Download as pdf or txt
Download as pdf or txt
You are on page 1of 16

Expert Systems With Applications 241 (2024) 122477

Contents lists available at ScienceDirect

Expert Systems With Applications


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

Novel wind-speed prediction system based on dimensionality reduction and


nonlinear weighting strategy for point-interval prediction
Xinyu Wang a, Jianzhou Wang a, *, Xinsong Niu b, Chunying Wu c
a
School of Statistics, Dongbei University of Finance and Economics, Dalian 116025, China
b
Academy of Mathematics and Systems Science, Chinese Academy of Sciences, Beijing 100190, China
c
School of Management Science and Engineering, Shandong University of Finance and Economics, Jinan, Shandong 250014, China

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

Keywords: In the context of today’s energy shortage, wind energy plays a crucial role as one of the most widely used
Combined model renewable energy sources. However, in order to fully utilize the potential of wind energy, it is necessary to build
Dimensionality reduction a precise and reliable wind-speed prediction system. Conventional wind-speed prediction models frequently rely
Multi-objective optimization algorithm
on a solitary model for forecasting, and sometimes these models cannot accurately fit the nonlinear character­
Neural networks
Wind-speed prediction
istics of the data, resulting in poor prediction performance. Therefore, to address this deficiency, a new wind-
speed point-interval prediction system is proposed in this study. The fuzzy information granulation technol­
ogy is adopted in this system to reduce the dimension of data and solve the problem of redundancy in wind-speed
data. In addition, this paper uses the multi-objective dragonfly algorithm and combines sub-models using the
multi-nonlinear weight strategy to address the issue of inadequate precision exhibited by an individual predic­
tion model. In order to verify the effectiveness of this wind-speed prediction system, the Penglai wind farm in
Shandong Province is used as the dataset to construct the model, and several reference models are selected to
conduct point prediction and interval prediction comparison experiments respectively. The experimental results
show that the combined prediction system has the best prediction performance, with MAPE values of 14.16%,
15.02% and 15.74%, respectively. Compared with other benchmark models, the average improvement rates for
the multi-step prediction were 14.97%, 16.48% and 19.11%, respectively.

1. Introduction delivering accurate and consistent predictions is critical for effective


dispatching and the safe operation of wind power installations (Akçay &
The Industrial Revolution brought about the rapid development of Filik, 2017).
the global economy, but it also led to an ongoing crisis of environmental In recent years, relevant domestic and international academics have
deterioration and the decreasing supply of fossil fuels. These factors recently investigated and analyzed wind speed forecasting techniques
have adversely affected the development of the world economy. To in-depth. By combing a large number of literatures, four main groups
address this dilemma, the development and use of green energy have can be used to broadly classify wind speed forecasting systems: (i)
been widely considered as successful solutions, attracting the interest of physical models, (ii) statistical models, (iii) artificial intelligence (AI)
countries all over the world. Wind energy, regarded as a form of models, and, and (iv) combined models (Wang , Zou, Liu, Zhang, & Liu,
renewable energy, has found wide application due to its inexhaustible 2021e).
nature, environmentally friendly attributes, and sustainable character­ A type of physical model known as the numerical weather prediction
istics (Yang, Hao, & Hao, 2023; Zhao et al., 2016a). According to the (NWP) has been extensively studied by numerous academics (Li, Wang,
World Wind Energy Association (WWEA), as of the end of 2020, the Zhang, & Li, 2022a). The model comprehensively considers physical
global wind capacity reached 744 GW, with an additional 93 GW of new characteristics such as terrain, density and temperature, and it is a useful
wind turbines installed. However, the fluctuation and volatility of wind technique for forecasting medium- and long-term wind speed. For
velocity exert a significant influence on the security and reliability of instance, Dong et al. (2016) established a model utilizing clustering
wind power generation (Liu, Wang, & Zhen, 2020). Therefore, analysis of NWP information. Sı̄le et al. (2014) verified that NWP model

* Corresponding author.
E-mail address: wangjz@dufe.edu.cn (J. Wang).

https://doi.org/10.1016/j.eswa.2023.122477
Received 6 September 2023; Received in revised form 26 October 2023; Accepted 4 November 2023
Available online 18 November 2023
0957-4174/© 2023 Elsevier Ltd. All rights reserved.
X. Wang et al. Expert Systems With Applications 241 (2024) 122477

is the main source of wind speed prediction in weather forecast, and that assessment. Ref. (Li, Zhu, Yang, & Li, 2019a) used multi-objective
it can also assess wind resources. Meanwhile, Cassola et al.(Cassola & Imperialist competitive algorithm (MOica) to establish a combined
Burlando, 2012) introduces the Kalman filter method to enhance the prediction model, which once again demonstrated the superior profi­
results of the NWP model. The outcomes indicate that long-term wind- ciency of the integrated algorithm in terms of precision and consistency.
speed forecasting accuracy is frequently satisfactory employing physical In addition, Zhang et al. (2022) and He et al. (2019) proposed to use
models (Zhu et al., 2021). However, the numerical model mechanism fuzzy information granulation technology (FIg) in the data preprocess­
requires extensive calculations to accurately describe environmental ing module of the prediction model to eliminate the uncertainty and
physical characteristics and wind farm characteristics, which can be volatility of the data. On the basis of deterministic prediction analysis,
limiting and difficult to achieve ideal results (Zhao et al., 2016b). the uncertainty analysis of wind speed series is developed and improved
Statistical models are based on probability theory and use specific by FIg, which provides a new feasible choice for prediction in this field.
mathematical techniques to show potential relationships among related Although the aforementioned models have achieved certain ad­
variables. In the realm of short-term forecasting, the statistical model vancements in the domain of forecasting wind speed, they still have
proves to be a more fitting choice in comparison to the physical some shortcomings. By comparing the characteristics of each model, as
approach. Two types of common statistical models are autoregressive shown in Table 1, several conclusions are attainable:
moving average (ARma) and autoregressive composite moving average
(ARima) (Nie, Liang, & Wang, 2021).Erdem et al.(Erdem & Shi, 2011) (a) The computational time required for physical models is long and
examined various approaches based on the ARma model and taking into therefore less efficient when dealing with large data sets.
account the wind speed and direction. Do et al. (Do, Lee, & Choi, 2016) Furthermore, physical models show limited efficacy in addressing
effectively employs the ARma model to accurately depict the random short-term forecasting needs, and their accuracy needs to be
data characteristics. However, it is important to note that these esti­ improved.
mation techniques count on data having a Gaussian distribution, which (b) Wind speed data are characterized by randomness and high
limits their predictive power in nonlinear models compared to linear volatility, while traditional statistical models have a relatively
time series models (Wang, Wang, Lu, & Zhao, 2021d; Qin, Li, & Du, simple structure and cannot capture the complex nonlinear fea­
2017). Therefore, when there are complex nonlinear relationships in the tures in the data, thus affecting the prediction results.
data, the traditional statistical model may not be able to obtain accurate (c) Artificial intelligence models can solve non-linear problems in
prediction results. data and excellent results have been achieved in data forecasting.
As artificial intelligence technology continues to evolve, researchers However, these models often suffer from overfitting and
have begun to apply it to wind speed prediction. Among these, a popular becoming trapped in local optima, which can impact prediction
technique is the artificial neural network (Ann), because it can accuracy in specific situations.
approximate any nonlinear function, so it satisfies the randomness of (d) Although combinatorial models have made some important
wind speed sequence (Mohammed & Al-Bazi, 2022; Wang, Li, Wang, & breakthroughs in the domain of forecasting wind-speed, most
Lu, 2021a). Many Anns have been utilized in wind-speed prediction, current prediction systems focus only on deterministic prediction
such as the Elman neural network (Liu, Tian, Liang, & Li, 2015a), and cannot analyze the uncertainty associated with a given wind
extreme learning machine (ELM)(Xing, Wang, Jiang, & Wang, 2023), speed prediction. Thus, although predictive models may provide
and general regression neural network (GRnn)(Wang , Niu, Liu, & accurate results in a given situation, they do not provide scholars
Zhang, 2020). In order to improve the prediction accuracy, the GA-Ann with information about the degree of confidence or uncertainty in
system was enhanced by Zhang et al. (2020) using variational mode the predicted results.
decomposition (Vmd), which has a strong self-learning capability. Guo
et al. (2011) presented a novel combined prediction method using a Considering the randomness of wind-speed data, the present study
backpropagation neural network (BPnn) with seasonal index to remove examined an innovative merged framework system that integrates data
seasonal effects from the wind-speed dataset. These artificial intelli­ preparation techniques and the multi-objective Dragonfly algorithm
gence models can be applied to different datasets and have higher pre­ (MOda) based on several prediction models (BPnn, ELM, Enn, ARima).
diction accuracy than other models. However, they may have some Below are outlined the benefits of the innovative combined
problems such as poor generalization ability, limited by local optimi­ system.
zation and overfitting.
Because of the randomness and irregularity of wind-speed time se­ (a) Advanced pre-processing strategies are applied to the raw
ries, a single model cannot precisely depict the distinct attributes of each data. The fuzzy information granulation preprocessing technique
series. Therefore, with the aim of enhancing the applicability of fore­ transforms the high-dimensional raw wind speed data into a low-
casting models, a combination system strategy is usually adopted (Gao, dimensional fuzzy subset, thus reducing the dimensionality of the
Wang, & Yang, 2022; Wang et al., 2023). The principle of this strategy is data and eliminating data redundancy for better extraction of key
to reduce the errors obtained from the training set provided by a single features in the data.
model and establish appropriate weight factors (Wang, Yang, Du, & Niu, (b) In order to solve the problem of data uncertainty, the com­
2018). For instance, Liu et al. (2015b) divided the original sequence into bined system proposed in this paper selects four prediction
several sub-layers using the FEemd algorithm and developed an Mlp models as sub-models, including neural networks and statistical
neural network and the Mea method to forecast the decomposed models. The system combines the strengths of individual sub-
sub-layers. According to the data, the hybrid FEemd-Mea-Mlp model models and effectively captures diverse features of wind-speed
appeared to have given the best prediction outcomes. Similarly, Qu et al. data, thereby enhancing its utility.
(2017) formulated a fusion flower pollination method using a bat search (c) In this paper, a weight optimization theory based on the
algorithm to determine the best weight coefficients (MOBSFPA). Niu dragonfly algorithm is applied. The algorithm adaptively ad­
and Wang (Niu & Wang, 2019) proposed a new model to further reduce justs the search parameters and strategies according to the
errors in data series. The prediction system first denoised the original different objective functions and constraints in the weight opti­
sequence and used a multi-objective grasshopper optimization algo­ mization problem, leading to quick convergence to the global
rithm (MOgoa) to select the best coefficients for the four Anns to optimal solution. Compared with other optimization algorithms,
improve the prediction accuracy. In combination forecasting research, it exhibits higher search efficiency and superior optimization
Nie (Nie et al., 2021) presented a bi-prediction framework that performance.
encompassed data preparation, integrated forecasting, and model

2
X. Wang et al. Expert Systems With Applications 241 (2024) 122477

Table 1
The characteristics of existing prediction methods.
Category Model Reference Results Advantage Disadvantage

Physical model NWP (Dong et al., A new model based on clustering NWP model can provide long-term It is necessary to simplify and approximate
2016; Sı̄le analysis of NWP information was wind speed prediction and provide the atmospheric physical process. There are
et al., 2014) proposed; the outcomes indicated high spatiotemporal resolution wind uncertainties in the observation data for
that it is effective for wind-speed speed prediction information. boundary conditions and validation.
prediction.
NWP models have been employed for
assessing the impact of wind speed on
model inaccuracies.

