You are on page 1of 11

Expert Systems With Applications 241 (2024) 122645

Contents lists available at ScienceDirect

Expert Systems With Applications


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

Hybrid deep learning and quantum-inspired neural network for day-ahead


spatiotemporal wind speed forecasting
Ying-Yi Hong a, *, Christian Lian Paulo P. Rioflorido a, Weina Zhang b
a
Dept. of Electrical Engineering, Chung Yuan Christian University, Taoyuan 32023, Taiwan
b
College of Computer Science and Technology, Shanghai University of Electric Power, Shanghai 201306, PR China

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

Keywords: Wind is an essential, clean and sustainable renewable source of energy; however, wind speed is stochastic and
Deep learning intermittent. Accurate wind power generation forecasts are required to ensure that power generation can be
Quantum computing scheduled economically and securely. This work proposes a hybrid deep learning technique that incorporates a
Quantum-inspired network
quantum-inspired neural network to predict wind speeds 24 h in advance. An innovative neural network tech­
Spatiotemporal data
Wind speed forecasting
nique is implemented by cascading parallel convolutional neural networks (CNNs) with a long short-term
memory (LSTM) and a quantum-inspired neural network (QINN). The proposed hybrid model is optimized
using two iterative loops: the outer loop is implemented by using quantum particle swarm optimization (QPSO)
to tune the structure of model and other hyperparameters/parameters. The inner loop uses an Adam optimizer to
tune the weights and biases of the proposed model. Spatiotemporal wind speeds at various locations provide the
2D input data. Simulation results reveal that the proposed method outperforms other methods for 24 h-ahead
wind speed forecasting.

1. Introduction energy trading; and long-term forecasts are used to optimize unit
commitment (Wang et al., 2021).
Reducing carbon dioxide emissions has become extremely important The three types of wind speed forecasting models that have been
as part of achieving the sustainable development goals (SDG) (particu­ proposed in the literature are physical approaches, statistical methods
larly goals 7 and 13) that were aggreed by members of the United Na­ and artificial intelligence (AI)-based models (Zhang et al., 2019). These
tions. To achieve these goals, a rapid transition from fossil fuel are detailed below.
consumption, which contributes to climate change, must be made. Wind
energy is an integral part of this effective environmental and energy (1) Physical approaches to wind forecasting include numerical
solution (IRENA, 2019). Wind is varying and intermittent, complicating weather prediction (NWP) (Xu et al., 2015; Buhan et al., 2016).
the operation of power systems, and particularly unit commitment and NWP solves dynamics and physics equations that express the
the scheduling of day-ahead generation. Developing highly accurate movements and variations of the atmosphere. Generally, super­
forecasts of renewable energy output may significantly minimize the computers are required to solve the complicated NWP equations.
risk of power system failures (Wang et al., 2018; Yang et al., 2020; Wu NWP is suitable for long-term wind speed forecasting.
et al., 2022). (2) Statistical methods that are used to forecast wind energy gener­
Wind speed forecast models can be categorized into very-short-term ation include probabilistic forecasting (Heng et al., 2022), poly­
(a few seconds to 30 min in advance), short-term (30 min to 6 h in nomial autoregressive modeling (Karakuş et al., 2017), and
advance), medium-term (6 h to 1 day in advance), and long-term (1 day sparse Bayesian learning (Wang et al., 2019). However, these
to 1 week or more in advance). Very-short-term forecasts are used to techniques do not have adaptive or learning capabilities.
control turbines and track loads in real time; short-term forecasts are (3) The most popular methods for wind power/speed forecasting that
used to plan load dispatch; medium-term forecasts are used to schedule are based on AI are machine learning and deep learning. Ahmadi
traditional power generation to satisfy the load demand and enable et al. (2020) presented a tree-based XGBoost algorithm to train

* Corresponding author.
E-mail addresses: yyhong@ee.cycu.edu.tw (Y.-Y. Hong), g11002803@cycu.edu.tw (C.L.P.P. Rioflorido), mszhangwn@shiep.edu.cn (W. Zhang).

https://doi.org/10.1016/j.eswa.2023.122645
Received 9 June 2023; Received in revised form 15 November 2023; Accepted 15 November 2023
Available online 22 November 2023
0957-4174/© 2023 Elsevier Ltd. All rights reserved.
Y.-Y. Hong et al. Expert Systems With Applications 241 (2024) 122645

