Multi-site charging load prediction method and system considering multi-variable heterogeneous characteristics

By designing a dedicated encoder to process time, space, and price features, and utilizing cross-attention mechanisms and price elasticity regularization constraints, the problems of loss of heterogeneous feature information and insufficient price response mechanisms in existing charging load forecasting are solved, achieving higher accuracy and economic rationality in charging load forecasting.

CN123121489APending Publication Date: 2026-11-17TIANJIN UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611240226.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-17
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing electric vehicle charging load forecasting methods suffer from information loss and insufficient price response mechanisms when dealing with heterogeneous features such as time, space, and price, resulting in poor prediction accuracy and economic interpretability.

Method used

A multi-site charging load prediction method considering the three heterogeneous features of time, space and price is adopted. Dedicated encoders are designed to process each type of feature, and a cross-attention mechanism is used to achieve the collaborative expression of heterogeneous features. At the same time, a regularization constraint based on price elasticity is constructed to enhance the prediction accuracy and economic interpretability of the model.

Benefits of technology

It improves the accuracy of multi-site charging load forecasting, enhances the economic interpretability of the model and the stability of the training process, and can more accurately predict charging demand in the next 24 hours.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123121489A_ABST
    Figure CN123121489A_ABST
Patent Text Reader

Abstract

This invention discloses a multi-site charging load prediction method and system considering diverse heterogeneous features, including: acquiring historical charging demand data, geographical data, and price data of multiple sites; constructing a multi-site charging load prediction network; the prediction network includes a cascaded heterogeneous feature three-branch encoder, a cross-attention collaborative expression module, a multi-scale periodic modeling module, a Transformer encoding module, and an output projection layer; training the multi-site charging load prediction network; and using the trained prediction network to predict the charging demand of multiple charging stations for the next 24 hours. This invention designs dedicated encoders for the different representation mechanisms of three types of heterogeneous features: historical load time features, site map structure features, and price features, and achieves collaborative expression between heterogeneous features and aggregation of multi-site information through cross-attention, overcoming the loss of feature information caused by traditional unified encoding that maps heterogeneous features to the same representation space.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of charging pile load prediction technology, and more specifically, to a method and system for predicting charging load at multiple sites that considers diverse heterogeneous characteristics. Background Technology

[0002] With the deepening of the global energy transition, the number of electric vehicles (EVs) is experiencing explosive growth. EV charging load exhibits significant randomness, volatility, and inter-station coupling: in the time dimension, influenced by user travel patterns, charging demand shows a clear peak-valley alternation characteristic; in terms of inter-station relationships, charging stations located in different functional areas have complex load correlations and mutual influences. Without accurate forecasting support, the large-scale, disorderly integration of EVs into the power grid may lead to local distribution network overload, increased system peak-shaving pressure, and even threaten the safe and stable operation of the power grid. Therefore, achieving accurate forecasting of EV charging load is of great value for optimizing the layout of charging infrastructure, rationally allocating grid reserve capacity, and supporting dynamic pricing and peak-shaving guidance.

[0003] To achieve accurate charging load prediction, research has explored various approaches, from machine learning to deep learning. Methods based on random forests, extreme gradient boosting, and multi-learner ensembles offer advantages in speed and robustness. The introduction of deep learning methods, such as Transformer, has significantly improved prediction accuracy and can support online real-time load prediction. Building on this foundation, related research has systematically reviewed charging load prediction and collaborative optimization strategies, highlighting that multi-source information fusion and deep learning methods have become important pathways to improve prediction performance.

[0004] With the increasing number of charging stations, it is necessary to predict the load distribution of multiple charging stations simultaneously. Significant inter-station load correlations exist, and incorporating these inter-station correlations as graph structure features into the model can further improve the accuracy of multi-site collaborative prediction. Some studies have introduced spatial features to model charging station networks, characterizing the spatiotemporal distribution of charging loads in urban areas. However, various heterogeneous features, such as temporal and spatial features, have fundamentally different representation mechanisms. Uniform encoding makes it difficult to fully express the unique feature information of each type of feature, limiting the model's ability to capture spatiotemporal coupling patterns.