Statistical ARma (Do et al., The traditional ARma model is better Statistical models have higher Statistical models have less accurate with
model 2016; Erdem at predicting wind speed. prediction accuracy than physical non-linear data and long-term forecasting.
& Shi, 2011) models on short-term prediction. It is sensitive to outliers. At the same time,
ARima (Qin et al., The proposed method produces the processing ability is weak for non-
2017) better forecasts for red tide stationary data and missing values.
forecasting.

Artificial BPnn (Guo et al., A new hybrid wind-speed forecasting Artificial intelligence models suitable The model usually requires a large number
intelligence 2011) theory using BPnn with low mean for both linear and non-linear series of high-quality and accurate training data,
model absolute error. and the predictive capability exhibits and requires complex algorithm training
Enn (Liu et al., Strong predictive performance of commendable performance. and optimization processes. Meanwhile, the
2015a) standard Elman neural network. model prone to local optima and over-
GRnn (Wang et al., The GRnn model can accurately fitting.
2020) describe wind speed.

Combined FEemd- (Liu et al., The FEemd-Mea-Mlp method can Combined models can overcome the Extant combined models unable to provide
model Mea-Mlp 2015b) predict wind speed with high disadvantages of an individual model. uncertainty information with more stable
accuracy. forecasting independent of the dataset.
CEEmDan- (Niu & Wang, The CEEmDan-MOgoa model
MOgoa 2019) outperforms conventional predictive
models in terms of predictive
accuracy.
BMs- (Wang, Wang, The BMs-MNOta-Qr model can
MNOta-Qr & Li, 2021b) guarantee precise point prediction
effectiveness and uncertainty
information.

(d) The wind-speed forecasting system framework is extended to Tong, & Westerdahl, 2022b). The key to this technology is to capture the
construct a new point-interval prediction system, which core features of the data by splitting the time series data. By accurately
quantifies the uncertainty of data and simplifies the computa­ extracting important information from the data and shielding unim­
tional complexity. Additionally, a scientific and reasonable portant or noisy interference, FIg can better describe the characteristics
research experiment is designed, and several comparison exper­ and trends of the data, and provide strong support for further analysis
iments are conducted to validate the reliability of the novel and decision making (Akram, Nawaz, & Kahraman, 2023). The infor­
combined system. The results demonstrate that the combined mation particles can be expressed as Eq. (1)
system possesses an enhanced capability for accurately predicting
(1)
Δ
wind-speed. η= (̃
γ is H’) is λ, γ̃ ∈ P, H’ ⊆ P

The subsequent sections are structured as follows. In Section 2, a where ̃γ is a variable in the domain P; a fuzzy concept H′ is a fuzzy set
comprehensive discussion is provided on the data preparation technol­ with P as the domain, and λ is a probability of likelihood.
ogy utilized in the combined system, along with an in-depth explanation FIg is a grain of information represented as a fuzzy set; there are two
of the underlying principles. Section 3 exhibits the three sets of exper­ key processes involved in fuzzy granulation.
iments designed to assess the forecasting accuracy in comparison to To divide a window, the time series is split into a number of small
relevant methods. In Section 4, the applicability of the combination subsequences, each of which is considered as a granulation window. The
system is evaluated. Section 5 provides this paper final conclusions. length of the granulation window is the temporal span of each
subsequence.
2. Design of prediction system Fuzzification creates a fuzzy particle η = Ψ ̃ ′(̃γ) for each window of
generated data and extracts essential information according to the
The introduced combined system integrates fuzzy information affiliation function, which acts as a rule for the granulation of
granulation methods with four Anns based on multi-objective optimi­ information.
zation (MOp) theory. This study uses a triangular fuzzy particle with an affiliation function
expressed as Eq. (2)

2.1. Fuzzy information granulation ⎪ 0, ̃γ < ξ


(̃γ − ξ)/(ϖ − ξ), ξ⩽̃γ ⩽ϛ
̃ ′
Ψ (̃γ , ξ, ϖ, ϛ) = (2)
Information granulation plays a significant role in quantifying data ⎪
⎪ (ϛ − ̃γ)/(ϛ − ϖ), ϖ < ̃γ⩽ϛ

uncertainty, and is an important tool in current information processing. 0, ̃γ > ϛ
The technology was introduced by Professor L.A. Zadeh and focuses on
the overall information by decomposing it into smaller parts (Li, Tong, where ξ, ϖ, and ϛ denote the lower, median, and upper limits of each

3
X. Wang et al. Expert Systems With Applications 241 (2024) 122477

granulation window, respectively. are in motion most of the time; their positions change constantly. The
In the wind speed series, each data point in the wind speed series is spatial vectors for dragonflies are expressed as Z
̃′ =Z ̃′ .
̃′ + ΔZ
t+1 t t+1
mapped to the corresponding fuzzy set using the corresponding mem­ The method of performing Lévy flight for dragonfly position updates
bership function. Each data point is assigned a set of membership values, when there is no local optimal solution. The updated location of the
indicating the degree to which it belongs to each fuzzy set, where ξ, ϖ
and ϛ respectively represent Low, R, and Up three parameters: Low de­ dragonfly is Z
̃′ =Z̃′ + Levy(ξ)ΔZ
t+1
̃′ . ξ denotes the dimension, and the
t t

scribes the minimum change in the wind speed series window, R de­ Lévy function is Levy(x) = 0.01 × (γ 1 × δ)/|γ2 |1/β , where δ can be
scribes the average value of the change, and Up describes the maximum expressed as Eq. (6)
change. After data normalization, Up, Low and R are used as variables
Γ(1 + β) × sin(π2β) 1/β
and applied to the time series wind speed input model to obtain the label δ = ( 1+β ) , Γ(x) = (x − 1)! (6)
Γ( 2 ) × β × 2(β−2 1)
vector and corresponding eigenmatrix of the input prediction model.
where γ 1 and γ 2 are arbitrary numbers within the interval [0, 1], and β is
2.2. Dragonfly algorithm a fixed constant. According to this series of descriptions, the pseudocode
is presented in Table 2.
In 2015, Mirjalili proposed the Dragonfly algorithm (DA), a novel
swarm intelligence algorithm (Hammouri et al., 2020; K.S.S.R. & Mur­
ugan, 2017). The algorithm mimics the natural actions of dragonflies in 2.3. Innovative combined prediction system
search of prey. Mirjalili reported five factors influencing the location
update of the dragonfly algorithm: separation, queuing, aggregation, In response to the shortcomings of the sub-model, this study estab­
predation, and avoidance of natural predators (J. Li, Lu, Yao, Cheng, & lishes a combined model with the structure shown in Fig. 1. The first
Qin, 2019b; L.-L. Li, Zhao, Tseng, & Tan, 2020). The mathematical section performs data preparation technology. In the second section,
model is expressed as follows. four classic single models are selected for forecasting. In the third sec­
Separation: Each individual dragonfly is separated from the group tion, the predictions of the sub-models are weighted based on the no-
using Eq. (3). negative constraint theory, expressed as Eq. (7) and Eq. (8)

̃ Se (k) = − ∑ X′(k) − X′ (k)


N
Table 2
Ω j (3)
The pseudocode of DA.
i
j=1

Algorithm: DA
̃ Se (k) is the location vector of the i dragonfly’s separation
where Ω Parameters:
i th
t–current iteration number.
behavior between similar species; X′j (k) is the current location of the
N–population size.
dragonfly; X′j (k) is the location of the jth dragonfly adjacent to X′(k), and Fi–the fitness function of ith dragonfly.
Itermax–the maximum number of iterations.
N is the number of dragonflies adjacent to X′(k).
ul–upper bound of weight.
Queuing: Each dragonfly matches the velocity of its neighboring bl–lower bound of weight.
̂ Al (k) = ∑N V
dragonflies with the velocity expression T ̃ ̂ Al 1:/*Initialize the parameters of DA.*/
i j=1 j (k)/N. T i (k)
2: WHILE t < Itermax
denotes the position vector of the ith individual dragonfly in line, and 3: FOR i = 1 to N
̃ j (t) refers to the velocity of the jth neighboring dragonfly.
V 4: Calculating the fitness value Fi;
Aggregation: The aggregation behavior of dragonflies grouped 5:/* Select the archive member situated in the least populated region as food sources
Co Co
to enhance coverage */

together can be expressed as Λi (k) = N j=1 X j (k)/N − X (k). Λi (k) de­
′ ′ 6: IF
7: Fi < FBest; /* The best fitness value in this iteration*/
notes the location vector of the ith dragonfly gathering.
8: Set the location of the individual as the food source X ̃ fo .
Predation: The predatory behavior of dragonflies can be calculated
9: END IF;
as Eq. (4) 10:/* Designate the archive member located in the most populated area as adversaries
fo */
(4)
Fo
̃ − X′(k)
Φi (k) = X 11: IF
12: Fi > FBest;
Fo ̂ en .
where Φi (k) is the location vector of the ith dragonfly at the time of 13: Set the location of the individual as the enemy X
14: END IF
̃ fo is the position of the food.
predation, and X 15: END FOR
Avoidance of natural predators: The dragonfly escapes from its 16: FOR i = 1 to N
En En ̂ en ;
̃ fo and enemy X
natural enemies using the equation, Ki (k) = X ̂ en + X′(k). K (k) rep­ 17: Update the food source X
i
18: Update σ′, τ′, λ′, φ′, ε′;
resents the directional vector of the ith dragonfly in avoiding a natural Se Al Co En
Fo
̂ en indicates the location of the natural enemy. 19: Calculate Ω̃ (k), T
̂ (k), Λ (k), Φ (k), K (k);
enemy, and X 20:
i i i
/*Update neighboring dragonfly*/;
i i

The position of food is the current optimal location, whereas the 21: IF At least one dragonfly is adjacent to a dragonfly
natural enemy’s position represents the current least favorable location. 22: /*Update velocity vector ΔZ ̃′ */
t+1
The step vector is the flight vector and step length of the dragonfly, Se
′ ̂ Al ′ Co ′ Fo
En
t+1 = (σ Ωi (k) + τ T i (k) + λ Λi (k) + φ Φi (k) + ε Ki (k)) + ωΔX t ;
̃′
ΔZ ′̃ ′ ̃′
expressed as Eq. (5)
23: /*Update position vector Z */
̃′
t+1

̃ Se (k) + τ′ T
̃′ = (σ′Ω
ΔZ
Co En
̂ Al (k) + λ′Λ (k) + φ′ΦFo (k) + ε′K (k)) + ωΔX
̃′ ̃′
Z t+1 = Z t + ΔZ t+1 ;
̃′ ̃′
t+1 t
i i i i i
24: ELSE
(5)
25: /*Update position vector Z
̃′ */
t+1

t+1 = Z t + Levy(ξ)ΔZ t ;
̃′
Z ̃′ ̃′
where σ′ is the separation weight; τ′ is the alignment weight; λ′ is the 26: END IF
aggregation weight; φ′ is the prey weight factor; ε′ is the separation 27: END FOR
weight; ω is the inertia weight, and t is the number of iterations. 28: END WHILE
29: END
In nature, dragonflies do not always remain in the same place. They

4
X. Wang et al. Expert Systems With Applications 241 (2024) 122477

Fig. 1. Main processes of the PCFS. Note: The wind-speed prediction system encompasses four main stages: In the data pre-processing section, a triangular
membership function is utilized to granulate the historical data within the selected time window, extracting the average values of the data variations within that time
window. Secondly, in the second stage, we selected four sub-models for forecasting. Subsequently, an optimization algorithm is employed to weight the prediction
results of the sub-models, yielding the comprehensive prediction results of the combined system. Finally, PCFS is employed for point prediction and interval pre­
diction experiments, followed by an evaluation of the model.

