Renewable Energy 246 (2025) 122945

Contents lists available at ScienceDirect

Renewable Energy
journal homepage: [Link]/locate/renene

Short-term multi-site solar irradiance prediction with

dynamic-graph-convolution-based spatial-temporal correlation capturing
Haixiang Zang *,1, Wenan Li , Lilin Cheng **,1 , Jingxuan Liu , Zhinong Wei , Guoqiang Sun
School of Electrical and Power Engineering, Hohai University, Nanjing, 210098, China

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

Keywords: To optimize the utilization of photovoltaic power sources, high-precision short-term solar irradiance predictions
Multi-site solar irradiance prediction are critically needed. This study presents a novel short-term solar irradiance prediction method utilizing dynamic
Deep learning graph convolution, which captures spatial-temporal correlations across multiple sites. The method takes his­
Attention mechanisms
torical solar irradiance and meteorological data as input. Initially, an adaptive feature selection approach based
Dynamic graph convolution
on squeeze-and-excitation networks (SENets) is introduced to assign appropriate weights to various meteoro­
logical features. Then, a self-attention mechanism is integrated into the adjacency matrix construction method
for graph convolutional networks (GCNs), allowing for the extraction of spatial features. Finally, the spatial-
temporal features extracted through a bi-directional long short-term memory (BiLSTM) network are used for
short-term prediction. Experimental results demonstrate that the proposed hybrid model exhibits superior
generalization capability and greater prediction accuracy compared to the benchmark methods, significantly
improving the accuracy of short-term global horizontal irradiance (GHI) prediction.

1. Introduction and machine learning (ML) models. However, these models only take
into account historical irradiance data and often ignore the influence of
Driven by economic development, the global demand for traditional other meteorological factors on solar irradiance. Instead, ML methods
resources such as oil and coal is growing. The environmental challenges can combine multiple meteorological variables to predict comprehen­
and non-renewable nature of these resources pose increasing risks to sively. Popular ML methods include artificial neural networks (ANNs)
daily life [1]. To mitigate these concerns, there is an urgent need to [7] and support vector machines [8], which have been widely applied
transition to cleaner energy sources. Among the various forms of clean for solar irradiance prediction [9]. As a sub-field of ML, deep learning
energy, the environmental protection and long-term availability of solar exhibits superior predictive capabilities compared to traditional
energy lead to its widespread use in power systems [2]. However, the methods. This approach has demonstrated strong performance in solar
output of photovoltaic systems is highly dependent on solar irradiance, irradiance prediction [10,11]. Studies have shown that models such as
which is inherently variable and unpredictable [3]. When photovoltaic long short-term memory (LSTM) networks and gated recurrent units
systems are integrated with the power grid, the output fluctuation of (GRUs) improve prediction accuracy by effectively capturing spatial
photovoltaic systems can threaten the reliability of the grid [4]. As a features [12,13]. This combination of convolutional neural networks
result, accurate short-term solar irradiance prediction is crucial for the (CNNs) with LSTM has been particularly effective in extracting spatio­
expansion of photovoltaic applications, as it aids in power system temporal features [14,15]. For example, Michael et al. [16] developed a
scheduling and helps to manage fluctuations effectively [5]. bidirectional long short-term memory (BiLSTM)-GRU algorithm to
Consequently, scholars from energy disciplines have increasingly handle multiple variables, such as weather conditions and irradiance.
focused on leveraging effective techniques for solar irradiance predic­ The algorithm achieved lower error rates compared to other models.
tion. Existing solar irradiance prediction methods can be broadly cate­ Deep learning methods frequently incorporate attention mechanisms to
gorized into three approaches: physical models [6], statistical models, enhance performance [17]. Jonathan et al. [18] integrated attention

* Corresponding author.
** Corresponding author.
E-mail addresses: zanghaixiang@[Link] (H. Zang), straw@[Link] (L. Cheng).
1
Present address: School of Electrical and Power Engineering, Hohai University, Nanjing 211,100, China.

[Link]
Received 22 October 2024; Received in revised form 18 March 2025; Accepted 20 March 2025
Available online 21 March 2025
0960-1481/© 2025 Elsevier Ltd. All rights are reserved, including those for text and data mining, AI training, and similar technologies.
H. Zang et al. Renewable Energy 246 (2025) 122945