[0005] To enrich predictive input information, some studies have begun to incorporate external data such as electricity prices and traffic flow, proposing physical information graph learning methods that simultaneously consider charging demand and electricity prices, thus improving the model's adaptability to external data. Analysis based on measured data shows that the demand for public charging of electric vehicles exhibits a quantifiable elastic response to prices, with significant differences in price elasticity across different regions and charging methods. Furthermore, existing research has provided reference ranges for the price elasticity of electricity demand through meta-analysis, offering a basis for quantitative modeling of price response mechanisms. However, existing research still has room for improvement in the handling of price features: most models treat price as a common input feature, lacking explicit constraints on the price response mechanism, resulting in a lack of economic interpretability in the price dimension and insufficient stability during training when price-related samples are sparse.

[0006] In summary, existing electric vehicle charging load forecasting methods suffer from two main shortcomings: first, the unified encoding of heterogeneous features such as time, space, and price leads to the loss of unique information for each feature, making it difficult to fully capture the spatiotemporal coupling patterns; second, the lack of explicit constraints on price response mechanisms results in insufficient interpretability of the model's economics and poor training stability in the price dimension. To address these shortcomings, this invention proposes a multi-site charging load forecasting method that considers the ternary heterogeneous features of time, space, and price. Summary of the Invention

[0007] The purpose of this invention is to address the shortcomings of existing electric vehicle charging load prediction methods, such as information loss due to unified encoding of heterogeneous features, insufficient characterization of price response mechanisms, and poor training stability of the price dimension. This invention proposes a multi-site charging load prediction method that considers the three heterogeneous features of time, space, and price. This method designs dedicated encoders for each of the three types of heterogeneous features—historical load time features, site map structure features, and price features—based on their different representation mechanisms. Within the Transformer encoding framework, a cross-attention mechanism is used to achieve collaborative expression among heterogeneous features and aggregation of multi-site information. Simultaneously, a regularization constraint based on price elasticity is constructed, setting an independent demand price elasticity coefficient for each charging site. The price elasticity mechanism is embedded as a physical constraint in the loss function, thereby improving the accuracy of multi-site charging load prediction while enhancing the economic interpretability and training stability of the model.

[0008] Specifically, the present invention provides a multi-site charging load prediction method considering diverse heterogeneous characteristics, comprising the following steps:

[0009] Acquire historical charging demand data, geographical data, and price data from multiple sites, process the data, and construct a historical charging power input sequence;

[0010] A multi-site charging load prediction network is constructed; the prediction network includes a cascaded heterogeneous feature three-branch encoder, a cross-attention collaborative expression module, a multi-scale periodic modeling module, a Transformer encoding module, and an output projection layer;

[0011] The multi-site charging load prediction network is trained using the obtained historical charging power input sequence to obtain the trained prediction network.

[0012] The trained prediction network is used to predict the charging demand of multiple charging stations for the next 24 hours.

[0013] Furthermore, the heterogeneous feature three-branch encoder includes a temporal branch encoder, a spatial branch encoder, and a price branch encoder; wherein, the temporal branch encoder includes a one-dimensional convolution with four sets of convolutional kernels, a convolutional attention gating unit, and a temporal self-attention layer; the spatial branch encoder includes a hybrid adjacency matrix construction module, a two-layer graph convolutional network, and a multi-head spatial attention layer; the price branch encoder includes a two-layer MLP price embedding network; the temporal branch encoder, spatial branch encoder, and price branch encoder respectively output three-branch features. , , .

[0014] Furthermore, the time-branch encoder, spatial-branch encoder, and price-branch encoder each output three-branch features. , , The method is as follows:

[0015] One-dimensional convolutions with four sets of kernels (1, 3, 5, and 7) are used to capture the features of charging demand at different time scales, including immediate fluctuations, short-term cycles, and long-term trends. A convolutional attention enhancement mechanism is introduced to generate feature-level attention weights through one-dimensional convolutions and adaptively gate the fused features.