T ∑
∑ m ∑
m objectives can be optimally satisfied at the same time, the set of Pareto
̂ = PT EP =
min N ωi ωj e∼’ it e∼’ jt (7) optimal solutions represents the optimal solutions (Ning, Zhang, Liu, &
Zhang, 2018).
t=1 j=1 i=1

{
BT P = 1 Definition 1. It is presumed that there are two solutions s1 and s2 . If s1 and
st. (8)
P⩾0 s2 satisfy ∀i ∈ [1, k], fi (s1 )⩽fi (s2 ), and ∃j ∈ [1, k], fj (s1 ) < fj (s2 ) and s1
dominates s2 (s1 ≻ s2 ).
where P = (ω1 , ω2 , ⋯, ωm )T is the weight vector; E = (Eij )m×m = e’Ti e’j is
Definition 2. If there is a solution s1 ∈ X and non-existent s2 satisfies
the error matrix, and B = (1, 1, ⋯, 1)T is the unit column vector. F(s2 ) ≻ F(s1 ), s1 is the Pareto optimal solution of MOp.
In the fourth section, simulation prediction experiments are per­
formed using the proposed combined forecasting system (PCFS), and the Definition 3. The set of all Pareto solutions is known as the Pareto optimal
model is evaluated. set.
To achieve MOp, the MOda system demonstrated the ability to find
Pareto optimal solutions using this definition.
2.4. Pareto optimal solution theory
Proof. The weight vector P is obtained from the following procedure
Eq. (10).
MOp problem entails optimizing two or more objectives simulta­
neously, with each objective indirectly impacting the others due to ∑
N

constraints. There may be objective constraints that result in multiple P = argmin Li (Fi , Ai )
solutions. The MOp problem can be mathematically expressed as Eq. (9)
i=1
∑N ∑⃒ ⃒}
= arg min {min ⃒ωn Fn,i − Ai ⃒ (10)
minF(X) = {f1 (X), f2 (X), ⋯, fk (X)}
(9) i=1 n∈N
s.t. gi (X)⩽0, i = 1, 2, …, h, X ∈ Rn ∑
s.t. ωn = 1, ωn ⩾0.
where fi (X) (1⩽i⩽k) is the objective function, and gi (X) denotes the n∈N

constraint with n decision variables, k objective functions, and h sig­ We assume that zn,i = min
∑ ⃒ ⃒
⃒ωn Fn,i − Ai ⃒. The equation can be
n∈N
nifies the constraints (Wang, Zhou, & Li, 2022). In cases where multiple

5
X. Wang et al. Expert Systems With Applications 241 (2024) 122477

rewritten as Eq. (11) variation ϛ, the average raw data variation ϖ, and the minimum raw
data variation ξ; ϖ was used for prediction. For wind-speed prediction,

N
P = arg min zn,i the dataset was split into training samples, consisting of 2488 data
i=1 points, with an additional 440 data points serving as test samples.
⎧ ∑
(11) Table 4 showcases the statistical characteristics of the original data
ωn = 1, ωn ⩾0

n∈N
collected from the three locations. Fig. 2 illustrates the primary nu­
s.t. merical characteristics of the three sites.
⎩ ⃒ ⃒
zn,i − ⃒Fn,i − Ai ⃒⩽0, n ∈ N Meanwhile, in recognition of the potential divergence between the
⃒ ⃒ final prediction and the real dynamic model due to cumulative errors in
If the constraint is not satisfied, then ∃zn,i > ⃒Fn,i − Ai ⃒ is applied. In
single-step prediction (Zhu & Zhu, 2024), this paper adopts a multi-step
other words, the existence of z′n,i = zn,i − ε, ε > 0 meets z′n,i = rolling prediction approach to forecast the state or outcome of multiple
⃒ ⃒
⃒Fn,i − Ai ⃒. The optimal solution at this point is z′n,i = future time steps based on the current state. We suppose a single-step
∑ ⃒ ⃒
min n∈N ⃒ω′n Fn,i − Ai ⃒. Thus, these assumptions are not valid. prediction model represented by the function g. Given the current
1− step 1− step
state Xt , it predicts the state of the next time step as Xt+1 = g(Xt ).
3. Experiments and analysis The multi-step prediction technique utilizes the aforementioned
model multiple times to anticipate the states of consecutive future time
The data sources and performance measures are discussed in this steps. For instance, to forecast the state of two future time steps, the one-
section. Point and interval forecasts were devised to showcase the step prediction model is applied consecutively as follows: {Xt+1 ,
2− step

exceptional efficacy of the innovative combined system. The experi­ 2− step 2− step 2− step 2− step 2− step
mental parameters are outlined in Table 3. In addition, the simulation Xt+2 |Xt+1 = g(Xt ), Xt+2 = g(Xt+1 )}. By extension, the fore­
experiments were carried out on MATLAB 2020b, using the Windows 11 casted state for m-step can be attained. This multi-step prediction
professional operating system. The hardware specifications are as fol­ methodology enhances the model’s resilience in long-term forecasting
lows: Intel(R) Core(TM) i7-11800H 2.30 GHz CPU and 16 GB RAM. scenarios.

3.2. Performance metrics


3.1. Datasets
Assessing the predictive ability of a model involves evaluating its
The study employed three sets of data extracted from the Penglai performance in making accurate predictions. To determine the best
wind farm located in Shandong Province, China. For experimental method effectively, four prediction accuracy evaluation indexes are
analysis of the predictive model, data were selected once every ten selected in this section for point prediction. These criteria are the sum of
minutes, for a total of 8784 data samples. The fuzzy information gran­ square error (SSE), mean square error (MSE) mean absolute error
ulation method was used to select triangular fuzzy particles for fuzzy (MAE), and mean absolute percentage error (MAPE), provide a clearer
information granulation. explanation of the performance metrics(Wang, Wang, Li, Peng, & Liu,
Three datasets (30 min each) were used as the fuzzy granulation 2016). The definitions and detailed calculations for these indicators are
window. Fuzzification of the raw data produces the maximum raw data listed in Table 5.
Evaluating the accuracy of the prediction interval involves consid­
Table 3 ering three key indicators, the prediction interval coverage probability
Experiment parameters of key models. (PICP), average interval score (AIS) and the prediction interval
Model Symbol Parameter Explain Value normalized averaged width (PINAW). The specific formula is shown in
BPnn Ni Input layer nodes 5 Table 5.
Nh Hidden layer nodes 5
No Output layer nodes 1 3.3. Test method
Et Epoch 100

The predictive performance of the combined model was assessed


ELM Ni Input layer nodes 5 using the Diebold–Mariano (DM) test, demonstrating its superiority to
Nh Hidden layer nodes 10
other models. The following section describes the basic principles of the
No Output layer nodes 1
fa Activation function ‘Sigmoid’ two tests.

Enn Ni Input layer nodes 5


Nh Hidden layer nodes 30
No Output layer nodes 1
Et Epoch 1000 Table 4
fa Activation Function ‘Tansig’ Data statistical characteristics.
Dataset Samples Sample size Statistical metric (m/s)
MOda Itermax The upper limit of iterations 200
Max Min Mean Std.
Arcsize Archive Size 500
Popsize Dragonfly Number 40 Site 1 All samples 2928 18.6 0.8 7.4318 3.1551
Training 2488 18.6 0.8 7.3526 3.0094
Testing 440 17.4 1.3 7.8800 3.8510
MOgoa, MOalo AS Archive Size 500
Itermax The upper limit of iterations 200
IN Individual Number 40 Site 2 All samples 2928 18.5 0.6 6.9518 3.1054
Training 2488 18.5 0.6 6.9491 3.0199
Note:To mitigate the impact of randomness on the models and ensure a valid
Testing 440 16.4 1.2 6.9668 3.5542
comparison of prediction performance among different models, the parameters
of each model were meticulously configured and optimized. This involved a
comprehensive process that encompassed thorough parameter setting, in-depth Site 3 All samples 2928 18.6 0.7 6.9099 3.1291
Training 2488 18.5 0.7 6.9186 3.0455
optimization, and a minimum of ten experimental iterations with subsequent
Testing 440 18.6 1.0 6.8605 3.5685
adjustments.

6
X. Wang et al. Expert Systems With Applications 241 (2024) 122477

Fig. 2. The main numerical features of the three sites.