mechanisms within convolutional neural networks (ATT_CNN) to boost multi-head attention mechanism, and temporal features from multiple
the prediction efficiency of global horizontal irradiance (GHI), direct sites are extracted using BiLSTM. The key contributions of this study
normal irradiance (DNI) and diffuse horizontal irradiance (DHI). Liang include the following.
et al. [19] applied an attention mechanism to train stacked dilated
convolutional networks (DCNs) for ultra-short-term renewable energy • A graph-based spatio-temporal feature construction method is
prediction. developed based on multi-site meteorological parameters and his­
In spite of the outstanding predictive capabilities of the above torical solar irradiance. The method can efficiently combine multi-
methods, most prior research has focused on single-site scenarios. site data into the united features, so as to reduce the learning diffi­
Additionally, advances have been made in multi-site data prediction culty of the prediction model.
across various fields, such as multi-site photovoltaic prediction [20], • Considering multi-sectional contributions of different factors, the
multi-site air pollution prediction [21], and multi-site domestic waste SENet module is introduced to modify the prediction model. Spe­
generation prediction [22]. However, multi-site solar irradiance pre­ cifically, it quantifies the importance contribution of different fea­
diction remains an under-explored area, highlighting the importance of tures to solar irradiance prediction at different time steps by
this paper’s focus on multi-site short-term GHI prediction. Given the adaptively generating time-varying feature weight matrix.
non-Euclidean nature of spatial information in multi-site data, tradi­ • In order to pay attention to the changes in the relationship between
tional Euclidean models like CNNs are ill-suited for spatial relationship sites, a dynamic graph convolution is innovatively proposed. Dy­
modeling [23]. Conversely, graph neural networks (GNNs) and graph namic graph convolution dynamically generates dynamic adjacency
convolutional networks (GCNs) excel in this domain by leveraging graph matrices at different time steps through a self-attention mechanism,
data and adjacency matrices to encode site dependencies. This study thus capturing the changing characteristics of spatial correlations
transforms site feature matrices into non-Euclidean graph structures to among different sites.
facilitate spatial information extraction. GCNs have demonstrated their
efficacy in extracting spatial features across diverse applications, The structure of the paper is as follows. Section 2 outlines the con­
including wind power prediction [24], photovoltaic power generation struction process of the proposed method and explains the principles
prediction [25], and traffic flow prediction [26]. For example, Zang behind each network. Section 3 details the datasets, performance eval­
et al. [27] utilized GCNs to extract spatial features from multiple time uation metrics, and benchmark methods used in this experiment. Section
series and constructed separate LSTM networks for each site to capture 4 compares the performance of the proposed method with that of
temporal features. Gao et al. [28] proposed a dynamic spatial-temporal benchmark methods and provides an analysis of additional capabilities.
interactive graph neural network (DSTIGNN), which used a dynamic Finally, Section 5 presents the conclusion of this paper.
graph convolution module to model and capture dynamic spatial cor­
relations between variables, facilitating the propagation of information 2. Methodology
in the spatial dimension. Using GCN, Xu et al. [29] leveraged geographic
information and dynamic time-warped (DTW) distances to construct an 2.1. Multi-site solar irradiance prediction framework
optimal weighted map and fully extract spatial features from the data. In
many GCN-based applications, adjacency matrices are computed based The hybrid method proposed in this study is designed to predict solar
on the geographical distance and location of sites. However, such irradiance across multiple sites by first performing adaptive feature se­
methods fail to capture the dynamic nature of spatial correlations be­ lection on input features, then extracting spatial relationships between
tween sites due to climate change or other factors. Furthermore, these stations, followed by time-series feature extraction for each station, and
methods do not account for the evolving relationships between adjacent finally generating GHI predictions for each site. Assuming there are N
sites. Therefore, improving adjacency matrix computation methods is sites, with each site having historical data for M features spanning T time
crucial for enhancing the performance of GCNs in multi-site prediction steps. The input dataset is a collection of M feature sequences across time
tasks. T for N sites. The dataset can be described as follows:
In addition to the accurate extraction of spatial information, effective ⎧ ⎫
⎪ [ m] ⎪
feature selection based on input variables plays a pivotal role in ⎪ z(1)m
⎪
⎪ , …, z(1) m
, …, z(1) ⎪
⎪
⎪
⎪ 1 t T ⎪
enhancing multi-site solar irradiance prediction performance. Feature ⎪
⎪
⎪…
⎪
⎪
⎪
⎪
⎨[ ⎪
⎬
selection has proven essential in various fields, including probabilistic m m m ]
Z= z(n)1 , …, z(n)t , …, z(n)T (m = 1, 2, …, M), (1)
prediction of electric vehicle (EV) charging demand [30] and wind ⎪
⎪ ⎪
⎪
⎪ … ⎪
power prediction [31]. For example, Zhang et al. [32] employed a ⎪
⎪[
⎪
⎪
⎪
⎪
⎪ m m m] ⎪
⎩ z(N)1 , …, z(N)t , …, z(N)T ⎪
⎪ ⎪
feed-forward neural network (FNN) to optimize feature subsets by ⎪ ⎭
considering their contribution to prediction errors, thereby refining
feature selection for improved accuracy. Gao et al. [33] utilized a where Z represents the input dataset of the method. z(n)m
t denotes the
combinatorial fuzzy strategy to preprocess initial solar irradiance data
dataset including m features for site n at time step t.
and applied a multi-objective optimization algorithm to search for
Based on the input dataset Z, the prediction process for the subse­
optimal parameters. Similarly, Castangia et al. [34] implemented mul­
quent time step can be formulated as:
tiple feature selection techniques to determine the most relevant input
variables for solar irradiance prediction. However, most of these feature P
̃t+1 = [P(1)
̃
t+1 , …, P(N)t+1 ] = f(Z)
̃ (2)
selection methods establish fixed importance parameters at the outset or
eliminate less important features outright [35]. This approach may not where P(n)
̃
t+1 signifies the predicted GHI value for site n at time step t +
fully capture the dynamic influence of various features on prediction 1, and P
̃t+1 represents the set of predicted GHI values for all sites at the
outcomes, particularly as these features change over time. Conse­ same time. f( ⋅) denotes a prediction task aimed at minimizing prediction
quently, there is a critical need to develop more advanced feature se­ errors.
lection methods. The complete process of the prediction task f( ⋅) is shown in Fig. 1. As
Inspired by the above methods, this study introduces a novel illustrated in Fig. 1, reconstruct the multi-site data firstly. The squeeze
approach. Firstly, a squeeze-and-excitation network (SENet) is used to excitation network (SENet) is used to adaptively weight the features.
adaptively select features. In addition, a dynamic adjacency matrices, Then, the attention mechanism is used to generate a dynamic adjacency
leveraging the self-attention mechanism, is proposed to enable GCN to matrix for the dynamic GCN to obtain graph data with spatial features.
better capture spatial features. Lastly, spatial features are optimized by

2
H. Zang et al. Renewable Energy 246 (2025) 122945

Fig. 1. The framework of the proposed hybrid GHI prediction method.

Finally, the multi-head attention mechanism was used to optimize the series. Meanwhile, M represents the number of features at each site,
spatial features, and the BiLSTM was used to extract the temporal fea­ which includes historical solar irradiance and meteorological parame­
tures. The subsequent subsections provide detailed explanations of each ters.
process involved. To reconstruct the spatial characteristics across multiple stations, the
original data from N stations is transformed into a three-dimensional
2.2. Construction of multi-site data data matrix based on climate type, succinctly expressed as
O ∈ RT×N×M . The process of converting multi-site data into a three-
The input utilized in this study consists of two-dimensional (2D) dimensional (3D) data matrix and displaying feature matric at time
matrices representing climate parameters for each station, with the raw step t is illustrated in Fig. 2. Consequently, the 3D matrix construction
data for each station denoted as Z(n) ∈ RT×M . Here, T indicates the total process of input dataset Z can be represented by the following formula:
number of samples collected at each site, reflecting the length of the time

3
H. Zang et al. Renewable Energy 246 (2025) 122945

Fig. 2. Process of 3D matrix construction based on input dataset.

O = fstack [Z(n)](i = 1, 2, …, N), (3)

where Z(n) ∈ RT×M signifies the original 2D data matrix for the n-th site.
N corresponds to the total number of selected sites for a particular
climate type. fstack [ ⋅] denotes the integration of these 2D matrices into a
3D structure, specifically along the dimension representing sites. As
illustrated in Fig. 2, the independent 2D dataset of each site is recon­
structed into a 3D matrix shaped like T × N × M along the site dimen­
sion, based on the fstack [ ⋅] operation. It is used for the subsequent
dynamic GCN to extract the spatial information of the input matrix.

2.3. Adaptive feature selection

Fig. 3. Process of adaptive feature selection based on SENet.