model using the mean and standard deviation of wind speeds problems (Li et al., 2019; Liu et al., 2019; Fallahi and Taghadosi, 2022;
measured at a height of 40 m. Zhang et al. (2015) used the Deep Kuranga and Pillay, 2022).
Boltzmann Machine (DBM), pretrained by the greedy algorithm This work presents a hybrid deep learning and quantum-inspired
and then fine-tuned by the back-propagation algorithm, to study neural network to improve the accuracy of the 24-hour ahead wind
several minutes- to days-ahead wind speed forecasting. Nair et al. speed forecasting. Two aspects of quantum-inspired computing are used.
(2017) firstly handled the linear features of raw wind speed These are (i) a Quantum-Inspired Neural Network (QINN) incorporating
datasets by the Autoregressive Integrated Moving Average deep-learning networks and (ii) a variant of particle swarm optimization
(ARIMA) and then used Artificial Neural Network (ANN) to (PSO), called quantum PSO (QPSO), which is used to optimize the
perform nonlinear mapping. Sun et al. (2020) used variational hybrid deep learning (CNN in conjunction with LSTM) and QINN. The
mode decomposition (VMD) to excavate implicit information in contributions of this paper can be summarized as follows.
original wind speed datasets and then used the random forest to
select the suitable inputs; they subsequently used various AI 1) A novel hybrid deep-learning and QINN is presented to perform 24-
techniques, such as Generalized Regression Neural Networks hour-ahead wind speed forecasting. The deep-learning network
(GRNN), to generate forecasts. Araya et al. (2020) used dense consists of parallel CNNs cascading with LSTM. The parallel CNNs
layers of a neural network to build sub-sequences of wind speed extract the wind speed features from the spatiotemporal data, while
data with various timescales which were used as inputs to mul­ the LSTM interprets those features.
tiple Long Short-Term Memory (LSTM-M) to conduct 24 h-ahead 2) Historical geographic (spatial) and temporal data from three ancil­
forecasting. Zhu et al. (2020) consolidated a unified framework lary locations and one target location are used to generate the 2D
integrated Convolutional Neural Network (CNN, which selects dataset used in the proposed method, which does not require addi­
spatial feature) with LSTM (which processes temporal depen­ tional signal pre-processing.
dence) to conduct wind speed forecasting. Wang et al. (2020) 3) To determine the number of dimensions of the 2D data, a maximum
used CNN to carry out short-term wind speed forecasting, using cross-correlation analysis is performed to account for the maximum
information from neighboring wind farms. These methods do not time lag of ancillary locations to be used as inputs with respect to the
require a predetermined mathematical model and can handle target location.
multiple inputs and outputs, and can extract the relevant features 4) The QPSO is used to determine the structural parameters and
of data. hyperparameters of the CNN, LSTM, and QINN simultaneously. This
QPSO keeps the swarm particles from converging prematurely and
Artificial intelligence is the most popular basis for forecasting wind becoming trapped at local optima.
speed; however, related methods have some limitations as follows. 5) The proposed method introduces a novel approach by replacing the
traditional fully-connected layer in a CNN with a QINN, leveraging
(1) The accuracy of regression results that are obtained by tree-based complex-number weights to enhance accuracy and expand the search
machine learning algorithms depends on the split criterion dimensionality of the neural network. This innovative technique, to
(Ahmadi et al., 2020), which must be set by trial-and-error. the best of the authors’ knowledge, marks the first application of a
(2) Only a supervised learning algorithm is implemented without QINN in 24-hour-ahead forecasting, surpassing the performance of
extracting features from datasets, resulting in long training time the conventional fully-connected layer with real-number weights.
(Zhang et al., 2015) and probably low accuracy (Ahmadi et al.,
2020; Zhang et al., 2015). This paper is organized as follows. Section 2 provides the theoretical
(3) Only temporal dependence is considered and the spatial corre­ background of the CNN, LSTM, PSO, and quantum computing. Section 3
lation of wind characteristics is ignored (Ahmadi et al., 2020; elucidates the proposed method with the QPSO, QINN, and network
Zhang et al., 2015; Nair et al., 2017; Sun et al., 2020; Araya et al., architecture. Section 4 presents comparative results that are obtained
2020). using the proposed method and the methods in the literature. Section 5
(4) Data pre-processing (such as by using ARIMA or VMD) is required draws conclusions.
(Nair et al., 2017; Sun et al., 2020).
(5) The hyperparameters/parameters of machine/deep learning 2. Theoretical background
models are set to fix the structure of predictive model, based on
experiences (Zhang et al., 2015; Nair et al., 2017; Sun et al., 2020; 2.1. Multi-Step spatiotemporal forecasting
Araya et al., 2020; Zhu et al., 2020;Wang et al., 2020).
Multi-step spatiotemporal forecasting presents a complex computa­
Whereas various techniques for doing computational learning are tional challenge that has garnered significant academic interest in recent
readily available, quantum computing represents a paradigm shift in times. This specialized forecasting approach is primarily focused on
computing technologies. Several quantum algorithms are believed to accurately predicting future states of variables over various spatial lo­
solve problems more efficiently than classical approaches, motivating cations and time periods. It takes into careful consideration the intricate
the use of quantum computing herein. The developers of classical al­ spatial and temporal dependencies within the data. This forecasting
gorithms and quantum algorithms can exploit the strengths of both sets technique has found applications across a wide spectrum of fields,
of algorithms and thereby deepen our understanding of computational including but not limited to its use in forecasting sea temperatures (Kim
limitation (Nielsen and Chuang, 2011). The research landscape has seen et al., 2023), predicting traffic patterns (Ta et al., 2022), and estimating
the development of various quantum-inspired neural networks (QINNs) renewable energy generation (Chai et al., 2020; Zhu et al., 2020).
in response to a wide range of computational challenges. These QINNs This paper utilizes the multi-horizon spatiotemporal forecasting
have been explored and refined in multiple studies (Ganjefar and framework to improve the accuracy of wind speed predictions. The
Tofighi, 2018; Konar et al., 2020; Patel et al., 2021; Yan et al., 2021; Shi methodology presented in this paper delves into the complexities of this
et al., 2022). These endeavors highlight the significant potential and forecasting approach with the goal of enhancing the precision and
value that quantum-inspired design principles bring to the realm of dependability of wind speed forecasts. This contributes to the reliability
neural networks. On the other hand, Quantum Particle Swarm Optimi­ of forecasting models used in the wind energy sector and related fields.
zation has emerged as a robust strategy for optimization tasks, drawing In this study, historical data from specific locations in China, the
inspiration from quantum principles. It excels in navigating complex Philippines, and Taiwan are utilized to forecast wind speeds 24 h ahead
parameter spaces and finding optimal solutions for diverse optimization at a target wind farm in Taiwan.

2
Y.-Y. Hong et al. Expert Systems With Applications 241 (2024) 122645

2.2. Convolutional neural network (CNN) Equations (3) and (4) involve determining what new information
will be stored in the cell state. First, it determines which values are to be
A typical convolutional neural network comprises of the input layer, updated by using a sigmoid function in Eq. (3). Then, the tanh function
the convolution layer, the pooling layer, and the fully connected (FC) generates a vector of new values, C ̃ t , in Eq. (4). To update the old cell
layer. The original input data are pre-processed into a one-dimensional state, Ct-1, into the new cell state, Eq. (5) is applied to yield a value that is
or multi-dimensional matrix in the forms of the input. The convolution stored in long-term memory, Ct.
operation is used to construct feature maps of input data (Wang et al., A sigmoid layer determines which parts of the cell state will be
2020). output as described by Eq. (6). Then the tanh function shifts the values of
The number of feature maps is determined by the number of cell state from − 1 to 1, and the output is again multiplied to yield a
convolution kernels, while the size of the feature maps is determined by sigmoid in Eq. (7). The symbol ot is the result that is applied to the
the size of the convolution kernels. The value at position (p,q) in the output gate. The forget gate, input gate, and output gate all follow the
feature map is calculated using Eq. (1): sigmoid function, which yields a value between zero and one, effectively
(

Mr ∑ Mc
) providing the ability to control each gate. The special symbol “⊙” in­
fmpq = σ w(p− r)(q− c) ζcr + bpq dicates element-wise multiplication.
r=1 c=1

{ 2.4. Particle swarm optimization (PSO)


1 ≤ p ≤ Mr + Nr
subject to (1)
1 ≤ q ≤ Mc + Nc Kennedy and Eberhat first proposed PSO in 1995. Owing to its ease of
implementation and fewer controls compared to other optimization al­
where ζcr is an element of the input matrix of Mr × Mc; wpq is the weight
gorithms, the PSO algorithm has attracted the attention of many aca­
element of the convolution kernel of size Nr × Nc; bpq is a bias, and σ is an
demics over the past decade. The sociological behavior of flocks of birds
activation function, which is typically a ReLU or sigmoid function. and schools of fish inspired the PSO algorithm (Kennedy and Eberhart,
The average/maximum pooling layer reduces the number of pa­ 1995; Dai et al., 2018).
rameters of the network and prevents overfitting during model training. The position vector and velocity vector of the i th particle in the
The average/maximum of the elements in the feature map is calculated search space can be expressed as Xi and Vi , respectively. In the kth
using average/maximum pooling. iteration of the PSO algorithm, the particle velocity and position are
The dropout layer, which prevents overfitting, randomly sets the given as follows (Kennedy and Eberhart, 1995; Dai et al., 2018).
input units to zero at a frequency of “Rate” in each step throughout the ( )
training phase. Inputs that are not set to zero are scaled up by 1− 1Rate so Vi (k + 1) = ωVi (k) + ϛ1 r1 (Pi − Xi (k) ) + ϛ2 r2 Pg − Xi (k) (8)
that the sum of all inputs remains constant.
In this paper, a combined convolutional layer and an average pooling Xi (k + 1) = Xi (k) + Vi (k + 1) (9)
layer are defined as a convolutional neural network. The FC layer is
where ω is the inertia weight, ϛ1 and ϛ2 are the self-learning coefficient
replaced by the LSTM and QINN.
and global learning coefficient of the ith particle, respectively; r1 and r2
are random numbers within [0,1]; Pi is the best known position of the ith
2.3. Long Short-Term memory (LSTM) particle, and Pg is the best known position of the swarm. The swarm is
comprised of positions of the individual particles.
LSTM is an extended version of recurrent neural network (RNN) and In this paper, the elements of Xi are the unknown parameters and
has three gates, which are input, output, and forget. The forget gate hyperparameters of the proposed model.
regulates the flow of data by adjusting the data threshold between
reservation and oblivion. Since the hidden layer of the RNN has only one 2.5. Quantum computing
state, gradient disappearance and gradient explosion are serious prob­
lems. LSTM adds the structure of the cell state to the basic RNN, enabling Quantum computing is based on quantum mechanical postulates and
long-term preservation of the cell state and supporting the powerful properties (such as qubits, interference, superposition, and entangle­
memory function of the LSTM network. LSTMs have been shown to be ment) for information processing. Quantum computing enables the
effective for forecasting a variety of time series data, including elec­ solving of complicated problems faster than can be achieved using
tricity demand, stock prices, and weather patterns. (Hossain and Mah­ traditional computing (Nielsen and Chuang, 2011; Lou and Cong, 2011).
mood, 2020; Lu et al., 2023). Let Wf , Wi , WC and Wo be the weight
matrices, and bf , bi , bC and bo be the bias vectors. Let t be the time step. 2.5.1. Qubits
The basic formulas for an LSTM are as follows. A qubit is a tiny unit that is used to process information in a quantum
( ) computer; it is analogous to a bit in classical computing. A qubit can also
ft = σ Wf • [ht− 1 , xt ] + bf (2) be in one state, zero state, or both states at the same time, which is a
phenomenon known as linear superposition. Qubits can be realized
it = σ(Wi • [ht− 1 , xt ] + bi ) (3) using a variety of physical systems, including photons with two polari­
zations, nuclear spins in magnetic fields, and electrons in ground and
̃ t = tanh(WC • [ht− 1 , xt ] + bC )
C (4) excited states. In quantum mechanics, the developments of closed sys­
tems fulfill the Schrödinger equation:
̃t
Ct = ft ⊙ Ct− 1 + it ⊙ C (5)
iℏ|ψ̇ 〉 = H|ψ 〉 (10)
ot = σ(Wo [ht− 1 , xt ] + bo ) (6)
where the state function |ψ 〉 is equivalent to the either of the states |0〉
ht = ot ⊙ tanh(Ct ) (7) and |1〉 which are the eigenstates of the Hamiltonian system. The pure
state of a single qubit may be represented as follows using the principles
Specifically, Eq. (2) indicates that the forget gate controls the amount of of superposition of quantum states:
history flows that is included the current history. ft is the result that is
|ψ 〉 = α|0〉 + β|1〉 (11)
applied to the forget gate; xt is the current input vector, ht is the hidden
vector which stores current outputs in a short-term memory.