3.3.1. DM test The original and alternate hypotheses are formulated as follows Eq.
In statistical analysis, when evaluating the performance of fore­ (12) (Liu et al., 2022):
casting models, the DM test serves as a common method, essentially
̃ ε1 )] = E[Y(
̃ ε2 )]
being a t-test. Proposed hypothesis values for the population parameters H0 : E[Y( t t
(12)
are suggested, and the sample data are employed to ascertain the val­ ̃ ε )] ∕ ̃ ε2 )]
1
H1 : E[Y( t = E[Y( t
idity of the hypothesis (Wang, Wang, Zeng, & Lu, 2022b). The following
paragraphs outline the main steps of the DM test. where εit represents εi = |Fi − Ai | and the prediction value calculated by
Building assumptions the novel model, and Y( ̃ εi ) represents the prediction error function.
t

7
X. Wang et al. Expert Systems With Applications 241 (2024) 122477

√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
Table 5 ̃ 1 (1 − ̃ 2 − (ω ̃ 1 )2 ).
ω ω
Performance metrics.
Metric Definition Equation
⃒ ⃒ 3.4. Point forecasting experiments and analysis
MAPE The mean absolute percentage 1 ∑N ⃒⃒Ai − Fi ⃒⃒
MAPE = × 100%
error N i=1 ⃒ Ai ⃒
MSE The average of error squares 1 ∑N To guarantee the precision of the prediction system, point fore­
MSE = × i=1 (Fi − Ai )2 casting experiments were carried out utilizing data from three units at
N

SSE The squared sum of the errors SSE = Ni=1 (Ai − Fi )2 the Penglai wind farm.
MAE The absolute mean error of the 1 ∑N Experiment I: Comparison with FIg-based models
MAE = |Fi − Ai |
forecasted outcomes N i=1
In the model proposed in this paper, the initial speed and speed
PICP The coverage probability of 1 ∑N
prediction interval PICP = ci , ci = boundary of Dragonfly algorithm are set as Si = 1, Ss = 4 respectively,
N i=1
{ [ α α] and the upper limit of iterations is Itermax = 200, Archive Size is Arcsize
1, ifAi ∈ Li , Ui
= 500, Dragonfly Number is Popsize = 40. This section confirms the
0, else
correctness of PCFS, based on a comparison with four models, FIg-BPnn,
AIS The average interval score 1 ∑N α
AIS =
N i− 1 i
S , FIg-ELM, FIg-Enn, and FIg-ARima. The final outcomes are showcased in
Sαi = Table 6 and Fig. 3. From the predicted outcomes, we derived the

⎨ − 2αξi − 4(Li − Ai ), ifAi < Li


α α α following analysis:
− 2αξi , ifAi ∈ [Li , Ui ]
α α α

(a) The novel combined forecasting model obtained the smallest

⎩ − 2αξα − 4(Ai − Uα ), ifAi > Uα
i i i

PINAW The normalized averaged width 1 ∑N error when combined with the sub-models. The novel system
PINAW = (Uα − L αi )
of prediction interval NR i=1 i provides perfect forecasting performance at Site 1. In the FIg-
Note: N Fi and Ai indicate the sample size, predicted value and true value. Uαi based model, the MAPE accuracies of FIg-BPnn, FIg-ARima, FIg-
and Lαi means the ith upper bound and lower bound, respectively. R means the ELM, and FIg-Enn decreased sequentially. The corresponding
predicted value range. ξαi is specifically calculated as ξαi = Uαi − L αi at the level of MAPE values wereMAPESite11− step ¼ [14.16 %, 14.56 %, 14.68 %,
α. 14.84 %, and 15.39 %].
(b) With one-step prediction at Site 2, the proposed model had SSE,
MSE, MAE, and MAPE values of [SSESite21− step , MSE1− step , MAE1− step ,
Site2 Site2

Determining the test statistic


1− step ] ¼ [322.06, 1.69, 0.91, 15.01 %]. Thus, the com­
MAPESite2
To determine whether the original hypothesis is true, a statistic is bined model remained perfect for FIg-based models. Similarly,
constructed. The distribution of this statistic is used to make the deter­
the PCFS with MAPE values ofMAPESite2 2− step ¼ 20.78 %
mination, known as the test statistic. The DM statistic is expressed in Eq.
(13). andMAPESite2
3− step ¼ 24.38 %, outperforms the other FIg-based
models.
1 ∑N √̅̅̅̅̅̅̅̅̅̅̅ (c) Considering one-step prediction at Site 3, perfect performance
DM = ̃ ε1 ) − Y(
[Y( ̃ ε2 )]/ S2 /N (13)
metrics were calculated using the novel combined model. The
t t
N t=1
MAPE values in one to three steps wereMAPESite3
1− step ¼ 15.74 %,
where S represents the Y( ̃ ε1 ) − Y(
̃ ε2 ) variance estimate, and N is the
2− step ¼19.47 %, andMAPE3− step ¼ 22.54 %. The MAPE of
MAPESite3
t t Site3

number of predicted values. the FIg models from lowest to highest was FIg-Enn, FIg-ELM, FIg-
Making judgments based on samples BPnn, and FIg-ARima.
If the original hypothesis is accepted, it can be considered that the
prediction errors of any two models are not significantly different. If the Remark: According to the results of experiment I, under different
alternative hypothesis is accepted, it indicates that the prediction errors prediction steps, the prediction evaluation index value of PCFS is lower
of the two models differ greatly. than that of the single model after fuzzy information granulation. The
experiment fully verifies the necessity of using MOda for weight opti­
3.3.2. Forecasting effectiveness
mization, which makes the model can synthesize the advantages of
The main principal of forecast effectiveness testing is described as multiple prediction models and effectively avoid the defects of using
follows.
only a single model.
̃k =
The kth order validity can be described by Eq. (14). ω
∑N ̃k
i=1 Qi Ωi , Experiment II: Comparison with traditional single model
where Qn represents a discrete probability distribution, but is not The second experiment included comparison with traditional single
∑N models BPnn, ELM, Enn, and ARima. The aim was to demonstrate the
generally available. We denote it as Qn = 1/N; i=1 Qi = 1. Ωn repre­
̃
excellent results of the PCFS. The optimal parameters of the model pa­
sents the prediction accuracy, expressed as Ω ̃ i = 1 − |ε″i |. rameters obtained through multiple experiments are shown in Table 3.
⎧ The anticipated outcomes are displayed in Table 7 and Fig. 4, and can be
⎨ − 1, (Ai − Fi )/Ai < − 1
described as follows.
ε″i = (Ai − Fi )/Ai , − 1 < (Ai − Fi )/Ai < 1 (14)

1, (Ai − Fi )/Ai > 1
(a) When it comes to Site 1, the innovative model exhibited the most
The kth order validity can also be expressed in matrix form. outstanding forecasting performance in one-step forecasting,
̃ ω̃1, ω
̃ 2 , ⋯, ω
̃k) achieving a MAPE value ofMAPESite1 1− step ¼ 14.46 %. The PCFS
H(
outperformed the classic single models in terms of predictive
If H
̃ is a univariate function, then the first-order validity test is the performance. The MAPE values of BPnn, ELM, Enn, and ARima
anticipated sequence of predictive precision, and H( ̃1) = ω
̃ ω ̃ 1 . If wereMAPESite1
1− step ¼ [18.39 %, 19.93 %, 20.11 %, and 21.7 %],
√̅̅̅̅̅̅̅̅̅̅̅̅̅
̃ y) = x(1 −
H(x, y − x ) is a bivariate function, the second-order pre­
2 respectively. The proposed model also had the most significant
dictive validity is the variation between the expected value and the MAPE with two- and three-step forecasting.
standard deviation of the prediction series; H( ̃1, ω
̃ ω ̃2) = (b) For Site 2, PCFS outperformed traditional single models, with
MAPE values ofMAPESite2 1− step ¼ 15.01 %, MAPE2− step ¼20.78 %,
Site2

8
X. Wang et al. Expert Systems With Applications 241 (2024) 122477

Table 6
The effectiveness of comparison with FIg-based models.
Dataset Model SSE MSE MAE MAPE(%)

1-step 2-step 3-step 1-step 2-step 3-step 1-step 2-step 3-step 1-step 2-step 3-step

Site 1 FIg-BPnn 620.1051 1224.8738 1676.3808 2.1747 3.9702 5.2719 1.0929 1.5532 1.7568 14.5698 22.6306 27.0415
FIg-ELM 639.8302 1190.7361 1734.6457 2.2817 3.7897 5.6024 1.1245 1.4996 1.8295 14.8448 22.0843 27.6443
FIg-Enn 697.7785 1234.0227 1703.48 2.5613 3.9694 5.3531 1.2097 1.5432 1.7742 15.3968 22.1724 27.2148
FIg -ARima 630.6154 1657.9399 2800.8843 2.1126 5.1761 8.7317 1.0951 1.7305 2.249 14.6869 24.6478 32.1049
Proposed Model 373.9944 618.3548 960.5896 2.0216 3.2545 5.0873 1.0428 1.3658 1.7548 14.4616 20.8199 24.5147

Site 2 FIg-BPnn 563.2518 1189.2457 1561.6495 1.6969 3.5975 4.4182 0.9919 1.3834 1.643 15.5814 24.1275 29.8698
FIg-ELM 566.4472 1160.5054 1568.2836 1.689 3.4239 4.402 0.98417 1.3445 1.6406 15.8785 24.4707 30.0587
FIg-Enn 556.5455 1177.108 1604.7722 1.659 3.5194 4.5791 0.9118 1.3776 1.6659 15.8191 24.7724 30.1431
FIg-ARima 857.0947 1417.8296 2186.4643 2.3545 3.8538 5.9624 1.0886 1.4275 1.7927 20.0123 26.004 32.4759
Proposed Model 322.0678 615.5749 822.1623 1.6951 3.2399 4.3272 0.9053 1.376 1.6379 15.0164 20.7845 24.383

Site 3 FIg-BPnn 567.7311 1078.4977 1622.1504 1.8624 3.4007 4.9401 0.9376 1.3703 1.626 16.5319 23.8441 30.2557
FIg-ELM 547.727 1082.9266 1655.116 1.7865 3.3376 5.157 0.95303 1.3414 1.6784 16.1805 23.8772 30.8344
FIg-Enn 549.1642 1079.2776 1599.3258 1.7994 3.344 4.8159 0.9522 1.3581 1.5921 16.0648 24.6145 29.619
FIg-ARima 919.7494 1237.4763 1643.7869 3.159 4.1833 5.5014 1.2907 1.5119 1.7363 19.0763 23.06 27.0224
Proposed Model 330.0868 518.7739 726.5712 1.7843 2.7304 3.8241 0.9238 1.2475 1.4924 15.7377 19.4686 22.5431

Note: The forecasting error results of FIg-based models are shown in table, including in FIg-BPnn, FIg-ELM, FIg-Enn, FIg-ARima and the proposed model. The
∑ 1 ∑ 1 ∑ 1 ∑N
performance metrics are specifically expressed as: SSE = Ni=1 (Ai − Fi )2 , MSE = × Ni=1 (Fi − Ai )2 , MSE = × Ni=1 (Fi − Ai )2 , MAPE = |(Ai − Fi )/Ai | ×
N N N i=1
100%.