Based on the 3D matrix shaped like T × N × M, solar irradiance is
influenced by a combination of M factors consisting of historical irra­
diance data and various meteorological. Each factor exerts a varying parameters of the two fully connected layers.
degree of influence on prediction outcomes. Moreover, the significance The feature selection weights St for a given time step t are derived
of the factors fluctuates over T time steps. Consequently, it is of utmost from the excitation output. The corresponding feature selection results
significance to utilize SENet to accurately and adaptively capture the xt are calculated by multiplying these weights by the original features
key features at each time step. This approach effectively highlights matrix Ot , as indicated in Equation (6):
important features while diminishing the impact of redundant ones, xt = Scale(Ot , St ) = St ⋅Ot , (6)
ultimately enhancing the model’s prediction performance. As an atten­
tion mechanism, SENet dynamically adjusts the focus on different fea­ where Scale( ⋅) is the reweight operation. In view of the foregoing, SENet
tures when processing the feature matrices of various sites, assigning generates normalized feature selection weights for each time step. After
feature importance, and weighting the feature information according to applying these weights and fusing the results using Equation (6), a
different levels of saliency [36]. Fig. 3 illustrates the process of adaptive feature sequence is produced that dynamically reflects the importance of
feature selection using SENet at each time step. various features over time.
As shown in Fig. 3, SENet generates adaptive weights for each
channel through squeeze-and-excitation operations. It applies these
2.4. Dynamic graph convolutional network
weights to each feature, thereby obtaining the results of adaptive feature
selection. SENet comprises two key components: squeeze and excitation.
2.4.1. Construction of dynamic adjacency matrices
During the squeeze operation, SENet applies average pooling across
In order to deal with the matrix after feature selection, it needs to be
each channel’s feature map, generating a value that reflects the impor­
described as a graph structure. In a network of nodes, a weighted un­
tance of the channel’s features. This process captures global information
directed graph is described as G = (V, E, A), where V = {v1 , v2 , …, vN }
for each channel. Given that the input matrix is O ∈ RT×N×M , the squeeze
represents the set of N nodes in the graph, and E is the set of edges. The
operation is as expressed in Equation (4).
adjacency matrix A ∈ RN×N quantitatively describes the correlation be­
1 ∑ H ∑ W tween different nodes. In this study, every site corresponds to a node in
zt = Ot (r, c), (4) G, and the elements in the adjacency matrix A provide edge weights
H × W r=1 c=1
between different stations, representing the spatial correlation between
where zt represents the result of extrusion at time step t. H and W are the sites. The feature matrix obtained through the feature selection process
height and width of the feature map, respectively. Ot is the input matrix based on the SENet is X ∈ RT×N×M consisting of M features of N sites. The
at time step t, with r and c indicating rows and columns. feature matrix at time step t is expressed as:
In the Excitation operation, the extruded features are adaptively
X = {xt }(t = 1, 2, …, T), (7)
weighted by passing through fully connected layers, allowing the
network to learn the importance of features in each channel dynami­ where xt ∈ RN×M represents the feature matrix at time step t. X is the set
cally. The process is represented by Equation (5): of feature matrices corresponding to all samples.
St = sigmoid(W2 δ(W1 zt )), (5) Adjacency matrices constructed from feature matrices are an effec­
tive way to describe spatial features in GCN. However, a fixed adjacency
where St is the incentive result, i.e., the weight of each channel at time matrix cannot account for the dynamic spatial correlation between
step t. δ( ⋅) is the RELU activation function. W1 ( ⋅) and W2 ( ⋅) are the feature matrices of each site over time [37]. This paper proposes a

4
H. Zang et al. Renewable Energy 246 (2025) 122945

dynamic adjacency matrix, incorporating an attention mechanism to able weight matrices of the GCN l-th layer. σ ( ⋅) is the activation func­
describe the dynamic correlations between feature matrices. The tion. C denotes the number of filters or channels in the GCN layer.
attention mechanism adaptively, rather than statically, calculates the Hl ∈ RN×M and Hl+1 ∈ RN×C represent the input and outputs of the GCN
fluctuating distribution of attention across sites and captures the in­ l-th layer, respectively.
teractions and influences between them. For a given input matrix xt , the Based on the above approximation, the calculation process of dy­
self-attention mechanism transforms the input xt into the query matrix namic stacked GCN in this paper is shown in Fig. 4. From Fig. 4, it can be
Qt , key matrix Kt , and value matrix Vt at time step t, as described by the observed that the graph data is initially input into the first convolutional
following equation: layer to aggregate the first-order site information. In this step, the model
captures the dynamic spatial dependencies of the sites with its neigh­
Qt = Wq xt
Kt = Wk xt , (8) boring sites. In the second step, to further aggregate the correlated solar
Vt = Wv xt irradiance data from nearby sites, the output of the first layer is passed
through a second convolutional layer. This layer aggregates the second-
where Wq , Wk , and Wv are the learnable parameter matrices for the order node information, allowing the model to learn broader spatial
linear transformations, respectively. The core concept of self-attention is correlations. After multiple graph convolution operations, each site
to calculate the similarity between input elements based on the inner captures both its autocorrelation characteristics and the spatial features
product of Qt and Kt . Therefore, the inter-element correlation of the of adjacent sites. When the activation function employs RELU, the
input matrix xt can be calculated similarly. By doing this, the dynamic process of the two-layer dynamic graph convolution model at time step t
adjacency matrix at different time steps is learned, allowing the spatial is expressed as follows:
features at varying time steps to be fully captured during model training.
Gt (A
̃ t , xt ) = ReLU(A
̃ t ReLU(A
̃ t xt W1 )W2 ), (11)
This also enables the model to explore the fluctuating correlation be­
tween sites. The dynamic adjacency matrix at time step t is computed as
where Gt is the output of dynamic GCN, W1 and W2 represent the
follows:
learnable weight matrices of the first and second GCN layers, respec­
( )
̃ t = Softmax Q tK
T
tively. xt is the feature matrix at the time step t and A
̃ t is the dynamic
A √̅̅̅̅t̅ , (9)
dk adjacency matrix at the time step t.

where dk is the dimension of Kt . The Softmax function normalizes the 2.5. Spatial-temporal feature optimization extraction module
attention scores to values between [0,1], yielding an adjacency matrix
A
̃ t that represents the site relevance at time step t. 2.5.1. Optimization of spatial features by multi-head attention mechanism
The feature optimization of the time series with spatial information is
2.4.2. Extraction of spatial features by dynamic GCN carried out. The output of dynamic GCN is used as the input of the multi-
On the basis of the above data preprocessing, dynamic GCN is head attention mechanism. As presented in Fig. 5, the multi-head
introduced in this paper. GCNs is widely employed for spatial feature attention mechanism combines multiple single-headed attentions to
extraction and have achieved significant success in image recognition.
GCNs are also effective in capturing the spatial distribution patterns
among stations with varying solar irradiance levels [38]. To enhance the
prediction accuracy of solar irradiance and fully extract spatial features,
the paper introduces dynamic GCN, which leverages the dynamic ad­
jacency matrix to capture spatial relationships across different time steps
among multiple sites. In GCN, convolution operations are performed by
constructing filters in the Fourier domain, which act on nodes and their
first-order neighbors to extract spatial features [39]. The GCN operation
can be approximated by expanding the first-order Chebyshev poly­
nomial as follows:
⎛ ⎞
1 1
Hl+1 = σ ⎝D− 2 AD− 2 Hl Wl ⎠, (10)

where A ∈ RN×N is the normalized adjacency matrix with self-loops, and

D represents the corresponding degree matrix. Wl ∈ RM×C is the learn­ Fig. 5. Process of multi-head attention mechanism.

Fig. 4. Implementation of a dynamic stacked GCN layer based on dynamic adjacency matrices.