3
Y.-Y. Hong et al. Expert Systems With Applications 241 (2024) 122645

when α and β are complex eigenstate coefficients and the squares of their where the δ enables the flipping of the phase information. The function
absolute values represent the probability of matching eigenstates under
f(x) can be expressed as eix , where the argument of f(x) is equal to − θ
measurements. They satisfy probability completeness:
when δ = 0 and 2π − θ when δ = 1.
‖α‖2 + ‖β‖2 = 1 (12)
3. Proposed method

3.1. Determination of time lag using Cross-Correlation analysis


Thus, Eq. (11) can be modified to be
As well as the time-series at the target location (the offshore wind
( ) farm Fuhai near Taiwan, denoted as L1 in Table 1), wind speeds at
|ψ 〉 = e iφ iζ
cosθ|0〉 + e sinθ|1〉 (13)
multiple wind farms (at locations L2, L3 and L4 in Table 1) are used to
provide spatial information. The selection of L2, L3 and L4 is based on
Since the global phase factor φ has no discernible effect on the compu­
the correlation values of the wind speed data and also the distance be­
tation of state function, Eq. (13) has an analogous form:
tween the target wind farm and L2, L3 and L4. Fig. 2 illustrates the
|ψ 〉 = cosθ|0〉 + eiζ sinθ|1〉 (14) spatial distribution of these specific locations.
Most of the wind farms near Fuhai have few wind turbines or low
where θ and ζ define a point on a unitary three-dimensional Bloch generation capacity. Only three other wind farms are considered.
sphere, as shown in Fig. 1. Table 1 provides information on the target and three others, including
the number of turbines and MW capacity of each (Wind Energy Data­
2.5.2. Quantum rotation gate base, 2022). Table 2 specifies wind speed statistics for the four locations
The logic function in quantum computation can be implemented by for one year.
applying a sequence of uniform transformations to the qubit states, the To study cross-correlation between the target and ancillary wind
effect of which is identical to that of the logic gate. Therefore, quantum speeds, Pearson’s rank-order correlation, which yields the degree and
services that involve logical changes at a certain interval of states are direction of association between two rank-order variables, is used. Since
called quantum gates and are the basis for performing quantum calcu­ the wind farms are located over a large region, time lags or delays
lations. A single rotation gate is defined by Eq. (15). The rotation can be (denoted as λ) must be considered in the cross-correlation study
visualized as in Fig. 1. (Bechrakis and Sparis, 2004).
[ ] For example, if both time-series S1(η) and S2(η-λ) include ℵ scalar
cosθ − sinθ
Ry (θ) = (15) observations (η = 1, 2, …, ℵ), then the correlation coefficient ρ(λ) be­
sinθ cosθ
tween S1(η) and S2(η-λ) is given by Eq. (17), where μ1 and μ2 are the
means of S1(η) and S2(η-λ), respectively.
2.5.3. Quantum-Inspired CNOT (Controlled NOT) gate
∑ℵ
This gate is a conditional gate that applies an X-gate on the second η=1 (S1 (η) − μ1 )(S2 (η − λ) − μ2 )
qubit (target) if the first qubit (control) is in state |1〉. A quantum- ρ(λ) = ∑√ ̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ (17)
ℵ 2 ∑ℵ 2
η=1 (S1 (η) − μ1 ) η=1 (S2 (η − λ) − μ2 )
inspired CNOT gate is represented by Eq. (16).
(π ) {
cosθ − isinθ (δ = 0) Assume that the wind speed time-series at L1 is S1(η) and that at L2,
f δ− θ = (16) L3 or L4 is S2(η-λ). The initial λ may be set to 25 (beginning the day
2 sinθ + icosθ (δ = 1)
following the target hours) and increased by one repeatedly to 97 (the
end of the fourth day). The maximum time lag for L2, L3 or L4 with
respect to L1 is the maximum of the λs that corresponds to the maximum
of ρ(λ) at L2, L3 and L4 with respect to L1. This process is called
maximum cross-correlation analysis in this paper. This maximum time
lag is set to the number of rows of 2D spatiotemporal dimensional data
(the number of columns is fixed at four because four locations are
considered).

3.2. Wind speed statistics of Fuhai and Cross-Correlation

Historical wind speeds at four locations were considered. Hourly


wind speeds (m/s) were measured from May 1, 2017, to April 30, 2018,
yielding a total of 8760 data points per location. Taiwan has four distinct
seasons - winter, spring, summer, and autumn. The dataset is stratified
into four distinct seasonal partitions. Subsequently, each seasonal subset
undergoes a process of partitioning into training and testing subsets,
adhering to an 80–20 split ratio, where 80 % of the data is allocated for
training and validation purposes, while the remaining 20 % is reserved
for testing.

Table 1
Target site and three other sites.
Names (label) Locations No. of Turbines Power (MW)

Fuhai, Taiwan (L1) (24.03◦ ,120.25◦ ) 30 120.0


Yueshan, China (L2) (22.43◦ ,112.51◦ ) 128 76.8
Burgos, Philippines (L3) (18.53◦ , 120.64◦ ) 29 87.0
Kuochangshan, China (L4) (28.81◦ ,120.93◦ ) 35 21.3
Fig. 1. Qubit on Bloch sphere.

