Sci 2023;642:119121. https://doi.org/
8:53–60. https://doi.org/10.1016/j.egyr.2022.05.110. 10.1016/j.ins.2023.119121.
[37] Wu Y, Zhang P. Online monitoring for power cables in DFIG-based wind farms [55] Guo J, Si Z, Xiang J. A compound fault diagnosis method of rolling bearing based
using high-frequency resonance analysis. IEEE Trans Sustain Energy 2022;13: on wavelet scattering transform and improved soft threshold denoising algorithm.
378–90. https://doi.org/10.1109/TSTE.2021.3113017. Measurement 2022;196:111276. https://doi.org/10.1016/j.
[38] Yu M, Niu D, Gao T, Wang K, Sun L, Li M, Xu X. A novel framework for ultra-short- measurement.2022.111276.
term interval wind power prediction based on RF-WOA-VMD and BiGRU optimized [56] Liu S, Ouyang Z, Chen G, Zhou X, Zou Z. Black-box modeling of ship maneuvering
by the attention mechanism. Energy 2023;269:126738. https://doi.org/10.1016/j. motion based on Gaussian process regression with wavelet threshold denoising.
energy.2023.126738. Ocean Eng 2023;271:113765. https://doi.org/10.1016/j.oceaneng.2023.113765.
[39] Wu Y, Meng X, Zhang J, He Y, Romo J, Dong Y, Lu D. Effective LSTMs with [57] Justen L, Ramlau R. A general framework for soft-shrinkage with applications to
seasonal-trend decomposition and adaptive learning and niching-based blind deconvolution and wavelet denoising. Appl Comput Harmon Anal 2009;26:
backtracking search algorithm for time series forecasting. Expert Syst Appl 2024; 43–63. https://doi.org/10.1016/j.acha.2008.02.002.
236:121202. https://doi.org/10.1016/j.eswa.2023.121202. [58] Xian S, Chen K, Cheng Y. Improved seagull optimization algorithm of partition and
[40] Li L, Zheng S, Tseng M, Liu Y. Performance assessment of combined cooling, XGBoost of prediction for fuzzy time series forecasting of COVID-19 daily
heating and power system operation strategy based on multi-objective seagull confirmed. Adv Eng Software 2022;173:103212. https://doi.org/10.1016/j.
optimization algorithm. Energy Convers Manag 2021;244:114443. https://doi. advengsoft.2022.103212.
org/10.1016/j.enconman.2021.114443. [59] Long W, Jiao J, Liang X, Xu M, Tang M, Cai S. Parameters estimation of
[41] Zhang Y, Zhao Y, Shen X, Zhang J. A comprehensive wind speed prediction system photovoltaic models using a novel hybrid seagull optimization algorithm. Energy
based on Monte Carlo and artificial intelligence algorithms. Appl Energy 2022;305: 2022;249:123760. https://doi.org/10.1016/j.energy.2022.123760.
117815. https://doi.org/10.1016/j.apenergy.2021.117815. [60] Liu L, Zhao D, Yu F, Heidari A, Li C, Ouyang J, Chen H, Mafarja M, Turabieh H,
[42] Jiang Y, Huang G, Yang Q, Yan Z, Zhang C. A novel probabilistic wind speed Pan J. Ant colony optimization with Cauchy and greedy Levy mutations for
prediction approach using real time refined variational model decomposition and multilevel COVID 19 X-ray image segmentation. Comput Biol Med 2021;136:
conditional kernel density estimation. Energy Convers Manag 2019;185:758–73. 104609. https://doi.org/10.1016/j.compbiomed.2021.104609.
https://doi.org/10.1016/j.enconman.2019.02.028. [61] Liu H, Song W, Niu Y, Zio E. A generalized cauchy method for remaining useful life
[43] Chen J, Liu H, Chen C, Duan Z. Wind speed forecasting using multi-scale feature prediction of wind turbine gearboxes. Mech Syst Signal Process 2021;153:107471.
adaptive extraction ensemble model with error regression correction. Expert Syst https://doi.org/10.1016/j.ymssp.2020.107471.