5
H. Zang et al. Renewable Energy 246 (2025) 122945

compute the input sequence in parallel. This method enables the input to distinct climate types in the United States, each containing data from
sequence to be processed simultaneously across multiple single-headed eight stations. The dataset spans 24-h periods from January 1, 2018, to
attention mechanisms, which results in a final output value. Unlike December 31, 2022, resulting in a total of 43,800 samples per site. Of
single-head attention, multi-head attention applies different weights to this data, 80 % was allocated for training the model, while the remaining
the input sequence at various positions, allowing the model to capture 20 % served as the test set for evaluating model performance. The
multifaceted relationships between input sequence features. This en­ selected states for analysis were Texas, Montana, and Ohio, which
hances the model’s capacity to represent features more effectively [40]. represent the climate types: hot semi-arid (BSH), mid-latitude cold
One of the key advantages of, multi-head attention is its ability to ac­ desert (BSK), and hot summer humid continental (DFA). Table 1 pre­
count for correlation between different positions of the input features at sents the geographic information and climate type of each site. The
various time steps of the sequence [41]. Therefore, multi-head attention climatic features considered for prediction include GHI and clear sky
is introduced to optimize the feature matrix with spatial information, GHI (CSI) for each station, along with eight meteorological parameters:
strengthen the expression of spatial information, and better explain cloud type (CT), dew point (DP), solar zenith angle (SZA), wind direc­
spatial correlation. With G as input, the multi-headIt is calculated as tion (WD), wind speed (WS), relative humidity (RH), temperature (T),
follows: and precipitable water (PW).
(̂ ̂T )
QK
A
̂ = Softmax √ ̅̅̅̅̅ V,
̂ (12) 3.2. Evaluating metrics
dk
̂
To comprehensively assess the prediction performance of the model,
xopt several evaluation metrics were selected, including root mean square
t = MultiHead(G) = Concat( A 1 , A 2 , …, A head )WO , (13)
̂ ̂ ̂
error (RMSE), normalized root mean square error (nRMSE), mean ab­
where Q,̂ K,
̂ and V̂ are the query, key, and value matrix obtained by solute error (MAE), normalized mean absolute error (nMAE), and the
correlation coefficient (R). These metrics are calculated as follows:
linear transformation, respectively. ̂
d k is the dimension of K.
̂ Â is the
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
output of the single-head attention mechanism. head represents the 1 ∑N
number of heads in the multi-head attention mechanism. WO is the RMSE = (̃yi − yi )2 , (17)
N i=1
parameter matrix, and Concat( ⋅) is the concatenation operation.
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
̅
MultiHead( ⋅) signifies the output produced by the multi-head attention ∑N 2
i=1 (yi − yi )
1 ̃
N
mechanism. xoptt represents the output matrix after optimization of nRMSE = × 100%, (18)
yave
spatial features.
1 ∑N
2.5.2. Extraction of temporal features by BiLSTM MAE = yi − yi |,
|̃ (19)
N i=1
In addition to the extraction and optimization of spatial feature, the
∑N
extraction of temporal features is also crucial for enhancing the pre­ 1
i=1 |yi
̃ − yi |
diction performance of solar irradiance. Thus, BiLSTM is introduced to nMAE = N × 100%, (20)
yave
extract the time information of the optimized feature matrix xopt
t . Unlike
⃒∑N ⃒
traditional recurrent neural networks (RNNs), LSTMs incorporate three ⃒
i=1 (yi − yave )⋅(yi − yave )
̃ ̃ ⃒
gates—an input gate, an oblivion gate, and an output gate—along with a R = √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
[∑N ] [ , (21)
∑ 2]
̃ ̃ 2
i=1 (yi − yave ) ⋅
N
i=1 (yi − yave )
cell state. The BiLSTM model comprises a forward LSTM that processes
inputs in a forward direction and a backward LSTM that processes inputs
in a reverse direction. It employs two hidden layers to capture both Where ̃ yi and yi denote the i-th predicted and observed GHI values,
historical and future data, finally connecting to a common output layer. respectively. ̃
yave and yave represent the averaged predicted and observed
Therefore, BiLSTM can effectively capture temporal information from GHI values, respectively. N signifies the total number of observations.
both the past and future, thereby improving the generalization capa­
bility and predictive accuracy of the models. The procedure of the 3.3. Baselines and benchmark models
BiLSTM can be detailed as follows:
→ ( ̅̅→) Eight comparison models were utilized to assess the prediction per­
ht = σ1 ω1 xopt
t + ω2 ht− 1 , (14) formance of the proposed hybrid model, including CNN, GRU, CNN-
BiGRU, CNN-LSTM [15], CNN-BiLSTM-Attention [17], and AGCLSTM
( ←̅̅ )
←
ht = σ2 ω3 xopt [27]. These models have demonstrated strong performance in
t + ω4 ht+1 , (15)
short-term solar irradiance prediction. Specifically, CNN captures tem­
[→ ←] poral features [43], while GRU [44] and LSTM [45] are designed to
Ht = ht , ht , (16) extract spatial features. The CNN-BiGRU model integrates CNN and
GRU, creating a hybrid approach that enhances the network’s depth and
where xopt t
→ ←
is the input of the BiLSTM at time step t. ht and ht represent offers superior capabilities in extracting both temporal and spatial fea­
the output of the forward and backward LSTM layers at time step t, tures. Moreover, CNN-LSTM and CNN-BiLSTM-Attention are SOTA
respectively. ω1 , ω2 , ω3 , and ω4 are the weights of each LSTM layer. methods that employ CNN and LSTM for temporal feature extraction,
σ 1 ( ⋅) and σ 2 ( ⋅) are activation functions of each layer. Ht denotes the further allowing dynamic GCN to effectively capture the spatial corre­
output of the BiLSTM layer at time step t. lation between different sites. The primary distinction with AGCLSTM
lies in the way the adjacency matrix is calculated. AGCLSTM utilizes the
3. Experiment feature matrix from different time steps and the parameters of the dense
layer to determine the spatial information between sites, using
3.1. Datasets self-attention to describe these spatial correlations. This approach not
only reflects relationships between neighboring sites but also describes a
The dataset utilized in this study is sourced from the national solar site’s relationship with itself. The grid search method was employed to
radiation database (NSRDB) [42], comprising three sets corresponding select the optimal parameters for the proposed model. Tables 2 and 3
present the compositions and hyperparameters and their respective

6
H. Zang et al. Renewable Energy 246 (2025) 122945

Table 1
Geographic information and climate types of the selected sites in Texas (BSH), Montana (BSK), and Ohio (DFA).
Climate type Location Latitude range (◦ ) Longitude range (◦ ) Mean elevation (m)

Min latitude Max latitude Min longitude Max longitude

BSH Texas 28.25 28.65 − 99.94 − 99.30 177.875

BSK Montana 48.18 48.54 − 109.65 − 110.11 959.25
DFA Ohio 39.88 40.24 − 82.93 − 83.25 266.375

Table 2
Compositions and hyperparameters of the benchmark methods.
Model Layers Hyperparameters Model Layers Hyperparameters