4
Y.-Y. Hong et al. Expert Systems With Applications 241 (2024) 122645

Table 3
Characteristics of wind speed in Fuhai, Taiwan.
Season Mean (m/ Standard deviation Maximum (m/ Minimum (m/
s) (m/s) s) s)

Winter 13.01 4.57 27.08 0.12


Spring 8.88 5.48 25.47 0.25
Summer 5.55 3.15 20.28 0.12
Autumn 7.29 5.19 25.47 0.12

Table 4
Time lag (h) obtained from maximum cross-correlations among locations.
Season L1 L2 L3 L4

Winter 25 25 25 71
Spring 25 25 25 70
Summer 25 95 25 47
Fall 25 45 97 47

deep learning and quantum model generates next 24 wind speeds.


Therefore, for each season, the maximum cross-correlation values with
the largest time lag are used to obtain the time lag for the proposed
method. The time lag varies by location and time of year because the
Fig. 2. Target wind farm L1 and wind farms at three ancillary locations L2 wind direction varies with locations and times of year. In summary, the
~ L4. time lags in the winter, spring, summer and autumn are 71, 70, 95 and
97 h, and the numbers of dimensions of the 2D spatiotemporal data are
71 × 4, 70 × 4, 95 × 4 and 97 × 4, respectively. Specifically, for the
Table 2 winter, the inputs of the proposed hybrid deep learning and quantum
Wind statistics of target site and three other sites.
model are wind speeds at t-71, t-70, …, t at L1 ~ L4 while the outputs are
Names (label) Mean Standard Maximum Minimum wind speeds at t + 1, t + 2, …, t + 24 at L1.
(m/s) deviation (m/s) (m/s)
(m/s)
3.3. Quantum particle swarm optimization (QPSO)
Fuhai, Taiwan (L1) 8.68 4.68 27.08 0.12
Yueshan, China (L2) 3.16 1.85 19.74 0.18
QPSO improves upon the standard PSO. The dynamic behavior of the
Burgos, Philippines 6.92 4.04 25.84 0.18
(L3) individual particle in QPSO differs significantly from that of an indi­
Kuochangshan, 2.02 5.78 8.24 0 vidual particle in conventional PSO due to the presence of uncertainty,
China (L4) primarily associated with particle movement rather than particle posi­
tion (Wang et al., 2008; Sun et al., 2004). Let the position of the ith
particle Xi = [χ i1 , χ i2 , ⋯, χ id , ⋯, χ iD ]. The best known position of the ith
The histogram in Fig. 3 shows the sampled wind speeds at the target
particle Pi = [pi1 , pi2 , ⋯, pid , ⋯, piD ] and the best known position of the
wind farm throughout the winter season. Table 3 shows the wind speed
swarm Pg = [pg1 , pg2 , ⋯, pgd ⋯, pgD ] r1 , r2 and r3 are three random
characteristics at the target wind farm. They suggest that the Fuhai
numbers that are sampled in the range of [0,1] for each particle
offshore wind farm has high potential for power generation owing to its
movement.
high mean wind speed value per season.
The dth element of particle Xi in QPSO moves in a D-dimensional
The cross-correlation analysis involves two locations. The target
space following the iterative equation,
location L1 (Fuhai) is compared individually to L2 (Yueshan, China), L3
( )
(Burgos, Philippines), and L4 (Kuocangshan, China). Table 4 shows the 1
results of the cross-correlation analysis to determine the best time lags χ id (k + 1) = pd ± L⋅ln (18)
r3
(λ) for the 2D spatiotemporal data. The value of λ yields the number of
rows in the 2D spatiotemporal data while the number of columns is fixed where
at four.
1
A time lag that exceeds 24 is preferred because the proposed hybrid L= (19)
g⋅|χ id (k) − pd |

r1 pid + r2 pgd
pd = (20)
r1 + r2
√̅̅̅
where g is a constant and g ≥ ln 2. If r3 > 0.5, then a “plus” sign is used
in Eq. (18); else, a “minus” sign is used (Wang et al., 2008; Sun et al.,
2004).
QPSO is as a powerful optimization method, notably employed for
the determination of structural parameters and hyperparameters of the
proposed neural network architectures, encompassing CNN, LSTM, and
QINN. QPSO holds several distinct advantages (Fallahi and Taghadosi,
2022):

Fig. 3. Histogram of wind speeds in winter at target wind farm.

5
Y.-Y. Hong et al. Expert Systems With Applications 241 (2024) 122645

1) QPSO can search intricate, high-dimensional, and non-linear 3.4.2. Hidden neurons of QINN
parameter spaces, such as multiple layers, units, and hyper­ The hidden layer is a rotation layer implemented by CNOT gates. The
parameters in complex configurations of neural networks. CNOT gate R(θ) is defined as follows:
2) QPSO integrates quantum mechanics with swarm intelligence, (π ) {
cosθ − isinθ (δ = 0)
achieving a balance between the exploration of diverse configura­ R(θ) = f δ − θ = (24)
2 sinθ + icosθ (δ = 1)
tions and the exploitation of promising ones. This facilitates the
discovery of optimal settings for neural network designs. In Eq. (24), the parameter δ plays a pivotal role in the transformation
process. Specifically, when δ is set to 0, the input data undergo a
3.4. Quantum-Inspired neural network (QINN) transformation where their magnitudes remain unaltered, and the phase
information is effectively neutralized. In contrast, when δ equals 1, a
A QINN is based on the two concepts of quantum computing and significant alteration occurs: the real and imaginary components of the
artificial neural networks (Shi et al., 2022). Fig. 4 shows the proposed input data are interchanged, mirroring a bitwise operation akin to the
quantum-inspired network. flipping mechanism elucidated in Eq. (13). It is worth emphasizing that
the assignment of δ is not deterministic but rather stochastic, with its
3.4.1. Input neurons of QINN value for each hidden neuron being chosen randomly. This deliberate
The input to this network is a complex number that represents the randomness introduces a crucial element of exploration within the
eigenstate coefficients in Eq. (11). All neurons in the first layer include a model. By allowing δ to vary unpredictably, the model gains the capacity
function that multiplies 2π to the input. A 2π pulse sends |0〉→|0〉+i|1〉
√̅̅ to the to probe diverse transformations, enabling it to decipher intricate pat­
2
terns within the data. This stochasticity serves as a strategic tool,
input layer. In this layer, the input scalar numbers xq1, xq2, …, xqM are
empowering the model to delve into multifaceted data nuances, ulti­
rescaled into the range [0, 1], and multiplies by π/2 to yield the phase of
mately enhancing its ability to discern and learn intricate, non-linear
the complex input in the range [0, π/2].
patterns.
An arbitrary one-qubit state can be expressed as Eq. (11) based on the
Let the numbers of input and hidden neurons be M and N,
fundamental quantum principle of state superposition. The revised
respctively, m = 1, 2, …, M and n = 1, 2, …, N. Eq. (25) represents the
equation for a qubit state is defined in Eq. (14).
complex output of a hidden neuron:
Phase factors are ignored in Eq. (14) and the magnitude information
is incorporated as a complex number, yielding a qubit-inspired repre­ ∑
M

sentation of data that is similar to Euler’s formula: Hn = R(θ) eiθm , n = 1, ⋯, N (25)


m=1

f (θ) = eiθ = cosθ + isinθ (21)