[0016] ;

[0017] In the formula, Let B be the feature at the k-th scale, and B be the batch size. The total dimension of time features This is a one-dimensional convolution with a kernel size of k, and ReLU is the activation function. For bias vectors, To fuse the weight matrix, For bias vectors, The merged time features are represented by the concatenation operation, where Concat represents the concatenation operation. The enhanced temporal features; H1, H3, H5, and H7 represent the values ​​of k when k is 1, 3, 5, and 7, respectively. ;

[0018] By modeling long-range dependencies within a sequence using a temporal self-attention mechanism, a temporal representation with both multi-scale local perception and global temporal modeling capabilities is generated.

[0019] ;

[0020] In the formula, , , These are query, key, and value matrices, respectively, all derived from... Linear projection yields, The dimension of the key vector. Here is the attention weight matrix. For self-attention output, Softmax is a normalized exponential function;

[0021] Establish geographical adjacency relationships based on the geographical distance between sites:

[0022] ;

[0023] In the formula, The geographical distance between site i and site j For Gaussian kernel bandwidth, For connection threshold, This is the element in the i-th row and j-th column of the geographic adjacency matrix;

[0024] The cosine similarity of power time series between sites is calculated, and a hybrid adjacency matrix is ​​obtained by fusing geographical and functional adjacency relationships. The adjacency matrix is ​​then normalized.

[0025] ;

[0026] In the formula: , Let i be the historical power time-series vector of stations i and j. Similarity scores are calculated only in the training set interval, and the functional adjacency matrix constructed accordingly remains fixed in the validation and test sets. For similarity threshold, The element in the i-th row and j-th column of the functional adjacency matrix. For geographical weight, For functional weights, It is a hybrid adjacency matrix. It is a diagonal matrix. It is the identity matrix. This is the normalized adjacency matrix; It is a geographic adjacency matrix; It is a functional adjacency matrix;

[0027] Charging station topology encoding is achieved using a two-layer graph convolutional network.

[0028] ;

[0029] In the formula: The site node feature matrix is ​​obtained by linearly mapping the historical power sequences of each site along the time dimension. The convolution weights for the first layer graph are... For the convolution weights of the second layer graph, and For bias, , These are the spatial features output by the first and second layer graph convolutions, respectively. For site node feature dimensions, , Output dimensions for two-layer graph convolution;

[0030] A multi-head spatial attention mechanism is deployed on top of graph convolutional features to calculate dynamic interaction weights between sites, enabling cross-site information aggregation. Through neighborhood aggregation and feature transformation, a deep nonlinear transformation is performed on the node features to obtain the final spatial feature output.

[0031] ;

[0032] In the formula: For the number of attention heads, , , Let m be the query, key, and value projection matrix of the attention head. To output the projection matrix, For the output of the m-th attention head, This is the structure-aware projection matrix. For the final spatial feature output, For a single attention head dimension, For the aggregated output of multi-head spatial attention, For structure-aware transformation features;

[0033] A price feature tensor is constructed using three characteristics: electricity price, service price, and price changes. :

[0034] ;

[0035] In the formula, Let be the standardized price feature vector of station i at time t. The standardized electricity price at station i at time t. The standardized service price for station i at time t. This represents the total price change;

[0036] Price representation obtained through two-layer MLP embedding Output as the price branch:

[0037] ;

[0038] In the formula, For the first layer of embedded weights, For the second layer of embedded weights, , It is a bias vector. For price embedding features, For the price MLP hidden layer dimension, Price is embedded as a dimension.

[0039] Furthermore, the cross-attention collaborative representation module includes three linear projection layers and three cross-attention layers, projecting the three-branch features onto a unified dimension. Then, the cross-attention mechanism is used for interaction, and the final fused feature is obtained by fusing the adaptive association weight network and the gating network.