CNN Conv2D filters = 64, kernel size = 3 × 3 GRU 2 GRU neurons = 64

Maxpooling pool size = 2
Flatten –
Fully-connected neuron = 128
CNN-BiGRU Conv2D filters = 64, kernel size = 3 × 3 CNN-LSTM Conv2D filters = 64, kernel size = 3 × 3
2 BiGRU neuron = 64 2 LSTM neuron = 64
Fully-connected neuron = 128 Fully-connected neuron = 128
Concatenate neuron = [256, 128] Concatenate neuron = [256, 128]
2 Fully-connected 2 Fully-connected
CNN-BiLSTM-Attention Conv2D filters = 64, kernel size = 3 × 3 AGCLSTM 2 GCN 64
BiLSTM neuron = 64 2 LSTM 64
Self-attention 32 MLP [256,128,64], dropout = 0.2

Table 3 Table 4
Composition and hyperparameters of the proposed model. Average prediction errors of different models under three climate types.
Layers Parameters Climate Model RMSE nRMSE MAE nMAE R
type (W/m2) (%) (W/m2) (%)
2 GCN neuron = 64
activation function: ReLU BSH CNN 57.902 5.590 32.143 14.101 0.9830
Multi-Head Attention head numbers = 5 GRU 57.012 5.504 27.311 11.982 0.9833
neuron = 64 CNN- 52.465 5.065 22.511 9.877 0.9859
2 BiLSTM neuron = 64 BiGRU
MLP neuron = [512, 128] CNN-LSTM 52.549 5.073 22.425 9.838 0.9859
Dropout rate = 0.2 activation function: ReLU CNN- 51.861 5.007 21.865 9.593 0.9863
Batch size 512 BiLSTM-
Learning rate 0.001 Attention
AGCLSTM 49.686 4.796 21.014 9.220 0.9878
Proposed 49.371 4.766 20.165 8.848 0.9878
benchmarks, with any unspecified hyperparameter settings consistent BSK CNN 70.706 7.174 38.840 24.600 0.958
GRU 70.219 7.124 33.783 21.396 0.959
with the proposed method.
CNN- 68.670 6.967 33.640 21.304 0.961
BiGRU
4. Results and discussion CNN-LSTM 67.576 6.856 31.571 19.995 0.962
CNN- 66.728 6.770 30.060 19.038 0.963
4.1. Comparison of the proposed method with the benchmark BiLSTM-
Attention
AGCLSTM 65.325 6.628 28.896 18.300 0.964
The baseline methods outlined in Section 3.3 were developed using Proposed 64.784 6.573 28.628 18.131 0.965
Tensorflow and Keras, with the Adam optimizer chosen for parameter DFA CNN 80.165 7.747 44.442 26.247 0.9519
optimization and MSE as the loss function. Each model was trained with GRU 82.375 7.961 41.328 24.409 0.9486
an epoch count of 80 and a batch size of 512. The initial learning rate CNN- 77.223 7.463 36.667 21.657 0.9556
BiGRU
was set to 0.001, the validation split to 0.1, and the time step to 24. All CNN-LSTM 76.910 7.433 36.361 21.475 0.9558
models were implemented on a personal computer running Windows 10, CNN- 75.906 7.336 35.065 20.710 0.9566
equipped with 12.00 GB of RAM and a 1.60 GHz CPU. BiLSTM-
Attention
AGCLSTM 71.663 6.926 33.344 19.693 0.9617
4.1.1. Comparison of overall performance under different climate types
Proposed 71.149 6.876 31.349 18.514 0.9620
To evaluate the average prediction performance across different sites
within the BSH, BSK, and DFA climate types, a comprehensive analysis
of the method’s prediction capability under these diverse conditions was models. The average MAE and R values for the three climate types were
conducted. Table 4 presents the corresponding results for all evaluation 20.165 W/m2, 28.628 W/m2, 31.349 W/m2, and 0.9878, 0.9650, and
metrics, with the optimal values highlighted in bold. All evaluation 0.9620, respectively, further demonstrating the superiority of the pro­
metrics are the average of the evaluation indicators for the eight sites posed method over alternative approaches. The results in Table 4 also
under the same climate type. For enhanced visualization, the results for reveal that the model focused solely on temporal feature extraction, such
RMSE, MAE, and R are depicted graphically in Fig. 6. As demonstrated in as CNN or spatial feature extraction, such as GRU performs the worst.
Table 4, the proposed hybrid method consistently achieves the best However, significant improvements are observed when spatial-temporal
prediction performance across all three climate types. The average features are integrated. Additionally, Table 4 underscores the strength of
RMSE values for BSH, BSK, and DFA are 49.371 W/m2, 64.784 W/m2, GCN in extracting spatial features. AGCLSTM, leveraging the spatial
and 71.149 W/m2, respectively, outperforming the other comparison modeling capabilities of GCN, exhibits average RMSE values of 49.686

7
H. Zang et al. Renewable Energy 246 (2025) 122945

Fig. 6. Average prediction errors of develop models for three climate type.

W/m2, 65.325 W/m2, and 71.663 W/m2 under the BSH, BSK, and DFA significant challenge, resulting in comparatively lower prediction ac­
climates, respectively, results that surpass CNN, GRU, and even CNN- curacy. Despite this, the proposed hybrid consistently demonstrates
BiLSTM-Attention. The dynamic adjacency matrix utilized in the pro­ notable improvements in the prediction metrics for all three climate
posed method further enhances its performance. types compared to other methods.
Fig. 6 illustrates that all methods perform best under the BSH
climate, likely due to factors such as altitude and specific climate 4.1.2. Comparison of overall performance for each site
characteristics. Meanwhile, the DFA climate type poses the most To provide a detailed comparison of the model’s performance, the

Fig. 7. Prediction errors of different models at twenty-four sites, measured in terms of RMSE and MAE.

8
H. Zang et al. Renewable Energy 246 (2025) 122945

prediction results for eight selected stations across three climate zones three climate types. Moreover, the proposed hybrid model stands out by
were analyzed collectively. Fig. 7 illustrates the performance of the 24 ensuring optimal prediction accuracy when generating GHI values for all
stations across three climate types, evaluated based on RMSE and MAE. stations simultaneously, which is more efficient compared to models
As depicted in Fig. 7, the proposed hybrid model demonstrates superior that predict one station at a time. This efficiency further highlights the
performance across all evaluation metrics. For the DFA climate type, the model’s strength in short-term solar irradiance prediction.
eight stations within this zone exhibited the best prediction results in Fig. 8 illustrates the prediction curve of the proposed model
terms of RMSE and MAE. The proposed model showed a clear advantage compared to the actual GHI curves for three distinct sites in Montana,
over other models, particularly with lower RMSE and MAE values. In the categorized under the BSK climate type, during the period from January
BSK zone, six of the eight stations (except for sites 1 and 6) demonstrated 1 to January 3, 2022. By analyzing the prediction results, it becomes
higher prediction accuracy than the other models. Although the pro­ evident that all comparison models tend to perform well when the GHI
posed model’s performance fluctuated more in BSH climate due to fac­ changes are minimal, with slow or no fluctuations. However, when
tors such as geographical location and climate variability, it still substantial GHI fluctuation occurs, most models struggle to adapt
achieved satisfactory results. quickly, resulting in slower predictions. In contrast, the proposed hybrid
Overall, the proposed model consistently excelled in MAE across the model demonstrates a superior ability to capture these rapid fluctua­
different climate types, with 23 out of 24 stations achieving the best tions, exhibiting a closer alignment with the actual GHI curve and
prediction results. The prediction outcomes varied across different smaller deviations, as highlighted in Fig. 8.
climate regions, but the hybrid model outperformed the comparison In conclusion, the advanced models incorporating adaptive feature
models at most sites. This demonstrates its superior flexibility across all selection, dynamic graph convolution, and attention mechanisms have

Fig. 8. Comparison of Actual and predicted GHI curves at three target sites from January 1, 2022 to January 3, 2022.