3.4.3. Output neurons of QINN
Euler’s formula provides the refined qubit state vector Let the symbol o and O represent index of the output neuron and total
[ ]
cosθ number of neurons or units in the complex dense layer, respectively. The
|ψ 〉 = cosθ|0〉 + eiϕ sinθ|1〉 = iϕ (22) weights {β1,1 , ..., βn,o , ⋯, βN,O } are assumed to have real values because
e sinθ
they are easier to use in the commonly used machine learning frame­
A single qubit state is known to correspond to a point on the surface
work. Complex-valued networks have greater representation capacity.
of a Bloch sphere with coordinates (θ, φ). The evolution of a single qubit
The complex dense layer (with neurons CD1, CD2, … in Fig. 4) processes
state can be modeled as a vector rotation on the Bloch sphere. In this
the complex output of Eq. (25), generating the output yo :
case, Eq. (21) corresponds to a ring on the surface of a Bloch sphere; data
that are obtained using Eq. (21) correspond to a rotation in the direction ∑
N
( )
yo = f βn,o Hn − f (bo ) (26)
of latitude. Thus, the data in Eq. (21) can be processed in a manner that
n=1
is comparable to quantum state evolution. Specifically,
U⋅f (θ0 ) = U⋅(cosθ0 + isinθ0 ) = cos(θ + θ0 ) + isin(θ + θ0 ) (23) where bias b is also a real number and o is the index of the output
neuron, o = 1, 2, …, O. The computation in Eq. (26) is similar to that of a
( )
where U = eiθ and f(θ0 ) represents another complex number. Then this classical neuron of a neural network but the weights f βn,o and bias
operation yields ei(θ+θ0 ) . This operation describes the transformation of f(bo ) are complex numbers, which are associated with Eq. (21). Eq. (27)
data points on the surface of a Bloch sphere. Through the combined defines the final output of a neuron:
rotations represented by U and f(θ0 ), a new data point emerges on the
output = |Im(yo ) |2 (27)
Bloch sphere, indicating the altered state of the data after successive
transformations. This formula encapsulates the essence of quantum- where Im(yo ) represents the imaginary part of the complex number yo .
inspired data evolution (Shi et al., 2022).

3.5. Architecture for proposed method and QPSO search space

The block diagrams in Fig. 5 present the architecture for the pro­
posed method. The input data are prepared as four columns (corre­
sponding to four locations) of two-dimensional data, with a time lag (λ)
that is determined by maximum cross-correlation analysis.
Fig. 5 presents an example of the proposed model for the spring. The
three CNN layers in parallel, associated with average pooling layers, are
connected to an LSTM. The QINN, which consists of a rotation (hidden)
layer and a complex dense layer, is cascaded to the LSTM. Several
possible architectures are available for the CNN layer: (a) one CNN
layer, (b) two CNN layers, (c) two CNN layers connected in parallel, (d)
three CNN layers, and (e) three CNN layers connected in parallel. Any
one of these options is cascaded to the LSTM and the QINN, as deter­
Fig. 4. Quantum-inspired neural network. mined by QPSO.

6
Y.-Y. Hong et al. Expert Systems With Applications 241 (2024) 122645

Fig. 5. Block diagram of the proposed method for spring season.

The proposed method involves two iterative loops. QPSO optimizes used.
the architecture of the proposed hybrid deep learning with QINN, while
the ADAM optimizer tunes the weights and biases. ADAM, based on
4.1. Performance indices
adaptive estimates of lower-order moments, is an algorithm for first-
order gradient-based optimization to minimize an objective function.
Three performance indices are obtained using various seasonal data
QPSO is used to optimize the architecture, number of kernels, kernel size
and trials; they are the Root Mean Squared Error (RMSE), Mean Absolute
and dropout value of CNN, the number of output units in LSTM, and the
Error (MAE), and R-Squared (R2). Let y, ̂ y and y be the forecasted values,
parameters of the QINN. An output of the QINN consists of 24 nodes that
the mean of the forecasted values and the actual values, respectively. Let
correspond to the forecasted wind speeds for the next 24 h.
K be the number of studied statistical data points, κ=1, 2, …, K. The
Table 5 presents a list of hyperparameters and parameters that can be
formulas for RMSE, MAE and R2 are given as follows.
used as elements in a particle, along with their respective limits. Filters 1
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
to 3 are allocated to three convolution layers, respectively, and they are √
√1 ∑ K
dependent on the structures mentioned earlier. The variables filter RMSE = √ (̂y κ − yκ )2 (28)
lengths 1 to 3 represent the kernel size, which specifies the height and K κ=1
width of the 2D convolution window. Dropout values are presented in
floating-point format, rounded to two decimal places. The LSTM unit 1 ∑K
MAE = y κ − yκ |
|̂ (29)
refers to the number of inner cells within the LSTM. The two QINN layers K κ=1
consist of the rotation layer and the complex dense layer, respectively.
The number of gates or units per layer in the QINN is optimized through SSE
R2 = 1 − (30)
the use of QPSO. The parameters for the neural architecture are deter­ SST
mined by QPSO.
∑ ∑
where SSE = y − y)2 and SST =
(̂ (y − y)2 .
4. Simulation results and discussions
4.2. PSO Comparison results
The proposed method was implemented in Python using the Keras
package with TensorFlow as the backend. A workstation that was used As mentioned in the preceding section, the proposed method in­
to run the simulations had an Ubuntu 18 operating system with an Intel volves two iterative loops. The outer loop uses QPSO to optimize the
Core i7-11700 processor at 2.40 GHz and 47.0 GB of RAM. Two graphics architecture, number of kernels, kernel size and dropout value of the
cards (NVIDIA GeForce RTX 3080 Ti) with 10 GB of RAM each were CNN, number of output units in the LSTM, and parameters of the QINN.
This method finds the optimal solution using agents, known as particles,
Table 5 whose orbits are modified by using a stochastic and deterministic
Hyperparameters/parameters to be optimized and their limits. component. The second optimizer (inner loop) is the ADAM, which
Parameters Limits optimizes weights and biases.
Structure (a, b, c, d, e)
The performances of the basic PSO (BPSO) (Kennedy and Eberhart,
Filter 1 1, 2, …, 5 1995) and two PSO variants (Chaos PSO (CPSO) (Song et al., 2007) and
Filter 2 1, 2, …, 5 QPSO (Sun et al., 2004)) are compared by running the aforementioned
Filter 3 1, 2, …, 5 outer loop optimization. Each run has 50 iterations and MAE is used as a
Filter Length 1 1 × 1, 2 × 2, …, 5 × 5
fitness function. For all seasons, the number of particles is 35.
Filter Length 2 1 × 1, 2 × 2, …, 5 × 5
Filter Length 3 1 × 1, 2 × 2, …, 5 × 5 Tables 6 to 9 show the optimal hyperparameters/parameters that are
Dropout 0 ~ 0.50 obtained using three PSO algorithms for the four seasons. QPSO yields
LSTM Units 100, 101, …, 500 the lowest MAE, which is the fitness function of three PSO algorithms.
No. of gates in QINN Layer 1 100, 101, …, 500 Based on the results presented in these tables, QPSO dynamically adapts
No. of gates in QINN Layer 2 100, 101, …, 500
its hyperparameters and parameters for different seasons in order to

7
Y.-Y. Hong et al. Expert Systems With Applications 241 (2024) 122645

Table 6 seasons, showcasing its adaptability and effectiveness in optimizing


Optimal hyperparameters/parameters for winter season. network architectures for seasonal data.
Parameters BPSO CPSO QPSO Figs. 6-9 display the iteration performances of three PSOs for each
season. QPSO consistently converges to the lowest MAE and has been
Structure c e e
Filter 1 4 3 1 proven to be the most effective optimization method for addressing the
Filter 2 4 5 5 specified problem. Notably, during the summer and autumn seasons,
Filter 3 – 4 1 while CPSO and QPSO exhibit very similar MAE values, QPSO demon­
Filter Length 1 5×5 3×3 5×5 strates a faster convergence rate than CPSO. This accelerated conver­
Filter Length 2 4×4 5×5 5×5
Filter Length 3 – 5×5 5×5
gence pattern further confirms the superior performance of QPSO when
Dropout 0.47 0.50 0.41 compared to the other two PSO algorithms.
LSTM Units 427 500 419
QINN Layer 1 354 498 184
QINN Layer 2 373 117 264 4.3. Comparison of results
MAE 0.0286 0.0276 0.0275