Appl 2022;207:117358. https://doi.org/10.1016/j.eswa.2022.117358. [62] Hua Z, Zhou Y, Pun C, Chen C. 2D Sine Logistic modulation map for image
[44] Ye L, Dai B, Li Z, Pei M, Zhao Y, Lu P. An ensemble method for short-term wind encryption. Inf Sci 2015;297:80–94. https://doi.org/10.1016/j.ins.2014.11.018.
power prediction considering error correction strategy. Appl Energy 2022;322: [63] Lai Q, Hu G, Erkan U, Toktas A. High-efficiency medical image encryption method
119475. https://doi.org/10.1016/j.apenergy.2022.119475. based on 2D Logistic-Gaussian hyperchaotic map. Appl Math Comput 2023;442:
[45] Zhang Y, Zhang J, Yu L, Pan Z, Feng C, Sun Y, Wang F. A short-term wind energy 127738. https://doi.org/10.1016/j.amc.2022.127738.
hybrid optimal prediction system with denoising and novel error correction [64] Wang J, Jiang W, Xu H, Wu X, Kim J. Image encryption based on Logistic-Sine self-
technique. Energy 2022;254:124378. https://doi.org/10.1016/j. embedding chaotic sequence. Optik 2022;271:170075. https://doi.org/10.1016/j.
energy.2022.124378. ijleo.2022.170075.
[46] Han L, Yang G, Yang X, Song X, Xu B, Li Z, Wu J, Yang H, Wu J. An explainable [65] Ko M-S, Lee K, Kim J-K, Hong CW, Dong ZY, Hur K. Deep concatenated residual
XGBoost model improved by SMOTE-ENN technique for maize lodging detection network with bidirectional LSTM for one-hour-ahead wind power forecasting. IEEE
based on multi-source unmanned aerial vehicle images. Comput Electron Agric Trans Sustain Energy 2021;12:1321–35. https://doi.org/10.1109/
2022;194:106804. https://doi.org/10.1016/j.compag.2022.106804. TSTE.2020.3043884.
[47] Gao Y, Xu X, Yan Z, Shahidehpour M. Gaussian mixture model for multivariate [66] Zheng L, Lu W, Zhou Q. Weather image-based short-term dense wind speed
wind power based on kernel density estimation and component number reduction. forecast with a ConvLSTM-LSTM deep learning model. Build Environ 2023;239:
IEEE Trans Sustain Energy 2022;13:1853–6. https://doi.org/10.1109/ 110446. https://doi.org/10.1016/j.buildenv.2023.110446.
TSTE.2022.3159391. [67] Chen Y, Dong Z, Wang Y, Su J, Han Z, Zhou D, Zhang K, Zhao Y, Bao Y. Short-term
[48] Ahn E, Hur J. A short-term forecasting of wind power outputs using the enhanced wind speed predicting framework based on EEMD-GA-LSTM method under large
wavelet transform and arimax techniques. Renew Energy 2023;212:394–402. scaled wind history. Energy Convers Manag 2021;227:113559. https://doi.org/
https://doi.org/10.1016/j.renene.2023.05.048. 10.1016/j.enconman.2020.113559.
[49] Cardarilli G, Di Nunzio L, Fazzolari R, Re M, Spanò S. AW-SOM, an algorithm for [68] Yang Y, Zhang Y, Cheng Y, Lei Z, Gao X, Huang Y, Ma Y. Using one-dimensional
high-speed learning in hardware self-organizing maps. IEEE Transactions on convolutional neural networks and data augmentation to predict thermal
Circuits and Systems II: Express Briefs 2020;67:380–4. https://doi.org/10.1109/ production in geothermal fields. J Clean Prod 2023;387:135879. https://doi.org/
TCSII.2019.2909117. 10.1016/j.jclepro.2023.135879.
[50] Li K, Huang W, Hu G, Li J. Ultra-short term power load forecasting based on [69] Imakura A, Kihira M, Okada Y, Sakurai T. Another use of SMOTE for interpretable
CEEMDAN-SE and LSTM neural network. Energy Build 2023;279:112666. https:// data collaboration analysis. Expert Syst Appl 2023;228:120385. https://doi.org/
doi.org/10.1016/j.enbuild.2022.112666. 10.1016/j.eswa.2023.120385.
[51] Bommidi B, Teeparthi K, Kosana V. Hybrid wind speed forecasting using [70] Du B, Huang S, Guo J, Tang H, Wang L, Zhou S. Interval forecasting for urban
ICEEMDAN and transformer model with novel loss function. Energy 2023;265: water demand using PSO optimized KDE distribution and LSTM neural networks.
126383. https://doi.org/10.1016/j.energy.2022.126383. Appl Soft Comput 2022;122:108875. https://doi.org/10.1016/j.
[52] Emeksiz C, Tan M. Wind speed estimation using novelty hybrid adaptive estimation asoc.2022.108875.
model based on decomposition and deep learning methods (ICEEMDAN-CNN).
Energy 2022;249:123785. https://doi.org/10.1016/j.energy.2022.123785.

19