9
H. Zang et al. Renewable Energy 246 (2025) 122945

demonstrated exceptional capabilities in extracting spatial-temporal 4.2. Analysis of model interpretability

features. The proposed hybrid model consistently achieved the best
prediction outcomes across various evaluation metrics and climate 4.2.1. Analysis of the results of adaptive feature selection
types, highlighting its effectiveness and superiority in short-term solar To further elucidate the impact of adaptive feature selection in the
irradiance prediction. proposed hybrid model, we visualize and analyze the results of this
feature selection process. Given the variability of each meteorological
factor, the weights associated with the prediction results fluctuate at

Fig. 9. The results of adaptive feature selection of different climate types at 3:30 on January 2, 2022.

10
H. Zang et al. Renewable Energy 246 (2025) 122945

different time steps. Therefore, it is essential to accurately compute the 4.3. Discussion
weight of each feature at each time step, enhancing the significance of
important features while diminishing the influence of secondary ones, 4.3.1. Analysis of running time
thereby improving the prediction accuracy. Table 6 presents a comparison of the running time of all models for
Fig. 9 illustrates the results of adaptive feature selection at 3:30 on training and testing in BSK. The calculation time for CNN and GRU is
January 2, 2022, across three climate types: BSH, BSK, and DFA. In the relatively shorter, while CNN-LSTM and CNN-BiGRU exhibit longer
BSH climate type, CSI is identified as the most significant influencing running times due to the additional time feature extraction involved.
factor at the fifth time step, but it becomes the least important at the AGCLSTM has an extended training duration as it employs the LSTM
subsequent time step, indicating a reduced contribution to the predic­ network to extract time features for each site individually. Although the
tion outcome. Conversely, as the prediction time approaches, the weight training time of the proposed hybrid model is longer than that of the
assigned to GHI increases, consistently holding the highest weight across other comparison models due to its complex calculations, this increased
all features, signifying that the proximity of GHI to the prediction time computational cost is justified by the superior prediction accuracy
enhances its importance. These findings underscore that the variations achieved. Importantly, the prediction time remains efficient within a 1 h
in climate type, station, and meteorological factors significantly influ­ prediction period, indicating that the hybrid model not only excels in
ence the input features, with changes in input features being a crucial prediction performance but also holds significant application value.
reason for discrepancies in feature selection results. This reflects the
adaptability inherent in the feature selection process. Overall, the 4.3.2. Sensitivity analysis of key hyperparameters
adaptive feature selection network proposed in this study effectively The hyperparameters in the model play a crucial role in determining
identifies significant features corresponding to different inputs at each the final prediction outcomes; therefore, analyzing their impact on the
time step, thereby enhancing the model’s predictive capability. proposed model’s performance is essential. Using the DFA dataset,
Table 7 presents eight different hyperparameter combinations of the
4.2.2. Analysis of dynamic adjacency matrix proposed method. Table 8 showcases the prediction performance of the
Table 5 presents a comparison of the prediction results for the BSH, model with these varying combinations compared to the proposed
BSK, and DFA climate types across eight sites using two distinct adja­ configuration. The visualizations that the optimal hyperparameters
cency matrix calculation methods: AGCN and DGCN. Fig. 10 displays the combination employed in this study yields strong results in terms of
adjacency matrices obtained at 9:00, 12:00, and 18:00 on January 1, RMSE, MAE, and R. Model 1 exhibited the highest MAE and RMSE
2022. Additionally, the actual curves and the GHI curves from GCN- values, recorded at 33.24 W/m2 and 73.25 W/m2, respectively, high­
based methods are compared for January 1, 2022 in Fig. 10. lighting that variations in the number of GCN neurons significantly
The analysis reveals that when the GHI curve experiences fluctua­ affect the MAE.
tions due to cloud cover and weather variations, both AGCN and DGCN In conclusion, the hyperparameter combination selected in this
exhibit corresponding changes, aligning the predicted GHI curve more paper optimizes the prediction performance of the proposed model,
closely with the actual GHI curve. However, despite AGCN’s capability maintaining superior results compared to most other models even with
to adaptively compute the adjacency matrix at different time steps, its slight adjustments to the hyperparameters. This underscores the stabil­
prediction accuracy for multi-site GHI remains slightly inferior to that of ity and robustness of the proposed model in achieving accurate
the proposed dynamic graph convolution. predictions.
Notably, the DGCN method offers a less complex and computation­
ally intensive approach compared to AGCN, making it a more viable 5. Conclusion
option when considering the prediction accuracy achieved. This sug­
gests that dynamic graph convolution is more effective at capturing the This paper proposes a novel method for predicting GHI to enhance
variations in input data, thereby positively influencing prediction the accuracy of short-term solar irradiance prediction. The method be­
outcomes. gins with the SENet-based adaptive feature selection network to prior­
In contrast, traditional GCN calculations primarily focus on the itize the significance of each meteorological parameter and historical
characteristics of individual sites, often neglecting the relationships with solar irradiance data. Subsequently, dynamic graph convolution is
neighboring sites. Although AGCN adapts the adjacency matrix over employed to extract spatial features from meteorological parameters
time to incorporate spatial information across different sites, its across multiple stations. These spatial features are further optimized
computational complexity can be a drawback. Furthermore, the adja­ through an attention mechanism, while temporal features are extracted
cency matrix generated by the proposed DGCN not only reflects tem­ using BiLSTM. Finally, the solar irradiance is predicted as MLP. This
poral changes in input features but also provides a more comprehensive novel spatiotemporal solar irradiance prediction approach leverages
representation of the connections between various sites, enhancing dynamic graph convolution, which calculates adjacency matrices
overall prediction performance. dynamically at various time steps to flexibly capture spatial and tem­
In summary, the dynamic adjacency matrix, which varies across poral features, improving the overall prediction accuracy of solar
different time steps, effectively captures spatial information in a more irradiation.
flexible manner. This adaptability enhances the prediction stability of The model’s performance was evaluated using three datasets, con­
the proposed hybrid model across various prediction scenarios, resulting taining spatiotemporal data from 24 locations across Texas, Montana,
in improved accuracy and reliability in GHI prediction. and Ohio each representing a different climate zone. The results were

Table 5
Average prediction errors at 8 sites using different adjacency matrix calculation methods.
Climate type Model RMSE (W/m2) nRMSE (%) MAE (W/m2) nMAE (%) R

BSH AGCN-Attention-BiLSTM 49.8363 4.8109 21.1320 9.2718 0.9879

Proposed 49.3708 4.7659 20.1651 8.8476 0.9878
BSK AGCN-Attention-BiLSTM 64.9675 6.5913 28.6689 18.1568 0.9646
Proposed 64.7841 6.5727 28.6276 18.1307 0.9648
DFA AGCN-Attention-BiLSTM 71.1912 6.8800 31.5902 18.6570 0.9619
Proposed 71.1489 6.8759 31.3491 18.5143 0.9620

11
H. Zang et al. Renewable Energy 246 (2025) 122945

Fig. 10. Adjacency matrices obtained by two GCN-based methods at 9:00, 12:00, and 18:00 in one day on January 1, 2022.