In this section, simulation results that are obtained by the proposed


method are compared with those obtained using six other methods and
Table 7 other benchmark models. Specifically, XGBoost (Ahmadi et al., 2020),
Optimal hyperparameters/parameters for spring season.
DBM (Zhang et al., 2015), ARIMA-ANN (Nair et al., 2017), GRNN (Sun
Parameters BPSO CPSO QPSO et al., 2020), LSTM-M (Araya et al., 2020), and CNN-LSTM (Zhu et al.,
Structure e e e 2020) are used for hourly wind speed forecasting. The hyperparameters
Filter 1 3 4 1 and parameters in these models are replicated as documented in the
Filter 2 2 5 5 cited papers. Subsequently, employing the same dataset, these models
Filter 3 5 2 5
undergo a process of reconstruction, enabling the day-ahead spatio­
Filter Length 1 2×2 3×3 4×4
Filter Length 2 4×4 4×4 4×4 temporal wind speed forecasting, which can be used for further
Filter Length 3 3×3 4×4 4×4 comparative analysis. Benchmark models, including Backpropagation
Dropout 0.50 0.50 0.43 Neural Network (BPNN), 1D CNN, Long- and Short-term Time-series
LSTM Units 427 499 295 Network (LSTNET), and Gated Recurrent Unit (GRU), are harnessed for
QINN Layer 1 424 499 192
QINN Layer 2 327 104 149
comparative analysis. The proposed method is also compared with
MAE 0.0305 0.0299 0.0292 equivalent architectures employing a fully connected layer (CNN-LSTM-
FC) and a complex-valued neural network without quantum inspiration
(CNN-LSTM-CVNN) as a subsequent connection after the LSTM layer.
Table 8
These benchmark models constitute conventional tools for assessing
Optimal hyperparameters/parameters for summer season. errors and R2 metrics in the realm of time-series forecasting and
regression problems.
Parameters BPSO CPSO QPSO
Tables 10-13 provide a rigorous assessment of these forecasting
Structure c e e methods for various seasons, comparing their accuracy in terms of MAE,
Filter 1 4 5 5
RMSE, and R2. Lower MAE and RMSE values indicate more accurate
Filter 2 5 3 2
Filter 3 – 1 1 forecasting methods, while higher R2 values indicate better fits to the
Filter Length 1 4×4 5×5 5×5 data.
Filter Length 2 4×4 1×1 4×4 The proposed method achieves lower RMSE and MAE values
Filter Length 3 5×5 1×1

compared to the other methods, indicating its superior accuracy in
Dropout 0.5 0.5 0.47
LSTM Units 417 499 187
predicting wind speeds. The proposed method consistently produce R2
QINN Layer 1 285 100 168 values exceeding 0.92 in each season, surpassing the R2 values obtained
QINN Layer 2 271 232 214 by other methods. This indicates a strong correlation between the pre­
MAE 0.0260 0.0248 0.0248 dicted and actual wind speed data, highlighting the reliability and
robustness of the proposed approach. It is evident from the above ob­
servations that the proposed method consistently outperforms the other
Table 9 methods across all seasons.
Optimal hyperparameters/parameters for autumn season. The significant improvement offered by the proposed method
Parameters BPSO CPSO QPSO compared to other methods in the literature is noteworthy. This
Structure c e e
Filter 1 4 5 4
Filter 2 3 4 2
Filter 3 – 5 3
Filter Length 1 5×5 5×5 3×3
Filter Length 2 2×2 5×5 5×5
Filter Length 3 – 3×3 4×4
Dropout 0.5 0.49 0.48
LSTM Units 467 500 479
QINN Layer 1 458 500 370
QINN Layer 2 355 500 344
MAE 0.0227 0.0224 0.0224

optimize neural network models. It demonstrates a tendency to favor


relatively compact filter configurations, dropout rates, and LSTM units
but makes adjustments based on the specific characteristics of each
season. QPSO consistently achieves competitive MAE values across all Fig. 6. Comparison of PSO variants applied to winter dataset.

8
Y.-Y. Hong et al. Expert Systems With Applications 241 (2024) 122645

Table 11
Comparison of results for spring season.
Method MAE RMSE R2

BPNN 0.1665 0.1998 0.0005


1D CNN 0.0741 0.0985 0.7758
LSTNET 0.0351 0.0591 0.9144
GRU 0.0372 0.0626 0.9105
XGBoost 0.1035 0.1401 0.5305
DBM 0.1154 0.1513 0.4968
ARIMA-ANN 0.0402 0.0641 0.9198
GRNN 0.1144 0.1529 0.4594
LSTM-M 0.0327 0.0523 0.9240
CNN-LSTM 0.0420 0.0672 0.9142
CNN-LSTM-FC 0.0381 0.0559 0.9275
CNN-LSTM-CVNN 0.0316 0.0507 0.9401
Proposed Method 0.0292 0.0491 0.9439
Fig. 7. Comparison of PSO variants applied to spring dataset.

Table 12
Comparison of results for summer season.
Method MAE RMSE R2

BPNN 0.1254 0.1564 − 0.0064


1D CNN 0.0617 0.0811 0.7636
LSTNET 0.0315 0.0450 0.8952
GRU 0.0270 0.0421 0.9225
XGBoost 0.0834 0.1125 0.4814
DBM 0.0913 0.1197 0.3949
ARIMA-ANN 0.0415 0.0594 0.8916
GRNN 0.0842 0.1103 0.4861
LSTM-M 0.0309 0.0443 0.9069
CNN-LSTM 0.0251 0.0498 0.8937
CNN-LSTM-FC 0.0299 0.0456 0.9114
CNN-LSTM-CVNN 0.0266 0.0422 0.9236
Proposed Method 0.0248 0.0406 0.9297
Fig. 8. Comparison of PSO variants applied to summer dataset.

Table 13
Comparison of results for autumn season.
Method MAE RMSE R2

BPNN 0.1645 0.2029 0.0125


1D CNN 0.0492 0.0657 0.9021
LSTNET 0.0363 0.0418 0.9105
GRU 0.0234 0.0372 0.9676
XGBoost 0.0719 0.0984 0.7998
DBM 0.0778 0.1045 0.7238
ARIMA-ANN 0.0277 0.0405 0.9166
GRNN 0.0753 0.1018 0.7594
LSTM-M 0.0277 0.0404 0.9364
CNN-LSTM 0.0263 0.0407 0.9045
CNN-LSTM-FC 0.0369 0.0501 0.9378
CNN-LSTM-CVNN 0.0244 0.0359 0.9598
Proposed Method 0.0224 0.0347 0.9708

Fig. 9. Comparison of PSO variants applied to autumn dataset.