Fig. 3. Analysis of experimental results in Experiment I.

andMAPESite2
3− step ¼ 24.38 %. The proposed combined forecasting
Remark: The evaluation criteria of the FIg-based model are better
model exhibited a decline, with DBPnn ¼ 7.08 %, DELM ¼ 7.64 %, than those of the single model. And proposed model outperforms typical
DEnn ¼ 7.929 %, and DARima ¼ 11.9415 %. The second-most single models. Thus, it is useful for forecasting wind speed.
effective model was Enn. The error of PCFS was smaller than Experiment III: Comparison among combined models employ­
that of the conventional single models at Site 2. ing diverse multi-objective optimization algorithms
(c) The optimum SSE, MSE, MAE, and MAPE values for the novel This experiment verified whether the novel model outperforms other
model are [SSESite3 combination forecasting models that use different MOp algorithms, such
1− step , MSE1− step , MAE1− step , MAPE1− step ] ¼
Site3 Site3 Site3
as MOgoa and MOalo. The MOgoa and MOalo were tested by adjusting
[330.08, 1.78, 0.9238, 15.73 %], respectively, serving as the
the parameters. All parameters are Arcsize = 500, Itermax = 200, Indnum
performance indicators at Site 3. Thus, PCFS is superior. The
= 40 to obtain the optimal results, seeing Table 3 for details. Table 8 and
proposed model exhibited a decline, with DBPnn ¼ 9.37 %, DELM
Fig. 4 present the performance metrics of these models. The experi­
¼ 8.05 %, DEnn ¼ 7.5 %, and DARima ¼ 9.24 %.
mental results are analyzed as follows.

9
X. Wang et al. Expert Systems With Applications 241 (2024) 122477

Table 7
The efficiency of comparison with traditional single model.
Dataset Model SSE MSE MAE MAPE (%)

1-step 2-step 3-step 1-step 2-step 3-step 1-step 2-step 3-step 1-step 2-step 3-step

Site 1 BPnn 996.0389 1527.8739 2001.485 3.3839 4.9531 6.4716 1.3587 1.6836 1.9164 18.3965 25.786 30.101
ELM 1271.5214 1551.0619 1940.2781 4.9046 4.9717 6.1995 1.5309 1.6765 1.87 19.9318 25.4157 29.2115
Enn 1150.4602 1593.4945 2230.3024 4.0687 5.2191 7.285 1.5227 1.7649 2.1082 20.1056 25.5555 31.2488
ARima 1373.6736 20.954975 3108.8829 4.5168 6.7753 10.0993 1.5024 1.888 2.2941 21.7046 27.0436 32.5843
Proposed 373.9944 618.3548 960.5896 2.0216 3.2545 5.0873 1.0428 1.3658 1.7548 14.4616 20.8199 24.5147
Model

Site 2 BPnn 1066.5815 1617.2613 1972.3226 3.7129 5.2019 6.0763 1.2896 1.6456 1.7836 22.0917 28.5648 33.0705
ELM 1253.6718 1425.1678 1853.7738 4.5895 4.2613 5.4743 1.313 1.5177 1.7198 22.6557 27.9915 32.7029
Enn 1030.1323 1378.833 1858.6518 3.2771 3.9997 5.4537 1.2693 1.4908 1.7014 22.9454 27.7027 32.5108
ARima 1631.8313 2347.2698 3650.9807 5.1169 7.462 11.3755 1.5726 1.8611 2.3326 26.9579 31.4868 40.3551
Proposed 322.0678 615.5749 822.1623 1.6951 3.2399 4.3272 0.9053 1.376 1.6379 15.0164 20.7845 24.383
Model

Site 3 BPnn 1052.8689 1543.0748 2039.1834 3.6723 4.9758 6.4591 1.3109 1.6525 1.814 21.334 28.8391 34.2557
ELM 1224.3869 1442.3131 1882.3912 4.6424 4.3974 5.7318 1.4277 1.5068 1.7296 22.0169 27.5152 32.2133
Enn 922.6567 1410.0095 18998.0211 3.0035 4.2399 5.7876 1.2239 1.4806 1.733 20.437 26.9733 32.088
ARima 1302.3276 1862.0829 2561.019 4.358 6.2381 8.5874 1.4293 1.7211 2.0192 23.7465 28.7048 33.4587
Proposed 330.0868 518.7739 726.5712 1.7843 2.7304 3.8241 0.9238 1.2475 1.4924 15.7377 19.4686 22.5431
Model

Note: The prediction error outcomes of classic individual models are shown in table, including in BPnn, ELM, Enn, ARima and PCFS. The performance metrics are
∑ 1 ∑ 1 ∑ 1 ∑N
specifically expressed as: SSE = Ni=1 (Ai − Fi )2 , MSE = × Ni=1 (Fi − Ai )2 , MSE = × Ni=1 (Fi − Ai )2 , MAPE = |(Ai − Fi )/Ai | × 100%.
N N N i=1

(a) The PCFS offers better prediction than the other models; the where 1 − α is the confidence level, L is the logistic distribution function,
MAPE value wasMAPE1− step
Low ¼ 14.16 % at Site 1. Among the and σ is the standard deviation. The prediction interval is defined by the
models evaluated, the MOalo-based combined model ranked upper bound (Uαi ) and lower bound (Lαi ) (Wang, Zhang, & Li, 2022a; Wu,
second in terms of effectiveness. The highest MAPE value Wang, & Hao, 2022).
wasMAPE1− step
High ¼ 14.56 %, obtained using the MOgoa-based This study used three evaluation metrics, PICP, AIS, and PINAW, as
combined model. shown in Table 5, to assess the effectiveness of interval prediction. For a
(b) For Site 2, the MAPE values of the other two models were given confidence level, the PICP and AIS should be as large as possible,
MAPEMOgoa-CFM ¼ 15.19 % and MAPEMOalo-CFM ¼ 15.25 % in and the PINAW should be as small as possible to improve the interval
one-step forecasting. The value for the PCFS decreased to DMOgoa- forecasting performance. The prediction outcomes are shown in Table 9
CFM ¼ 0.82 % in the three-step model.
and Fig. 5, with PINC ¼ [85 %, 90 %, 95 %].
(c) For Site 3, the MAPE value of PCFS was less than those of the In this experiment, a combination model of different optimization
models using different MOp algorithms. The decreased MAPE algorithms is selected as the benchmark model. Taking Site 1 as an
values of the different combined models were DMOgoa-CFM ¼ example, in 1-step and 2-step prediction, the combined prediction sys­
0.38 % and DMOgoa-CFM ¼ 0.78 % compared with different MOp tem proposed in this paper has the best performance in PINAW evalu­
algorithms in two-step forecasting. Of the three models, the ation index. At this time, the PINAW values of the combined prediction
FINC FINC
MOgoa-based combination model was the second-most accurate. ⇀
model arePINAW 1− step = [0.27, 028, 0.29] andPINAW 2− step = [0.25,

0.26, 0.28], respectively. It shows that the prediction interval of this


Remark: In experiment III, different optimization algorithms are
model is relatively narrow. However, MOgoa-CFM performs better in
introduced. Although the errors of both FIg-based and single prediction ⇀ FINC
models are optimized, the error value of PCFS is the smallest. This paper AIS evaluation indexAIS 2− step = [− 2.43, − 1.99, − 1.48]. The reason is
further proves the superiority of MOda in the prediction process, and that the residual error of the model is large, resulting in a large interval
provides solid support for its feasibility in practical application. obtained from the kernel density estimation curve. In 3-step, PCFS has
the best performance in PICP and AIS evaluation indicators, indicating
3.5. Interval forecasting experiments and results analysis that the model indicates that the model’s measurement interval
coverage is truly worth the most. Although the benchmark model per­
In this experimental segment, the interval prediction accuracy of the forms better on individual index values, comparing different confidence
interval estimation results was assessed using AIS, PICP, and PINAW. By levels and multi-step predictions at three sites can verify that the com­
fitting the fluctuation characteristics of point forecasts, forecasting in­ bined prediction model proposed in this paper is better than other
tervals with different forecasting interval nominal confidence (FINC) benchmark models as a whole.
values were calculated; MOgoa-CFM and MOalo-CFM were selected for In this experiment, the interval prediction results of the proposed
comparative experiments. prediction system are compared with the benchmark model. In the in­
The interval estimate can be expressed as Eq. (15) terval prediction tests with α = 0.01, α = 0.05 and α = 0.1 confidence
⎧ factors, it can be concluded that the proposed combined model has good
σ
⎪ α interval prediction performance at the significance level in most ex­
⎨ Li = Fi − √̅̅̅̅
⎪ Lα/2
N
(15) periments.

⎪ σ
⎩ Uαi = Fi + √̅̅̅̅ L1− α/2
N

10
X. Wang et al. Expert Systems With Applications 241 (2024) 122477

Fig. 4. Comparison of model prediction results in Experiment II and Experiment III.

Table 8
Comparison of model prediction results in Experiment III.
Dataset Model SSE MSE MAE MAPE(%)

1-step 2-step 3-step 1-step 2-step 3-step 1-step 2-step 3-step 1-step 2-step 3-step

Site 1 MOgoa-CFM 376.3974 684.1638 1028.9496 2.0284 3.6009 5.4155 1.0336 1.4797 1.8293 14.5616 21.0205 24.8504
MOalo-CFM 376.3866 696.5078 999.8321 2.0284 3.6658 5.2623 1.0311 1.4918 1.8054 14.522 21.0156 24.7708
Proposed Model 373.9944 618.3548 960.5896 2.0216 3.2545 5.0873 1.0428 1.3658 1.7548 14.4616 20.8199 24.5147

Site 2 MOgoa-CFM 344.8705 619.9108 837.3446 1.8151 3.2627 4.4071 0.9563 1.3715 1.6742 15.1927 20.8276 25.2069
MOalo-CFM 358.4268 618.4519 874.2215 1.8865 3.255 4.6012 0.9821 1.3695 1.7041 15.2598 20.817 24.7232
Proposed Model 322.0678 615.5749 822.1623 1.6951 3.2399 4.3272 0.9053 1.376 1.6379 15.0164 20.7845 24.383

Site 3 MOgoa-CFM 345.4288 553.5198 825.7744 1.818 2.9133 4.3462 0.9515 1.3157 1.642 15.8803 19.8531 23.7697
MOalo-CFM 345.578 589.7747 912.9018 1.8188 3.1041 4.8047 0.9428 1.3757 1.755 15.8806 20.2485 24.7087
Proposed Model 330.0868 518.7739 726.5712 1.7843 2.7304 3.8241 0.9238 1.2475 1.4924 15.7377 19.4686 22.5431

Note: The prediction error outcomes of combined models are shown, including in MOgoa-CFM, MOgoa-CFM and PCFS. The performance metrics are specifically
∑ 1 ∑ 1 ∑ 1 ∑N
expressed as: SSE = Ni=1 (Ai − Fi )2 , MSE = × Ni=1 (Fi − Ai )2 , MSE = × Ni=1 (Fi − Ai )2 , MAPE = |(Ai − Fi )/Ai | × 100%. The most satisfying results are
N N N i=1
highlighted in bold.