Table 6 Table 7
The running time of all models for training and testing in BSK. Eight different hyperparameter combinations of the proposed method.
Climate type Model Training time(s) Testing time(s) Model Number of GCN Number of Attention Number of BiLSTM
units heads units
BSK CNN 368.007 0.976
GRU 449.734 2.651 Model 8 – –
CNN-BiGRU 679.426 3.988 1
CNN-LSTM 328.121 2.700 Model 16 – –
CNN-Attention-BiLSTM 1232.422 5.672 2
AGCLSTM 5424.621 24.448 Model 32 – –
Proposed method 2662.762 11.097 3
Model – 4 –
4
rigorously assessed across various seasons and sky conditions. The Model – 6 –
proposed hybrid model consistently outperformed other models across 5
Model – – 8
the 24 sites and 3 datasets. For the BSH dataset, the model achieved a 6
reduction in RMSE and MAE by 0.63–14.73 % and 4.04–37.26 %, Model – – 16
respectively. For the BSK dataset, the RMSE and MAE were reduced by 7
0.83–8.38 % and 0.93–26.29 %, respectively. In the DFA dataset, RMSE Model – – 32
8
and MAE were reduced by 0.72–11.25 % and 5.98–29.46 %, respec­
tively. The study also highlights the contribution of adaptive feature
selection and dynamic adjacency matrix generation, which enhance historical irradiance to improve spatiotemporal feature extraction,
both the model’s interpretability and prediction flexibility. In conclu­ optimize feature selection techniques, denoise the features, and develop
sion, this hybrid model demonstrates robust generalization and high new neural network architectures to further boost prediction
accuracy in short-term solar irradiance prediction. Future work will performance.
explore the weighted integration of meteorological features and