indicates that incorporation of QINN elements, particularly the complex
weights among the gates and rotation layer, has a substantial positive
Table 10
Comparison of results for winter season. impact on the accuracy of wind speed forecasts. These complex weights
may capture intricate relationships within the data that conventional
Method MAE RMSE R2
methods or models lacking such elements cannot obtain a high accuracy.
BPNN 0.1358 0.1690 − 0.00002 In the winter season, accurate predictions are often challenging due
1D CNN 0.0614 0.0807 0.7642
to the unique characteristics of cold weather. The QPSO method dem­
LSTNET 0.3654 0.0501 0.9021
GRU 0.0315 0.0486 0.9166 onstrates a commendable RMSE of 0.0431, showcasing its capability to
XGBoost 0.0903 0.1213 0.4897 model complex seasonal dynamics. However, it’s noteworthy that in this
DBM 0.1038 0.1321 0.4076 specific season, the CNN-LSTM method exhibits superior performance,
ARIMA-ANN 0.0360 0.0523 0.9101 achieving an RMSE of 0.0407.
GRNN 0.0993 0.1269 0.4503
The proposed method achieves the highest R2 value, exceeding 0.97,
LSTM-M 0.0359 0.0499 0.9152
CNN-LSTM 0.0315 0.0407 0.8249 during the autumn season. This can be attributed to the frequent
CNN-LSTM-FC 0.0373 0.0520 0.9029 occurrence of intermediate to high wind speeds during this season. This
CNN-LSTM-CVNN 0.0304 0.0449 0.9275 superior forecasting accuracy may stem from the proposed method’s
Proposed Method 0.0274 0.0431 0.9333
capacity to capture intricate patterns and fluctuations in wind speed.
Conversely, the R2 values are at their lowest during the summer

9
Y.-Y. Hong et al. Expert Systems With Applications 241 (2024) 122645

season. This could be attributed to the presence of typhoons and the work reported in this paper.
generally low wind speeds in summer, which can be challenging to
predict accurately. Nevertheless, it’s noteworthy that the proposed Data availability
method continues to outperform other models even during this season,
which further underscores its effectiveness. Data will be made available on request.
The proposed method was compared with equivalent architectures
employing a fully connected layer (CNN-LSTM-FC) and a complex- Acknowledgements
valued neural network without quantum inspiration (CNN-LSTM-
CVNN) as a subsequent connection after the LSTM layer. As a result, the This work is sponsored by National Science and Technology Council,
proposed method continues to outperform these two models in terms of Taiwan, under the grant NSTC 112-2218-E-008-011 and NSTC 112-
MAE, RMSE, and R2. 2221-E-033-010-MY3.
In summary, the results presented in Tables 10-13 conclusively
demonstrate that the proposed method, bolstered by the QINN with References
complex weights, excels in wind speed forecasting across all seasons. Its
sustained achievement of lower MAE and RMSE values, coupled with Ahmadi, A., Nabipour, M., Mohammadi-Ivatloo, B., Amani, A. M., Rho, S., & Piran, M. J.
higher R2 values, serves as robust validation of its superiority over other (2020). Long-term wind power forecasting using tree-based learning algorithms.
IEEE Access, 8, 151511–151522. https://doi.org/10.1109/access.2020.3017442
methods. The integration of QINN significantly contributes to capturing Araya, I. A., Valle, C., & Allende, H. (2020). A multi-scale model based on the long short-
intricate relationships within the data, thereby establishing it as a term memory for Day Ahead hourly wind speed forecasting. Pattern Recognition
valuable tool for precise wind speed prediction in diverse weather Letters, 136, 333–340. https://doi.org/10.1016/j.patrec.2019.10.011
Bechrakis, D. A., & Sparis, P. D. (2004). Correlation of wind speed between neighboring
conditions. measuring stations. IEEE Transactions on Energy Conversion, 19(2), 400–406. https://
The average training time is 16.47 s for a season using NVIDIA CUDA doi.org/10.1109/tec.2004.827040
cores in an RTX3080 Ti, with a batch size of 2000 and 10,000 iterations. Buhan, S., Ozkazanc, Y., & Cadirci, I. (2016). Wind pattern recognition and reference
wind mast data correlations with NWP for improved wind-electric power forecasts.
Utilization rate of the GPU is 70 % on average.
IEEE Transactions on Industrial Informatics, 12(3), 991–1004. https://doi.org/
10.1109/tii.2016.2543004
5. Conclusion and future works Chai, S., Xu, Z., Jia, Y., & Wong, W. K. (2020). A robust spatiotemporal forecasting
framework for photovoltaic generation. IEEE Transactions on Smart Grid, 11(6),
5370–5382. https://doi.org/10.1109/tsg.2020.300608
This work develops a hybrid deep learning and quantum-inspired Dai, H. P., Chen, D. D., & Zheng, Z. S. (2018). Effects of random values for particle swarm
neural network to increase the accuracy of 24-hour-ahead wind speed optimization algorithm. Algorithms, 11(2), 23. https://doi.org/10.3390/a11020023
forecasting. Essential innovations and findings are summarized as Fallahi, S., & Taghadosi, M. (2022). Quantum-behaved particle swarm optimization
based on solitons. Scientific Reports, 12, 13977. https://doi.org/10.1038/s41598-
follows. 022-18351-0
Ganjefar, S., & Tofighi, M. (2018). Optimization of quantum-inspired neural network
(a) The dimensionality of the spatiotemporal 2D data is determined using memetic algorithm for function approximation and chaotic time series
prediction. Neurocomputing, 291, 175–186. https://doi.org/10.1016/j.
by both the number of wind farms and the maximum cross- neucom.2018.02.074
correlation approach. Heng, J., Hong, Y., Hu, J., & Wang, S. (2022). Probabilistic and deterministic wind speed
(b) QPSO can generally obtain better hyperparameters/parameters forecasting based on non-parametric approaches and wind characteristics
information. Applied Energy, 306, Article 118029. https://doi.org/10.1016/j.
of the proposed hybrid model than BPSO and CPSO algorithms. apenergy.2021.118029
QPSO also converges faster and yields a smaller MAE than the Hossain, M. S., & Mahmood, H. (2020). Short-term photovoltaic power forecasting using
other two PSO algorithms. an LSTM neural network and synthetic weather forecast. IEEE Access, 8,
172524–172533. https://doi.org/10.1109/access.2020.3024901
(c) Three CNN layers, connected in parallel, are found to be more Irena. (2019). Future of wind: Deployment, investment, technology, grid integration and socio-
effective than a single CNN layer for extracting data features. One economic aspects (A Global Energy Transformation paper). Abu Dhabi: International
or more filters with fixed filter lengths are required in each CNN Renewable Energy Agency.
Karakuş, O., Kuruoğlu, E. E., & Altınkaya, M. A. (2017). One-Day ahead wind speed/
layer.
power prediction based on polynomial autoregressive model. IET Renewable Power
(d) The integration of LSTM with QINN to replace the traditional Generation, 11(11), 1430–1439. https://doi.org/10.1049/iet-rpg.2016.0972
fully-connected layer to produce the next 24 wind speeds has a Kennedy, J., & Eberhart, R. (1995). Particle swarm optimization. Proceedings of ICNN’95
promising accuracy. The QINN can obtain a larger subspace of - International Conference on Neural Networks. https://doi.org/10.1109/
icnn.1995.488968.
complex numbers for weight tuning, improving the accuracy of Kim, J., Kim, T., Ryu, J.-G., & Kim, J. (2023). Spatiotemporal graph neural network for
wind speed prediction. multivariate multi-step ahead time-series forecasting of sea temperature. Engineering
(e) The proposed hybrid deep learning and quantum-inspired neural Applications of Artificial Intelligence, 126, Article 106854. https://doi.org/10.1016/j.
engappai.2023.106854
network outperforms XGBoost, DBM, ARIMA-ANN, GRNN, Konar, D., Bhattacharyya, S., Gandhi, T. K., & Panigrahi, B. K. (2020). A quantum-
LSTM-M, CNN-LSTM, CNN-LSTM-FC and CNN-LSTM-CVNN inspired self-supervised network model for automatic segmentation of brain MR
based on application to wind speed forecasting in four seasons. images. Applied Soft Computing, 93, Article 106348. https://doi.org/10.1016/j.
asoc.2020.106348
Kuranga, C., & Pillay, N. (2022). A comparative study of nonlinear regression and
Future research will delve into diverse quantum neural network ar­ autoregressive techniques in hybrid with particle swarm optimization for time-series
chitectures, encompassing quantum convolutional neural networks forecasting. Expert Systems with Applications, 190, Article 116163. https://doi.org/
10.1016/j.eswa.2021.116163
(QCCNs) and quantum recurrent neural networks (QRNNs), with the
Li, Y., Xiao, J., Chen, Y., & Jiao, L. (2019). Evolving deep convolutional neural networks
objective of attaining even higher accuracy in predictions. The study by quantum behaved particle swarm optimization with binary encoding for image
will also investigate the influence of additional meteorological and classification. Neurocomputing, 362, 156–165. https://doi.org/10.1016/j.
neucom.2019.07.026
environmental variables, such as wind direction and the roughness of
Liu, G., Chen, W., Chen, H., & Xie, J. (2019). A quantum particle swarm optimization
the Earth’s surface, on the accuracy of forecasting. Such a comprehen­ algorithm with teamwork evolutionary strategy. Mathematical Problems in
sive consideration will be designed to capture the intricate spatiotem­ Engineering, 2019, 1–12. https://doi.org/10.1155/2019/1805198
poral dynamics that govern wind speed fluctuations. Lou, Y., & Cong, S. (2011). State transfer control of quantum systems on the Bloch
sphere. Journal of Systems Science and Complexity, 24(3), 506–518. https://doi.org/
10.1007/s11424-011-9066-4
Declaration of competing interest Lu, K. D., Wu, Z. G., & Huang, T. (2023). Differential Evolution-based three stage
dynamic Cyber-Attack of cyber-physical power systems. IEEE/ASME Transactions on
Mechatronics, 28(2), 1137–1148. https://doi.org/10.1109/tmech.2022.3214314
The authors declare that they have no known competing financial Nair, K. R., Vanitha, V., & Jisma, M. (2017). Forecasting of wind speed using ann, Arima
interests or personal relationships that could have appeared to influence and hybrid models. In 2017 International Conference on Intelligent Computing,