[0040] Furthermore, the method for obtaining the final fusion feature is as follows:

[0041] Projecting time, space, and price characteristics onto a unified dimension :

[0042] ;

[0043] in, , , These are projection matrices representing temporal, spatial, and price characteristics, respectively. , , Features after projection;

[0044] The cross-attention mechanism is used to achieve collaborative representation among heterogeneous features. Spatiotemporal features, as the predictive backbone, are actively queried by price features to inject price response information, while space, as a structural representation, is queried by time to supplement structural information.

[0045] ;

[0046] in, For temporal-spatial cross-attention features, For price-time cross-attention features, Price-space cross-attention features;

[0047] The adaptive association weights of the three-way cross features are learned and then gated to obtain the final fused features:

[0048] ;

[0049] In the formula, For adaptive association weight vector, Control vector, For element-wise product, Features after splicing This is the final fusion feature.

[0050] Furthermore, the multi-scale periodic modeling module includes a fast Fourier transform, an inverse Fourier transform, and a multi-head attention and gating fusion layer at three scales.

[0051] The multi-scale periodic modeling module performs frequency domain analysis on the fused features and extracts the Top-K main periodic frequency components, retaining the first four main frequency components. The main periodic components are then weighted with the original fused features using inverse Fourier transform to obtain periodic enhancement features, and multi-head attention is applied at the short-term, medium-term, and long-term scales. All three scales of multi-head attention use the complete sequence as the query and the corresponding time slice as the key and value, with a uniform output length. By learning the fusion weights of three time scales through a gating mechanism, hierarchical attention fusion features are obtained.

[0052] Furthermore, when training the multi-site charging load prediction network, the mean squared error loss and the price elasticity regularization term are weighted to form the total error loss, and the network is trained end-to-end:

[0053] ;

[0054] In the formula, For mean square error loss, For the total loss, For price elasticity regularization, As the weight of the regularization term, For the actual needs of station i at time t in batch n, Let B represent the predicted demand for station i at time t in batch n, where B is the batch size and N is the number of charging stations. To predict the sequence length.

[0055] Furthermore, the method for obtaining the price elasticity regularization term is as follows:

[0056] By using automatic differential calculations to predict the partial derivative of demand with respect to electricity prices, a local sensitivity index of the network to price input is constructed:

[0057] ;

[0058] In the formula, For site i, the local diagnostic price elasticity at a single moment; The predicted demand is obtained by backpropagation of the automatic differential pair with respect to the input electricity price. and All values ​​are the original values ​​after standardization. Take 0.01;

[0059] The reference target value is constructed by superimposing the comprehensive literature value on a bounded range given by the site's functional attributes:

[0060] ;

[0061] In the formula, The reference target elasticity for site i; Take 0.15; Based on the prior knowledge of the site's functional areas:

[0062] Using the deviation between the diagnostic price sensitivity index and the reference target value as the regularization loss term, directional regularization is applied to the network's local response in the price dimension under the batch average meaning:

[0063] ;

[0064] In the formula, Obtained through automatic differentiation techniques: For the predicted demand of station i at time t in batch n, For the corresponding electricity price, take the electricity price at the end of the input window. As a price benchmark.

[0065] Furthermore, the weights of the regularization terms are dynamically adjusted during the training process based on the training phase. It provides strong directional guidance in the early stages and maintains a weak regularization effect in the later stages.

[0066] ;

[0067] In the formula, The training rounds are designed to provide directional guidance for parameter updates in the price dimension during the phase with a weight of 1.0; data-driven improvement dominates when the weight is reduced to 0.1; and weak regularization is retained during the phase with a weight of 0.01, ultimately achieving a balance between prediction accuracy and training process stability.

[0068] A multi-site charging load prediction system considering diverse and heterogeneous characteristics includes:

[0069] The historical charging power input sequence acquisition module is used to acquire historical charging demand data, geographical data, and price data from multiple sites, process the data, and construct a historical charging power input sequence.

[0070] A prediction network construction module is used to construct a multi-site charging load prediction network; the prediction network includes a cascaded heterogeneous feature three-branch encoder, a cross-attention collaborative expression module, a multi-scale periodic modeling module, a Transformer encoding module, and an output projection layer;

[0071] The prediction network training module is used to train the multi-site charging load prediction network using the obtained historical charging power input sequence to obtain the trained prediction network.