11
X. Wang et al. Expert Systems With Applications 241 (2024) 122477

Table 9
The result of interval forecasting.
Step FINC(%) Proposed Model MOgoa-CFM MOalo-CFM

PICP AIS PINAW PICP AIS PINAW PICP AIS PINAW

Site 1 1 85 83.78 − 1.7239 0.2651 84.32 ¡1.701 0.2666 84.3243 − 1.7091 0.266
90 86.49 − 1.3098 0.2758 85.41 ¡1.2839 0.2774 84.3243 − 1.2869 0.2767
95 87.57 − 0.8519 0.2942 88.1081 − 0.8251 0.2957 88.6486 ¡0.8226 0.2951
2 85 67.8947 − 2.4979 0.2494 69.4737 ¡2.4319 0.25 68.4211 − 2.4609 0.2509
90 69.4737 − 2.0569 0.2595 70.5263 ¡1.9944 0.2602 69.4737 − 2.0183 0.2612
95 71.0526 − 1.5356 0.2769 72.1053 ¡1.4774 0.2776 72.1053 − 1.4985 0.2787
3 85 57.8947 ¡3.2783 0.2344 53.1579 − 3.5251 0.2324 57.5947 − 3.3448 0.2382
90 58.9474 ¡2.8543 0.2437 55.7895 − 3.1009 0.2417 57.8947 − 2.9089 0.2479
95 62.6316 ¡2.3464 0.2597 57.8947 − 2.5843 0.2576 61.5789 − 2.3789 0.2644

Site 2 1 85 85.7895 ¡1.631 0.2559 84.2105 − 1.6854 0.2536 83.6842 − 1.712 0.2545
90 85.7895 ¡1.2739 0.2664 84.2105 − 1.3352 0.2639 84.2105 − 1.3618 0.2649
95 86.3158 ¡0.8706 0.2842 86.3128 − 0.9402 0.2817 85.2632 − 0.9676 0.2828
2 85 65.2632 − 2.3612 0.2368 67.8947 − 2.3477 0.2383 68.4211 ¡2.3453 0.2383
90 67.3684 − 1.9725 0.2466 71.0526 − 1.9581 0.2482 71.0526 ¡1.9563 0.2483
95 71.0526 − 1.5152 0.2634 73.1579 − 1.5037 0.2653 73.1579 ¡1.5024 0.2653
3 85 53.1579 − 3.1161 0.2335 54.2105 ¡2.9977 0.2389 49.4737 − 3.3237 0.2344
90 53.6842 − 2.7336 0.2431 55.7895 ¡2.5789 0.2486 50 − 2.9357 0.244
95 57.3684 − 2.2643 0.2594 61.0526 ¡2.0738 0.2601 53.1579 − 2.4551 0.2603

Site 3 1 85 88.1081 ¡1.6961 0.2468 87.5676 − 1.6967 0.2504 87.5676 − 1.7011 0.247
90 88.6484 ¡1.322 0.257 88.1081 − 1.323 0.2609 88.1081 − 1.3241 0.2572
95 89.1892 − 0.9052 0.2745 88.6486 ¡0.8999 0.2787 89.1892 − 0.9246 0.2748
2 85 83.6842 ¡1.965 0.2429 81.0526 − 1.9996 0.2411 81.0526 − 2.0457 0.2409
90 83.6842 ¡1.5405 0.2535 84.2105 − 1.5705 0.2517 83.1579 − 1.6176 0.2516
95 86.3158 ¡1.0599 0.2716 86.3158 − 1.0879 0.2698 86.8421 − 1.1317 0.2698
3 85 76.8421 ¡2.4143 0.238 71.0526 − 2.6219 0.2367 67.3684 − 2.8246 0.237
90 78.9474 ¡1.9563 0.2487 72.1053 − 2.1416 0.2475 68.4211 − 2.3382 0.2479
95 80 ¡1.423 0.267 76.3158 − 1.579 0.266 72.6316 − 1.7611 0.2665
{ [ ]
1 ∑N 1, if Ai ∈ Lαi , Uαi 1 ∑N α 1 ∑N
Note: The metrics are calculated as follows: PICP = ci ,ci = , AIS = S , PINAW = (Uα − L αi ) .The bold part indicates the
N i=1 0, else N i− 1 i NR i=1 i
optimal result for each performance metric.

4. Discussion novel combined system was significantly superior to that of the


comparative models.
The advantages of the new combined forecasting model are tested
using DM tests, prediction validity, index improvement rate, and com­
parison with the average strategy. 4.2. Improvement ratio

The purpose of this section is to further quantitatively compare the


4.1. Results of effectiveness test improvement range of PCFS and other models on indicators through the
index improvement rate test, so as to further confirm the superiority of
First, the innovative forecasting system is evaluated for its validity this model. We proposed the concept of PMAPE and evaluated the
using the DM test. The novel system was compared with the comparison improved efficiency of the novel combined model. The results allowed
models at various confidence levels based on the fundamental of DM for comprehensive evaluation. PMAPE can be defined by Eq. (16).
test, demonstrating the advantages of PCFS. The DM test outcomes are ⃒ ⃒
presented in Table 10. ⃒MAPEc − MAPEp ⃒
IRMAPE =⃒⃒ ⃒ × 100% (16)

At the α = 0.01 significance level, there are 14 values less than MAPEc
Z0.01/2 = 2.58; the remaining 16 DM test statistics reject the original
hypothesis with a probability of 99 %. In addition, there are ten DM where MAPEc represents the MAPE value of the comparison model, and
statistic values less than Z0.05/2 = 1.96 at the α = 0.05 significance level; MAPEp represents the MAPE value of the novel combined system. By
calculating IRMAPE , the degree of improvement in the predictive per­
the remaining prediction models accept the alternative hypothesis with
formance of the model can be visually assessed. A higher IRMAPE value
95 % probability. At the α = 0.1 significance level, there are 21 DM test
indicates a better model improvement, while a lower IRMAPE value may
statistics with values greater than Z0.1/2 = 1.645; the original hypothesis
indicate that the improvement is not significant or needs further opti­
H0 does not hold, and the original hypothesis H1 is rejected with 90 %
mization( Wang, Zhang, Wang, & Liu, 2021c). The calculation results
probability. The outcomes demonstrate that PCFS is outperforms the ten
are presented in Table 12.
comparative sub-models at the α = 0.1 significance level.
Predictive effectiveness was employed to evaluate the prediction
(a) For all models, a more significant improvement in predictive
accuracy, thereby providing further evidence to substantiate the supe­
validity was observed compared to the FIg-based models. FIg-
riority of the novel combined system (Hao, Yang, & Yin, 2023). The
ARima model has the largest improvement ratio in forecasting
mean predictive validity values are presented in Table 11. The devised
performance. Multi-step forecasting diminished the ability to
combined system exhibits superior first-order and second-order predic­
make accurate forecasts of the FIg-ARima model by an average
tive validity values compared to the comparison models, except for the
⇀ step
first-order validity in three-step forecasting. The effectiveness of the ofIR FIg− ARima ¼ [14.67 %, 17.06 %, 21.71 %].

12
X. Wang et al. Expert Systems With Applications 241 (2024) 122477

Fig. 5. The result of interval forecasting.

Table 10 Table 11
The result of DM test. The result of forecasting effectiveness.
Model 1-step 2-step 3-step Model 1-step 2-step 3-step

FIg-BPnn 0.8515 1.4669 1.0065 1- 2- 1- 2- 1- 2-


FIg-ELM 0.5326 0.9422 1.4536 order order order order order order
FIg-Enn 1.4836 1.3646 1.2628
FIg-BPnn 0.8556 0.7135 0.7828 0.6383 0.7489 0.5762
FIg-ARima 2.5770** 2.8756*** 2.9133***
FIg-ELM 0.8557 0.7158 0.7848 0.6404 0.7678 0.5747
BPnn 3.7784*** 3.0703*** 2.7675***
FIg-Enn 0.8525 0.7197 0.7842 0.6405 0.7672 0.5762
ELM 2.9526*** 2.6413*** 2.3987**
FIg-ARima 0.8328 0.6728 0.7678 0.5741 0.7167 0.5141
Enn 4.4230*** 2.8398*** 3.1890***
BPnn 0.8137 0.6478 0.7704 0.6016 0.7495 0.5733
ARima 4.3196*** 3.5742*** 3.5163***
ELM 0.8006 0.6224 0.7775 0.6085 0.7505 0.5752
MOgoa-CFM 1.9225* 2.5124** 3.3157***
Enn 0.8079 0.6454 0.7739 0.6086 0.7397 0.5652
MOalo-CFM 2.6945*** 2.5314** 5.3759***
ARima 0.7744 0.5796 0.7336 0.5287 0.6929 0.4801
Note: The null and alternative hypothesis are presented below: MOgoa-CFM 0.8565 0.7176 0.7830 0.6190 0.7328 0.5591
̃ ε1 )] = E[Y(̃ ε2 )] MOalo-CFM 0.8561 0.7170 0.7782 0.6124 0.7254 0.5527
H0 : E[Y( t t
. The DM statistic is defined as:DM = Proposed 0.8593 0.7197 0.7883 0.6411 0.7509 0.5766
̃ ε1 )] ∕
H1 : E[Y( ̃ ε2 )]
t = E[ Y( t Model
√̅̅̅̅̅̅̅̅̅̅̅
1 ∑ N
̃ ε1 ) − Y(
̃ ε2 )]/ S2 /N.
[Y( t t ̃k =
Note: The kth order validity can be described as below: ω
∑N ̃k
N t=1 i=1 Qi Ωi .Where

Qn = 1/N, and Ω
̃ i = 1 − |ε″i |.