10
Y.-Y. Hong et al. Expert Systems With Applications 241 (2024) 122645

Instrumentation and Control Technologies (ICICICT). https://doi.org/10.1109/ Wang, Y., Zou, R., Liu, F., Zhang, L., & Liu, Q. (2021). A review of wind speed and wind
icicict1.2017.8342555 power forecasting with Deep Neural Networks. Applied Energy, 304, Article 117766.
Nielsen, M. A., & Chuang, I. L. (2011). Quantum Computation and Quantum Information. https://doi.org/10.1016/j.apenergy.2021.117766
Cambridge University Press. Wang, Z., Zhang, J., Zhang, Y., Huang, C., & Wang, L. (2020). Short-term wind speed
Patel, O. P., Bharill, N., Tiwari, A., & Prasad, M. (2021). A novel quantum-inspired fuzzy forecasting based on information of neighboring wind farms. IEEE Access, 8,
based neural network for data classification. IEEE Transactions on Emerging Topics in 16760–16770. https://doi.org/10.1109/access.2020.2966268
Computing, 9(2), 1031–1044. https://doi.org/10.1109/tetc.2019.2901272 Wind energy database. (2022). https://www.thewindpower.net/index.php.
Shi, S., Wang, Z., Cui, G., Wang, S., Shang, R., Li, W., … Gu, Y. (2022). Quantum-inspired Wu, Y. K., Wu, Y. C., Chang, H. L., & Hong, J. S. (2022). Using extreme wind-speed
complex convolutional neural networks. Applied Intelligence, 52(15), 17912–17921. probabilistic forecasts to optimize unit scheduling decision. IEEE Transactions on
https://doi.org/10.1007/s10489-022-03525-0 Sustainable Energy, 13(2), 818–829. https://doi.org/10.1109/tste.2021.3132342
Song, Y., Chen, Z., & Yuan, Z. (2007). New chaotic PSO-based neural network predictive Xu, Q., He, D., Zhang, N., Kang, C., Xia, Q., Bai, J., & Huang, J. (2015). A short-term wind
control for nonlinear process. IEEE Transactions on Neural Networks, 18(2), 595–601. power forecasting approach with adjustment of numerical weather prediction input
https://doi.org/10.1109/tnn.2006.890809 by data mining. IEEE Transactions on Sustainable Energy, 6(4), 1283–1291. https://
Sun, J., Feng, B., & Xu, W. (2004). Particle swarm optimization with particles having doi.org/10.1109/tste.2015.2429586
quantum behavior. In Proceedings of the 2004 Congress on Evolutionary Computation Yan, P., Li, L., Jin, M., & Zeng, D. (2021). Quantum probability-inspired graph neural
(IEEE Cat. No.04TH8753). https://doi.org/10.1109/cec.2004.1330875 network for document representation and classification. Neurocomputing, 445,
Sun, N., Zhang, S., Peng, T., Zhou, J., & Sun, X. (2020). A composite uncertainty 276–286. https://doi.org/10.1016/j.neucom.2021.02.060
forecasting model for unstable time series: Application of wind speed and streamflow Yang, M., Zhang, L., Cui, Y., Zhou, Y., Chen, Y., & Yan, G. (2020). Investigating the wind
forecasting. IEEE Access, 8, 209251–209266. https://doi.org/10.1109/ power smoothing effect using set pair analysis. IEEE Transactions on Sustainable
access.2020.3034127 Energy, 11(3), 1161–1172. https://doi.org/10.1109/tste.2019.2920255
Ta, X., Liu, Z., Hu, X., Yu, L., Sun, L., & Du, B. (2022). Adaptive spatio-temporal graph Zhang, C. Y., Chen, C. L., Gan, M., & Chen, L. (2015). Predictive deep boltzmann machine
neural network for traffic forecasting. Knowledge-Based Systems, 242, Article 108199. for multiperiod wind speed forecasting. IEEE Transactions on Sustainable Energy, 6(4),
https://doi.org/10.1016/j.knosys.2022.108199 1416–1425. https://doi.org/10.1109/tste.2015.2434387
Wang, X., Chen, J., Wu, Z., & Pan, F. (2008). Modeling of fermentation process based on Zhang, L., Dong, Y., & Wang, J. (2019). Wind speed forecasting using a two-stage
QDPSO-SVM. In 2008 Fourth International Conference on Natural Computation. forecasting system with an error correcting and nonlinear ensemble strategy. IEEE
https://doi.org/10.1109/icnc.2008.176 Access, 7, 176000–176023. https://doi.org/10.1109/access.2019.2957174
Wang, Y., Wang, H., Srinivasan, D., & Hu, Q. (2019). Robust functional regression for Zhu, Q., Chen, J., Shi, D., Zhu, L., Bai, X., Duan, X., & Liu, Y. (2020). Learning temporal
wind speed forecasting based on sparse bayesian learning. Renewable Energy, 132, and spatial correlations jointly: A unified framework for wind speed prediction. IEEE
43–60. https://doi.org/10.1016/j.renene.2018.07.083 Transactions on Sustainable Energy, 11(1), 509–523. https://doi.org/10.1109/
Wang, Y., Zhou, Z., Botterud, A., & Zhang, K. (2018). Optimal wind power uncertainty tste.2019.2897136
intervals for electricity market operation. IEEE Transactions on Sustainable Energy, 9
(1), 199–210. https://doi.org/10.1109/tste.2017.2723907

11

You might also like