[0072] The charging demand prediction module is used to predict the charging demand of multiple charging stations for the next 24 hours using a trained prediction network.

[0073] The beneficial technical effects of this invention are as follows:

[0074] This invention constructs a multi-site charging load prediction framework that considers the ternary heterogeneous features of time, space, and price. It designs dedicated encoders for the different representation mechanisms of the three types of heterogeneous features: historical load time features, site map structure features, and price features. Cross-attention is used to achieve collaborative expression among heterogeneous features and aggregation of multi-site information, overcoming the loss of feature information caused by traditional unified encoding that maps heterogeneous features to the same representation space. Simultaneously, it introduces a regularization constraint based on price elasticity, automatically differentiating the network to calculate the local sensitivity of price and aligning it with the site-level reference target value. The price elasticity mechanism is embedded as a physical constraint in the loss function, compensating for the shortcomings of traditional black-box models in characterizing price mechanisms and improving the economic interpretability of the model. Attached Figure Description

[0075] Figure 1 This is a flowchart illustrating the multi-site charging load prediction method considering diverse heterogeneous characteristics provided in an embodiment of the present invention.

[0076] Figure 2 This is a charging demand forecast map of the top six charging stations in Shenzhen provided in Embodiment 2 of the present invention;

[0077] Figure 3 This is a price elasticity coefficient distribution diagram for three types of sites provided in Embodiment 2 of the present invention;

[0078] Figure 4(a) shows the evolution of the elasticity coefficients of the three types of sites as a function of training, as provided in Embodiment 2 of the present invention.

[0079] Figure 4(b) is a representative site resilience coefficient tracking diagram provided in Embodiment 2 of the present invention. Detailed Implementation

[0080] Example 1

[0081] This embodiment discloses a multi-site charging load prediction method that considers diverse heterogeneous characteristics, integrating data-driven heterogeneous characteristic encoding with price elasticity mechanisms, including the following steps:

[0082] Historical charging demand data, geographical data, and price data are acquired. The data is then processed, including missing value imputation, outlier site removal, and normalization preprocessing, to construct a historical charging power input sequence. ;

[0083] Among them, the historical charging power input sequence Represented as:

[0084] ;

[0085] In the formula, The input sequence length is given, the input window is fixed at 7 days, and N is the number of charging stations. , It is a column vector that includes the power of each charging station at time t.

[0086] A multi-site charging load prediction network is constructed. The prediction network is an end-to-end trainable whole, consisting of five cascaded parts:

[0087] 1. A heterogeneous feature three-branch encoder, comprising a temporal branch encoder, a spatial branch encoder, and a price branch encoder; wherein the temporal branch encoder comprises a one-dimensional convolution with four sets of convolutional kernels, a convolutional attention gating unit, and a temporal self-attention layer; the spatial branch encoder comprises a hybrid adjacency matrix construction module, a two-layer graph convolutional network, and a multi-head spatial attention layer; the price branch encoder comprises a two-layer MLP price embedding network, and a price elasticity regularization term is derived by an automatic differentiation module. The branch encoder, spatial branch encoder, and price branch encoder each output three-branch features. , , Specifically:

[0088] One-dimensional convolutions with four sets of kernels (1, 3, 5, and 7) are used to capture features of charging demand at different time scales, such as immediate fluctuations, short-term cycles, and long-term trends. A convolutional attention enhancement mechanism is introduced to generate feature-level attention weights through one-dimensional convolutions and adaptively gate the fused features.