12
H. Zang et al. Renewable Energy 246 (2025) 122945

Table 8 gated LSTM recurrent network with differential covariance matrix adaptation
Prediction performance of the model with different combinations of evolution strategy, Energy 278 (2023) 127701.
[14] J. Liu, X. Huang, Q. Li, Z. Chen, G. Liu, Y. Tai, Hourly stepwise forecasting for solar
hyperparameters. irradiance using integrated hybrid models CNN-LSTM-MLP combined with error
Model RMSE (W/m2) MAE (W/m2) R correction and VMD, Energy Convers. Manag. 280 (2023) 116804.
[15] H. Zang, L. Liu, L. Sun, L. Cheng, Z. Wei, G. Sun, Short-term global horizontal
Proposed 71.14 31.34 0.9620 irradiance forecasting based on a hybrid CNN-LSTM model with spatiotemporal
Model 1 73.25 33.24 0.9596 correlations, Renew. Energy 160 (2020) 26–41.
Model 2 72.33 32.30 0.9607 [16] N.E. Michael, R.C. Bansal, A.A.A. Ismail, A. Elnady, S. Hasan, A cohesive structure
Model 3 72.01 32.91 0.9611 of Bi-directional long-short-term memory (BiLSTM) -GRU for predicting hourly
Model 4 72.05 32.77 0.9615 solar radiation, Renew. Energy 222 (2024) 119943.
Model 5 71.45 32.25 0.9618 [17] W. Liu, Z. Mao, Short-term photovoltaic power forecasting with feature extraction
Model 6 71.65 32.51 0.9615 and attention mechanisms, Renew. Energy 226 (2024) 120437.
Model 7 71.89 33.24 0.9615 [18] A.L. Jonathan, D. Cai, C.C. Ukwuoma, N.J.J. Nkou, Q. Huang, O. Bamisile,
A radiant shift: attention-embedded CNNs for accurate solar irradiance forecasting
Model 8 71.46 32.14 0.9616
and prediction from sky images, Renew. Energy (2024) 121133.
[19] J. Liang, W. Tang, Ultra-short-term spatiotemporal forecasting of renewable
resources: an attention temporal convolutional network-based approach, IEEE
CRediT authorship contribution statement Trans. Smart Grid 13 (5) (2022) 3798–3812.
[20] Y. Wen, S. Pan, X. Li, Z. Li, W. Wen, Improving multi-site photovoltaic forecasting
with relevance amplification: DeepFEDformer-based approach, Energy 299 (2024)
Haixiang Zang: Software, Methodology, Conceptualization. Wenan
131479.
Li: Writing – original draft, Software, Data curation. Lilin Cheng: [21] K.K.R. Samal, A.K. Panda, K.S. Babu, S.K. Das, Multi-output TCN autoencoder for
Writing – review & editing. Jingxuan Liu: Writing – review & editing, long-term pollution forecasting for multiple sites, Urban Clim. 39 (2021) 100943.
[22] M. Cubillos, Multi-site household waste generation forecasting using a deep
Software. Zhinong Wei: Writing – review & editing. Guoqiang Sun:
learning approach, Waste Manag. 115 (2020) 8–14.
Writing – review & editing. [23] H. Zhang, Z. Yi, L. Kang, Y. Zhang, K. Wang, A novel supercapacitor degradation
prediction using a 1D convolutional neural network and improved informer model,
Protection and Control of Modern Power Systems 9 (4) (2024) 51–68.
Declaration of competing interest [24] Y. Zhao, H. Liao, S. Pan, Y. Zhao, Interpretable multi-graph convolution network
integrating spatio-temporal attention and dynamic combination for wind power
The authors declare that they have no known competing financial forecasting, Expert Syst. Appl. 255 (2024) 124766.
[25] W. Lai, Z. Zhen, F. Wang, W. Fu, J. Wang, X. Zhang, H. Ren, Sub-region division
interests or personal relationships that could have appeared to influence based short-term regional distributed PV power forecasting method considering
the work reported in this paper. spatio-temporal correlations, Energy 288 (2024) 129716.
[26] Y. Ma, H. Lou, M. Yan, F. Sun, G. Li, Spatio-temporal fusion graph convolutional
network for traffic flow forecasting, Inf. Fusion 104 (2024) 102196.
Acknowledgments [27] H. Zang, Y. Zhang, L. Cheng, T. Ding, Z. Wei, G. Sun, Multi-site solar irradiance
forecasting based on adaptive spatiotemporal graph convolutional network, Expert
This work was supported in part by the National Natural Science Syst. Appl. 236 (2024) 121313.
[28] Z. Gao, Z. Li, H. Zhang, J. Yu, L. Xu, Dynamic spatiotemporal interactive graph
Foundation under Program No. 52477086 and Program No. U24B2088. neural network for multivariate time series forecasting, Knowl. Base Syst. 280
(2023) 110995.
References [29] X. Xu, S. Hu, H. Shao, P. Shi, R. Li, D. Li, A spatio-temporal forecasting model using
optimally weighted graph convolutional network and gated recurrent unit for wind
speed of different sites distributed in an offshore wind farm, Energy 284 (2023)
[1] Q.u. Zaman, Y. Zhao, S. Zaman, A.H. Shah, Examining the symmetrical effect of
128565.
traditional energy resources, industrial production, and poverty lessening on
[30] T. Cao, Y. Xu, G. Liu, S. Tao, W. Tang, H. Sun, Feature-enhanced deep learning
ecological sustainability: policy track in the milieu of five neighboring Asian
method for electric vehicle charging demand probabilistic forecasting of charging
economies, Resour. Policy 83 (2023) 103606.
station, Appl. Energy 371 (2024) 123751.
[2] J. Liu, H. Zang, T. Ding, L. Cheng, Z. Wei, G. Sun, A principle-constrained wind
[31] J. Wang, X. Tang, W. Jiang, A deterministic and probabilistic hybrid model for
field image generation framework for short-term wind power forecasting, IEEE
wind power forecasting based improved feature screening and optimal Gaussian
Trans. Power Syst. 40 (2) (2025) 1790–1801.
mixed kernel function, Expert Syst. Appl. 251 (2024) 123965.
[3] P. Singh, U. Kumar, N.K. Choudhary, N. Singh, Advancements in protection
[32] R. Zhang, X. Ma, C. Zhang, W. Ding, J. Zhan, GA-FCFNN: a new forecasting method
coordination of microgrids: a comprehensive review of protection challenges and
combining feature selection methods and feedforward neural networks using
mitigation schemes for grid stability, Protection and Control of Modern Power
genetic algorithms, Inf. Sci. 669 (2024) 120566.
Systems 9 (6) (2024) 156–183.
[33] Y. Gao, P. Li, H. Yang, J. Wang, A solar radiation intelligent forecasting framework
[4] L. Monjo, L. Sainz, J. Pedra, Model of quasi-Z-source inverter-based PV power
based on feature selection and multivariable fuzzy time series, Eng. Appl. Artif.
systems for stability studies of multi-terminal AC grid-connected PV power
Intell. 126 (2023) 106986.
systems, Int. J. Electr. Power Energy Syst. 155 (2024) 109639.
[34] M. Castangia, A. Aliberti, L. Bottaccioli, E. Macii, E. Patti, A compound of feature
[5] Y. Chu, Y. Wang, D. Yang, S. Chen, M. Li, A review of distributed solar forecasting
selection techniques to improve solar radiation forecasting, Expert Syst. Appl. 178
with remote sensing and deep learning, Renew. Sustain. Energy Rev. 198 (2024)
(2021) 114979.
114391.
[35] H. Acikgoz, A novel approach based on integration of convolutional neural
[6] Y. Nie, Q. Paletta, A. Scott, L.M. Pomares, G. Arbod, S. Sgouridis, J. Lasenby,
networks and deep feature selection for short-term solar radiation forecasting,
A. Brandt, Sky image-based solar forecasting using deep learning with
Appl. Energy 305 (2022) 117912.
heterogeneous multi-location data: dataset fusion versus transfer learning, Appl.
[36] G. Zhang, D. Choi, J. Jung, Development of continuous cuffless blood pressure
Energy 369 (2024) 123467.
prediction platform using enhanced 1-D SENet–LSTM, Expert Syst. Appl. 242
[7] T.C. McCandless, S.E. Haupt, G.S. Young, A regime-dependent artificial neural
(2024) 122812.
network technique for short-range solar irradiance forecasting, Renew. Energy 89
[37] P. Zhou, Y. Cui, X. Guo, J. Wei, H. Cao, Phase-wise attention GCN for
(2016) 351–359.
recommendation denoising, Appl. Soft Comput. 163 (2024) 111910.
[8] A. Zendehboudi, M.A. Baseer, R. Saidur, Application of support vector machine
[38] Y. Liu, S. Rasouli, M. Wong, T. Feng, T. Huang, RT-GCN: Gaussian-based
models for forecasting solar and wind energy resources: a review, J. Clean. Prod.
spatiotemporal graph convolutional network for robust traffic prediction, Inf.
199 (2018) 272–285.
Fusion 102 (2024) 102078.
[9] Y. Xu, S. Zheng, Q. Zhu, K.-c. Wong, X. Wang, Q. Lin, A complementary fused
[39] T.N. Kipf, M. Welling, Semi-Supervised Classification with Graph Convolutional
method using GRU and XGBoost models for long-term solar energy hourly
Networks (2016) 02907 arXiv:1609.
forecasting, Expert Syst. Appl. 254 (2024) 124286.
[40] C. Jiang, Y. Duan, Elasticity unleashed: fine-grained cloud scaling through
[10] Mustaqeem, M. Ishaq, S. Kwon, A CNN-Assisted deep echo state network using
distributed three-way decision fusion with multi-head attention, Inf. Sci. 660
multiple Time-Scale dynamic learning reservoirs for generating Short-Term solar
(2024) 120127.
energy forecasting, Sustain. Energy Technol. Assessments 52 (2022) 102275.
[41] Z. Li, L. Li, J. Chen, D. Wang, A multi-head attention mechanism aided hybrid
[11] O. El Alani, M. Abraim, H. Ghennioui, A. Ghennioui, I. Ikenbi, F.-E. Dahr, Short
network for identifying batteries’ state of charge, Energy 286 (2024) 129504.
term solar irradiance forecasting using sky images based on a hybrid CNN–MLP
[42] M. Sengupta, Y. Xie, A. Lopez, A. Habte, G. Maclaurin, J. Shelby, The national solar
model, Energy Rep. 7 (2021) 888–900.
radiation data base (NSRDB), Renew. Sustain. Energy Rev. 89 (2018) 51–60.
[12] Z. Wang, Y. Zhang, G. Li, J. Zhang, H. Zhou, J. Wu, A novel solar irradiance
forecasting method based on multi-physical process of atmosphere optics and
LSTM-BP model, Renew. Energy 226 (2024) 120367.
[13] M. Neshat, M.M. Nezhad, S. Mirjalili, D.A. Garcia, E. Dahlquist, A.H. Gandomi,
Short-term solar radiation forecasting using hybrid deep residual learning and

13
H. Zang et al. Renewable Energy 246 (2025) 122945

[43] J. Duan, M. Chang, X. Chen, W. Wang, H. Zuo, Y. Bai, B. Chen, A combined short- [45] S.Z. Xu, J. Liu, X.Q. Huang, C.L. Li, Z.Q. Chen, Y.H. Tai, Minutely multi-step
term wind speed forecasting model based on CNN–RNN and linear regression irradiance forecasting based on all-sky images using LSTM-InformerStack hybrid
optimization considering error, Renew. Energy 200 (2022) 788–808. model with dual feature enhancement, Renew. Energy 224 (2024) 120135.
[44] Y. Xiao, C. Zou, H. Chi, R. Fang, Boosted GRU model for short-term forecasting of
wind power with feature-weighted principal component analysis, Energy 267
(2023) 126503.

14