(b) The novel forecasting system exhibited significant enhancements


when compared to the four classical individual models. For one-
step forecasting, the novel combined system reduced the mean 4.3. Averaging strategy
model MAPE by IRBPnn ¼ 26.55 %, IRELM ¼ 29.89 %, IREnn ¼
28.54 %, and IRARima ¼ 37.13 % compared with the four single In this part, the simple average strategy is used to discuss the pre­
models BPnn, ELM, Enn, and ARima, indicating that PCFS can diction effect of the model. Specifically, the arithmetic average is
produce perfect prediction results. calculated for the evaluation indicators of the other ten benchmark
(c) The combined models using different MOp algorithms showed models selected in this study. Then, the results are compared with the
the PCFS had slight improvements in predictive validity. The results of the new combined system proposed in this study, and the
most significant enhancement in accuracy was observed in three- specific comparison results are shown in Table 13.
step forecasting; the MAPE values were reduced byIR3− step At each of the three sites, the prediction accuracy of the model is
consistently better than the simple average results of the other bench­
MOgoa− CFM

¼ 3.26 % andIR3− step


MOalo− CFM ¼ 3.72 %, respectively. mark models. In the site1, for 1-step prediction the arithmetic average

13
X. Wang et al. Expert Systems With Applications 241 (2024) 122477

Table 12
The model of improvement radio.
Model Site 1 Site 2 Site 3 Average

1-step 2-step 3-step 1-step 2-step 3-step 1-step 2-step 3-step 1-step 2-step 3-step

FIg-BPnn 0.7426 8.0011 9.3442 3.6261 13.8556 18.3691 4.8040 18.3505 25.4914 3.0576 13.4024 17.7349
FIg-ELM 2.5814 5.7253 11.3210 5.4294 15.0637 18.8821 2.7366 18.4636 26.8898 3.5825 13.0842 19.0309
FIg-Enn 6.0740 6.0999 9.9214 5.0742 16.0982 19.1092 2.0361 20.9060 23.8897 4.3948 14.3680 17.6401
FIg-ARima 1.5340 15.5304 23.6419 24.9641 20.0719 24.9197 17.5013 15.5742 16.5762 14.6665 17.0588 21.7126
BPnn 21.3894 19.2589 18.5585 32.0270 27.2374 26.2696 26.2318 32.4923 34.1917 26.5494 26.3295 26.3399
ELM 27.4446 18.0825 16.0786 33.7191 25.7471 25.4409 28.5199 29.2442 30.0193 29.8945 24.3579 23.8462
Enn 28.0718 18.5306 21.5499 34.5559 24.9730 25.0003 22.9941 27.8227 29.7460 28.5406 23.7755 25.4321
ARima 33.3708 23.0136 24.7653 44.2968 33.9898 39.5789 33.7262 32.1765 32.6241 37.1313 29.7266 32.3228
MOgoa-CFM 0.6867 0.9543 1.3509 1.1604 0.2069 3.2685 0.8980 1.9367 5.1604 0.9150 1.0327 3.2599
MOalo-CFM 0.4159 0.9312 1.0339 1.5950 0.1561 1.3760 0.8998 3.8516 8.7645 0.9703 1.6463 3.7248
⃒( ⃒
Note: PMAPE can be defined as the following equation: PMAPE = ⃒ MAPEc − MAPEp )/MAPEc ⃒ × 100%. Where MAPEc denotes the MAPE value of the comparison model
and MAPEp denotes the MAPE value of the proposed combination model.

Site2
10.28 %] and△MAPE = [6.15 %, 8.54 %, 10.46 %] respectively. Through
Table 13
Comparison with the simple average strategy. the discussion of average strategy, the prediction accuracy of PCFS is
again verified to a certain extent compared with other benchmark
Dataset Multi- Model SSE MSE MAE MAPE
models, in which the least MAPE value changes by 6.15 %.
step

Site 1 1-step Average 1197.924 4.2185 1.4787 20.0346


4.4. Practical applications
strategy
Proposed 373.9944 2.0216 1.0428 14.4616
model The natural unpredictability and inconsistency of wind-speed pat­
2-step Average 1173.346 5.4798 1.7533 25.9502 terns make it more challenging to incorporate wind energy into the
strategy power grid, resulting in unstable power output from wind farms, which
Proposed 618.3548 3.2545 1.3658 20.8199
model
affects the quality of power output and threatens the grid. The proposed
3-step Average 2320.237 7.5139 2.0472 30.7864 combined model can compensate for these characteristics and effec­
strategy tively increase the economic benefits of wind farms (Li et al., 2019a;
Proposed 960.5896 5.0873 1.7548 24.5147 Tascikaraoglu & Uzunoglu, 2014). The specific application contribu­
model
tions are as follows:

Site 2 1-step Average 1245.5542 4.1741 1.3611 23.6627 (a) Improving economic efficiency of power systems. Accurate
strategy
Proposed 322.0678 1.6951 0.9053 15.0164
wind-speed predictions allow for more efficient turbine sched­
model uling; by effectively organizing generation schedules, the grid
2-step Average 1692.1329 5.2312 1.6288 28.9365 dispatching department can minimize the need for excessive
strategy system reserves, resulting in enhanced economic efficiency of
Proposed 615.5749 3.2399 1.376 20.7845
grid operations. Typically, a 1 % error in prediction eventually
model
3-step Average 2333.9322 7.0950 1.8844 34.6598 results in a 3 % error. Using the proposed model can significantly
strategy reduce wasted power and economic costs.
Proposed 822.1623 4.3272 1.6379 24.383 (b) Improved operation safety. An accurate prediction system can
model provide effective adjustment of wind-turbine operation. By pre­
dicting fluctuations in wind power, managers can adjust wind
Site 3 1-step Average 1125.5600 3.9191 1.3480 21.8836 turbines in a timely manner and reasonably arrange operating
strategy
methods and countermeasures to enhance grid safety and
Proposed 330.0868 1.7843 0.9238 15.7377
model
dependability. When the capacity of a wind turbine is less than
2-step Average 1564.3701 4.9628 1.5903 28.0081 the actual value, the wind turbine should be shut off as soon as
strategy possible to avoid machine damage and safety issues.
Proposed 518.7739 2.7304 1.2475 19.4686 (c) Ensuring grid operation stability. The signals of wind-power
model
systems can sometimes change suddenly. In such cases, fore­
3-step Average 2095.1537 6.6415 1.8240 33.0039
strategy casting results can be used as a reference for managers to guide
Proposed 726.5712 3.8241 1.4924 22.5431 the energy supply to ensure smooth operation. A dependable
model wind-speed forecast model helps ensure grid stability.

1− step 1− step 1− step 5. Conclusion


value of the benchmark model evaluation index is[A
̃
SSE , AMSE , AMAE ,
̃ ̃
̃ 1− step ] = [1197.92, 4.22, 1.48, 10.03], and the evaluation index value
A With the rapid advancement of wind power technology, precise
MAPE
of PCFS is decreased to a certain extent, which is[△SSE , △MSE ,
1− step 1− step wind-speed prediction has the potential to enhance the safety and reli­
1− step 1− step ability of grid operations. Furthermore, reducing wind-speed forecast
△MAE , △MAPE ] = [823.93, 2.20, 0.44, 5.58]. In 2-step and 3-step
errors can provide managers with accurate adjustment strategies and
prediction, the evaluation index value of the combined system pro­
increase economic benefits. However, due to the inherent randomness
posed in this paper also decreases. Similarly, in the site 2 and site 3
and instability, wind-speed prediction remains a challenging and com­
experimental results, the simple average MAPE values of other bench­
plex task. Many existing prediction models do not adequately consider
Site1
mark models are higher than PCFS, which are△MAPE = [8.65 %, 8.15 %, the importance of data preparation or the constraints of individual