[0089] ;

[0090] In the formula, Let B be the feature at the k-th scale, and B be the batch size. That is, both training and inference are performed in mini-batch mode, and B samples are input at the same time in a single forward propagation. The total dimension of time features This is a one-dimensional convolution with a kernel size of k, and ReLU is the activation function. For bias vectors, To fuse the weight matrix, For bias vectors, The merged time features are represented by the concatenation operation, where Concat represents the concatenation operation. The enhanced temporal features; H1, H3, H5, and H7 represent the values ​​of k when k is 1, 3, 5, and 7, respectively. , respectively, represent the instantaneous fluctuations of charging demand (k=1), local changes within a short period (k=3), short-term cycles at the hour level (k=5), and long-term trends across time periods (k=7); where the multiplication of the three-dimensional feature tensor with the two-dimensional weight matrix represents a linear transformation along the last dimension of the tensor, that is, multiplying each two-dimensional slice on the batch dimension and the sequence dimension by the weight matrix on the right; the addition of the feature tensor with the bias vector represents the element-wise addition of the bias vector to the tensor after it is broadcast along the last dimension, that is, the bias is shared on the batch dimension and the sequence dimension.

[0091] Adaptive weighting of multi-scale local features in the channel dimension was achieved, but its receptive field is limited by the convolution kernel size, making it unable to characterize the dependence of charging demand on the trans-day scale. Therefore, [further research is needed]. Using a common input, query, key, and value matrices are generated through three independent linear projections: , , , These are three independent linear projection weight matrices. , , For the corresponding bias vector The three elements share the same input but not the same parameters, which makes the gating features at the same time step have different representations when they are the query object and when they are the query initiator. Thus, the local multi-scale features extracted in formula (2) are converted into a representation space that can be used for global similarity measurement. Based on this, the attention weight between any two time steps is calculated according to formula (3), and long-range dependencies can be established within the entire input window.

[0092] By modeling long-range dependencies within a sequence using a temporal self-attention mechanism, a temporal representation with both multi-scale local perception and global temporal modeling capabilities is generated.

[0093] ;

[0094] In the formula, , , These are query, key, and value matrices, respectively, all derived from... Linear projection yields, Here is the attention weight matrix. For self-attention output, Softmax is a normalized exponential function.

[0095] Establish geographical adjacency relationships based on the geographical distance between sites:

[0096] ;

[0097] In the formula, The geographical distance between site i and site j For Gaussian kernel bandwidth, For connection threshold, It is the element in the i-th row and j-th column of the geographic adjacency matrix.

[0098] The cosine similarity of power time series between sites is calculated, and a hybrid adjacency matrix is ​​obtained by fusing geographical and functional adjacency relationships. The adjacency matrix is ​​then normalized.

[0099] ;

[0100] In the formula: , Let i be the historical power time-series vector of stations i and j. Similarity scores are calculated only in the training set interval, and the functional adjacency matrix constructed accordingly remains fixed in the validation and test sets. For similarity threshold, The element in the i-th row and j-th column of the functional adjacency matrix. For geographical weight, For functional weights, It is a hybrid adjacency matrix. It is a diagonal matrix. It is the identity matrix. This is the normalized adjacency matrix; It is a geographic adjacency matrix; This is a functional adjacency matrix.

[0101] Charging station topology encoding is achieved using a two-layer graph convolutional network.

[0102] ;

[0103] In the formula: The site node feature matrix is ​​obtained by linearly mapping the historical power sequences of each site along the time dimension. The convolution weights for the first layer graph are... For the convolution weights of the second layer graph, and For bias, , These are the spatial features output by the first and second layer graph convolutions, respectively. For site node feature dimensions, , The output dimension of the two-layer graph convolution is represented by the adjacency matrix multiplied by the three-dimensional node feature tensor, which means that the adjacency matrix is ​​broadcast along the batch dimension and then left-multiplied by the node feature matrix of each sample.

[0104] A multi-head spatial attention mechanism is deployed on top of graph convolutional features to calculate dynamic interaction weights between sites, enabling cross-site information aggregation. Through neighborhood aggregation and feature transformation, a deep nonlinear transformation is performed on the node features to obtain the final spatial feature output.