14
X. Wang et al. Expert Systems With Applications 241 (2024) 122477

prediction models, resulting in inaccurate predictions. Therefore, this Hammouri, A. I., Mafarja, M., Al-Betar, M. A., Awadallah, M. A., & Abu-Doush, I. (2020).
An improved Dragonfly Algorithm for feature selection. Knowledge-Based Systems,
paper proposes an innovative combined prediction system that includes
203, Article 106131. https://doi.org/10.1016/j.knosys.2020.106131
fuzzy information granulation of the raw data, selection of optimal Hao, Y., Yang, W., & Yin, K. (2023). Novel wind speed forecasting model based on a deep
weighting coefficients using the multi-objective dragonfly algorithm, learning combined strategy in urban energy systems. Expert Systems with Applications,
and integration of four sub-models. The experimental results and anal­ 219, Article 119636. https://doi.org/10.1016/j.eswa.2023.119636
He, Y., Yan, Y., & Xu, Q. (2019). Wind and solar power probability density prediction via
ysis show that, compared with other benchmark models, this model fuzzy information granulation and support vector quantile regression. International
significantly improves prediction accuracy and stability. The MAPE Journal of Electrical Power & Energy Systems, 113, 515–527. https://doi.org/10.1016/
values for this model are 14.46 %, 15.02 %, and 15.74 %, respectively, j.ijepes.2019.05.075
K.S., S. R., & Murugan, S. (2017). Memory based Hybrid Dragonfly Algorithm for
with average improvement rates of 14.97 %, 16.48 %, and 19.11 %. By numerical optimization problems. Expert Systems with Applications, 83, 63–78.
accurately predicting wind speed using the combined system, it provides 10.1016/j.eswa.2017.04.033.
a reliable basis for power dispatching, guaranteeing the secure func­ Li, J., Lu, J., Yao, L., Cheng, L., & Qin, H. (2019b). Wind-Solar-Hydro power optimal
scheduling model based on multi-objective dragonfly algorithm. Energy Procedia,
tioning of wind turbines and effectively reducing economic costs. 158, 6217–6224. https://doi.org/10.1016/j.egypro.2019.01.476
Although this paper has made some achievements in wind speed Li, Y., Tong, Z., Tong, S., & Westerdahl, D. (2022b). A data-driven interval forecasting
prediction, further expansion and improvement are still needed. Firstly, model for building energy prediction using attention-based LSTM and fuzzy
information granulation. Sustainable Cities and Society, 76, Article 103481. https://
the proposed combined forecasting system does not consider the effects doi.org/10.1016/j.scs.2021.103481
of unconventional factors such as terrain, temperature differences, and Li, J., Wang, J., Zhang, H., & Li, Z. (2022a). An innovative combined model based on
air pressure. Secondly, the combined system in this paper only selects multi-objective optimization approach for forecasting short-term wind speed: A case
study in China. Renewable Energy, 201, 766–779. https://doi.org/10.1016/j.
four sub-models. There may be other model combination methods with
renene.2022.10.123
higher accuracy among the many available prediction models, and the Li, L.-L., Zhao, X., Tseng, M.-L., & Tan, R. R. (2020). Short-term wind power forecasting
identification of a better model requires further study. In the future, with based on support vector machine with improved dragonfly algorithm. Journal of
the continuous development of science and technology, we can utilize Cleaner Production, 242, Article 118447. https://doi.org/10.1016/j.
jclepro.2019.118447
more advanced statistical and machine learning algorithms to model Li, C., Zhu, Z., Yang, H., & Li, R. (2019a). An innovative hybrid system for wind speed
and predict the spatiotemporal changes of wind speed. By incorporating forecasting based on fuzzy preprocessing scheme and multi-objective optimization.
multi-source data, considering geographical environment and climate Energy, 174, 1219–1237. https://doi.org/10.1016/j.energy.2019.02.194
Liu, J., Wang, F., & Zhen, Z. (2020). Deep Learning Based Visualized Wind Speed Matrix
factors, and analyzing the characteristics and patterns of wind fields, we Forecasting Model for Wind Power Forecasting. 2020 IEEE 3rd Student Conference on
can enhance the accuracy and reliability of wind speed prediction. Electrical Machines and Systems (SCEMS), 952–958. 10.1109/
SCEMS48876.2020.9352407.
Liu, H., Tian, H., Liang, X., & Li, Y. (2015a). New wind speed forecasting approaches
Declaration of Competing Interest using fast ensemble empirical model decomposition, genetic algorithm, Mind
Evolutionary Algorithm and Artificial Neural Networks. Renewable Energy, 83,
1066–1075. https://doi.org/10.1016/j.renene.2015.06.004
The authors declare that they have no known competing financial
Liu, H., Tian, H., Liang, X., & Li, Y. (2015b). Wind speed forecasting approach using
interests or personal relationships that could have appeared to influence secondary decomposition algorithm and Elman neural networks. Applied Energy, 157,
the work reported in this paper. 183–194. https://doi.org/10.1016/j.apenergy.2015.08.014
Liu, Z., Jiang, P., Wang, J., & Zhang, L. (2022). Ensemble system for short term carbon
dioxide emissions forecasting based on multi-objective tangent search algorithm.
Data availability Journal of Environmental Management, 302, 113951. https://doi.org/10.1016/j.
jenvman.2021.113951
The datasets generated during and/or analysed during the current Mohammed, N. A., & Al-Bazi, A. (2022). An adaptive backpropagation algorithm for
long-term electricity load forecasting. Neural Computing and Applications, 34(1),
study are available from the corresponding author on reasonable 477–491. https://doi.org/10.1007/s00521-021-06384-x
request. Nie, Y., Liang, N., & Wang, J. (2021). Ultra-short-term wind-speed bi-forecasting system
via artificial intelligence and a double-forecasting scheme. Applied Energy, 301,
Article 117452. https://doi.org/10.1016/j.apenergy.2021.117452
Acknowledgement Ning, J., Zhang, B., Liu, T., & Zhang, C. (2018). An archive-based artificial bee colony
optimization algorithm for multi-objective continuous optimization problem. Neural
This work was supported by the National Natural Science Foundation Computing and Applications, 30(9), 2661–2671. https://doi.org/10.1007/s00521-
016-2821-7
of China (grant number 71671029). Niu, X., & Wang, J. (2019). A combined model based on data preprocessing strategy and
multi-objective optimization algorithm for short-term wind speed forecasting.
References Applied Energy, 241, 519–539. https://doi.org/10.1016/j.apenergy.2019.03.097
Qin, M., Li, Z., & Du, Z. (2017). Red tide time series forecasting by combining ARIMA and
deep belief network. Knowledge-Based Systems, 125, 39–52. https://doi.org/10.1016/
Akçay, H., & Filik, T. (2017). Short-term wind speed forecasting by spectral analysis from
j.knosys.2017.03.027
long-term observations with missing values. Applied Energy, 191, 653–662. https://
Qu, Z., Zhang, K., Mao, W., Wang, J., Liu, C., & Zhang, W. (2017). Research and
doi.org/10.1016/j.apenergy.2017.01.063
application of ensemble forecasting based on a novel multi-objective optimization
Akram, M., Nawaz, H. S., & Kahraman, C. (2023). Rough Pythagorean fuzzy
algorithm for wind-speed forecasting. Energy Conversion and Management, 154,
approximations with neighborhood systems and information granulation. Expert
440–454. https://doi.org/10.1016/j.enconman.2017.10.099
Systems with Applications, 218, Article 119603. https://doi.org/10.1016/j.
Sı̄le, T., Beķere, L., Cepı̄te-Frišfelde, D., Seņņikovs, J., & Bethers, U. (2014). Verification
eswa.2023.119603
of Numerical Weather Prediction Model Results for Energy Applications in Latvia.
Cassola, F., & Burlando, M. (2012). Wind speed and wind energy forecast through
Energy Procedia, 59, 213–220. https://doi.org/10.1016/j.egypro.2014.10.369
Kalman filtering of Numerical Weather Prediction model output. Applied Energy, 99,
Tascikaraoglu, A., & Uzunoglu, M. (2014). A review of combined approaches for
154–166. https://doi.org/10.1016/j.apenergy.2012.03.054
prediction of short-term wind speed and power. Renewable and Sustainable Energy
Do, D.-P.-N., Lee, Y., & Choi, J. (2016). Hourly Average Wind Speed Simulation and
Reviews, 34, 243–254. https://doi.org/10.1016/j.rser.2014.03.033
Forecast Based on ARMA Model in Jeju Island, Korea. Journal of Electrical Engineering
Wang, J., Li, H., Wang, Y., & Lu, H. (2021a). A hesitant fuzzy wind speed forecasting
and Technology, 11(6), 1548–1555. https://doi.org/10.5370/JEET.2016.11.6.1548
system with novel defuzzification method and multi-objective optimization
Dong, L., Wang, L., Khahro, S. F., Gao, S., & Liao, X. (2016). Wind power day-ahead
algorithm. Expert Systems with Applications, 168, Article 114364. https://doi.org/
prediction with cluster analysis of NWP. Renewable and Sustainable Energy Reviews,
10.1016/j.eswa.2020.114364
60, 1206–1212. https://doi.org/10.1016/j.rser.2016.01.106
Wang, J., Lv, M., Li, Z., & Zeng, B. (2023). Multivariate selection-combination short-term
Erdem, E., & Shi, J. (2011). ARMA based approaches for forecasting the tuple of wind
wind speed forecasting system based on convolution-recurrent network and multi-
speed and direction. Applied Energy, 88(4), 1405–1414. https://doi.org/10.1016/j.
objective chameleon swarm algorithm. Expert Systems with Applications, 214,
apenergy.2010.10.031
119129. https://doi.org/10.1016/j.eswa.2022.119129
Gao, Y., Wang, J., & Yang, H. (2022). A multi-component hybrid system based on
Wang, J., Niu, X., Liu, Z., & Zhang, L. (2020). Analysis of the influence of international
predictability recognition and modified multi-objective optimization for ultra-short-
benchmark oil price on China’s real exchange rate forecasting. Engineering
term onshore wind speed forecasting. Renewable Energy, 188, 384–401. https://doi.
Applications of Artificial Intelligence, 94, Article 103783. https://doi.org/10.1016/j.
org/10.1016/j.renene.2022.02.005
engappai.2020.103783
Guo, Z., Wu, J., Lu, H., & Wang, J. (2011). A case study on a hybrid wind speed
Wang, J., Wang, S., & Li, Z. (2021b). Wind speed deterministic forecasting and
forecasting method using BP neural network. Knowledge-Based Systems, 24(7),
probabilistic interval forecasting approach based on deep learning, modified
1048–1056. https://doi.org/10.1016/j.knosys.2011.04.019

15
X. Wang et al. Expert Systems With Applications 241 (2024) 122477

tunicate swarm algorithm, and quantile regression. Renewable Energy, 179, Wu, C., Wang, J., & Hao, Y. (2022). Deterministic and uncertainty crude oil price
1246–1261. https://doi.org/10.1016/j.renene.2021.07.113 forecasting based on outlier detection and modified multi-objective optimization
Wang, H. Z., Wang, G. B., Li, G. Q., Peng, J. C., & Liu, Y. T. (2016). Deep belief network algorithm. Resources Policy, 77, Article 102780. https://doi.org/10.1016/j.
based deterministic and probabilistic wind speed forecasting approach. Applied resourpol.2022.102780
Energy, 182, 80–93. https://doi.org/10.1016/j.apenergy.2016.08.108 Xing, Q., Wang, J., Jiang, H., & Wang, K. (2023). Research of a novel combined
Wang, S., Wang, J., Lu, H., & Zhao, W. (2021d). A novel combined model for wind speed deterministic and probabilistic forecasting system for air pollutant concentration.
prediction – Combination of linear model, shallow neural networks, and deep Expert Systems with Applications, 228, Article 120117. https://doi.org/10.1016/j.
learning approaches. Energy, 234, Article 121275. https://doi.org/10.1016/j. eswa.2023.120117
energy.2021.121275 Yang, W., Hao, M., & Hao, Y. (2023). Innovative ensemble system based on mixed
Wang, J., Zhou, Y., & Li, Z. (2022). Hour-ahead photovoltaic generation forecasting frequency modeling for wind speed point and interval forecasting. Information
method based on machine learning and multi objective optimization algorithm. Sciences, 622, 560–586. https://doi.org/10.1016/j.ins.2022.11.145
Applied Energy, 312, 118725. https://doi.org/10.1016/j.apenergy.2022.118725 Zhang, Y., Pan, G., Chen, B., Han, J., Zhao, Y., & Zhang, C. (2020). Short-term wind speed
Wang, K., Wang, J., Zeng, B., & Lu, H. (2022b). An integrated power load point-interval prediction model based on GA-ANN improved by VMD. Renewable Energy, 156,
forecasting system based on information entropy and multi-objective optimization. 1373–1388. https://doi.org/10.1016/j.renene.2019.12.047
Applied Energy, 314, Article 118938. https://doi.org/10.1016/j. Zhang, L., Wang, J., Li, Z., Zeng, B., & Huang, X. (2022). Uncertainty quantification of
apenergy.2022.118938 PM2.5 concentrations using a hybrid model based on characteristic decomposition
Wang, J., Yang, W., Du, P., & Niu, T. (2018). A novel hybrid forecasting system of wind and fuzzy granulation. Journal of Environmental Management, 324, Article 116282.
speed based on a newly developed multi-objective sine cosine algorithm. Energy https://doi.org/10.1016/j.jenvman.2022.116282
Conversion and Management, 163, 134–150. https://doi.org/10.1016/j. Zhao, J., Guo, Z.-H., Su, Z.-Y., Zhao, Z.-Y., Xiao, X., & Liu, F. (2016a). An improved multi-
enconman.2018.02.012 step forecasting model based on WRF ensembles and creative fuzzy systems for wind
Wang, J., Zhang, L., & Li, Z. (2022a). Interval forecasting system for electricity load speed. Applied Energy, 162, 808–826. https://doi.org/10.1016/j.
based on data pre-processing strategy and multi-objective optimization algorithm. apenergy.2015.10.145
Applied Energy, 305, Article 117911. https://doi.org/10.1016/j. Zhao, W., Wei, Y.-M., & Su, Z. (2016b). One day ahead wind speed forecasting: A
apenergy.2021.117911 resampling-based approach. Applied Energy, 178, 886–901. https://doi.org/10.1016/
Wang, J., Zhang, L., Wang, C., & Liu, Z. (2021c). A regional pretraining-classification- j.apenergy.2016.06.098
selection forecasting system for wind power point forecasting and interval Zhu, X., Liu, R., Chen, Y., Gao, X., Wang, Y., & Xu, Z. (2021). Wind speed behaviors
forecasting. Applied Soft Computing, 113, Article 107941. https://doi.org/10.1016/j. feather analysis and its utilization on wind speed prediction using 3D-CNN. Energy,
asoc.2021.107941 236, Article 121523. https://doi.org/10.1016/j.energy.2021.121523
Wang, Y., Zou, R., Liu, F., Zhang, L., & Liu, Q. (2021e). A review of wind speed and wind Zhu, G., & Zhu, F. (2024). Draw on advantages and avoid disadvantages by making a
power forecasting with deep neural networks. Applied Energy, 304, Article 117766. multi-step prediction. Expert Systems with Applications, 237, Article 121345. https://
https://doi.org/10.1016/j.apenergy.2021.117766 doi.org/10.1016/j.eswa.2023.121345

16

You might also like