[0105] ;

[0106] In the formula: For the number of attention heads, , , Let m be the query, key, and value projection matrix of the attention head. To output the projection matrix, For the output of the m-th attention head, This is the structure-aware projection matrix. For the final spatial feature output, For a single attention head dimension, The aggregated output of multi-head spatial attention is provided by The dynamic interaction weights between stations are calculated in parallel by multiple attention heads, weighted and aggregated, and then mapped through the output projection matrix. This is used to characterize the inter-station load correlations adaptively learned from the data outside of the fixed graph structure. The structure-aware transformation features are output by the second layer of graph convolution. It is obtained by linear transformation of the structure-aware projection matrix and activation by ReLU, and is used to preserve the unique topological properties of each charging station obtained by graph convolution encoding.

[0107] A price feature tensor is constructed using three characteristics: electricity price, service price, and price changes. :

[0108] ;

[0109] In the formula, Let be the standardized price feature vector of station i at time t. The standardized electricity price at station i at time t. The standardized service price for station i at time t. This represents the total price change.

[0110] Price representation obtained through two-layer MLP embedding Output as the price branch:

[0111] ;

[0112] In the formula, For the first layer of embedded weights, For the second layer of embedded weights, , It is a bias vector. For price embedding features, For the price MLP hidden layer dimension, Price is embedded as a dimension.

[0113] 2. Cross-attention collaborative representation module: This module includes three linear projection layers and three cross-attention layers, which project the three-branch features to a unified dimension. Then, a cross-attention mechanism is used for interaction, followed by fusion between an adaptive association weight network and a gating network; specifically,

[0114] Projecting time, space, and price characteristics onto a unified dimension :

[0115] ;

[0116] in, , , These are projection matrices representing temporal, spatial, and price characteristics, respectively. , , These are the projected features.

[0117] A cross-attention mechanism is used to achieve collaborative representation among heterogeneous features. Spatiotemporal features, as the predictive backbone, are actively queried by price features to inject price response information, while space, as a structural representation, is queried by time to supplement structural information. Spatial features have already been aggregated and are no longer used as query sources to reduce redundant interactions.

[0118] ;

[0119] in, For temporal-spatial cross-attention features, For price-time cross-attention features, This represents the price-space cross-attention feature.

[0120] The adaptive association weights of the three-way cross features are learned and then gated to obtain the final fused features:

[0121] ;

[0122] In the formula, For adaptive association weight vector, Control vector, For element-wise product, Features after splicing This is the final fusion feature.

[0123] 3. Multi-scale periodic modeling module: It consists of fast Fourier transform, inverse Fourier transform, and a multi-head attention and gating fusion layer at three scales;

[0124] Frequency domain analysis was performed on the fusion features, and the Top-K main periodic frequency components were extracted, retaining the first four main frequency components:

[0125] ;

[0126] In the formula, Here, f is the frequency index and j is the imaginary unit, representing the Fast Fourier Transform operator. The first four main frequency components are retained.

[0127] The main periodic components are weighted with the original fused features after inverse Fourier transform to obtain periodic enhancement features, and multi-head attention is applied at the short-term, medium-term, and long-term scales respectively:

[0128] ;

[0129] In the formula, For inverse Fast Fourier Transform, Weights are fused to learnable periodic features. Output features for TimesBlock, This indicates taking the features from the most recent n time steps. , , These represent the number of attention heads in the short, medium, and long term, respectively. , , Attention output at three scales.

[0130] Multi-head attention at all three scales uses the complete sequence as the query and the corresponding time slice as the key and value, with a uniform output length. By learning the fusion weights at three time scales through a gating mechanism, hierarchical attention fusion features are obtained:

[0131] ;

[0132] In the formula, For dynamic weight vectors, For sequence averaging, This is a feature for hierarchical attention fusion.

[0133] 4. Transformer Encoding Module: This module consists of a positional encoding layer and 3-5 stacked encoding layers. Each encoding layer contains a multi-head self-attention sub-layer and a feedforward network sub-layer. Each sub-layer employs residual connections and layer normalization. Specifically, after the hierarchical fused features are stacked with positional encoding, sequence modeling is performed through the Transformer encoding layer composed of multi-head self-attention and a feedforward network. The encoded features are then output through residual connections and layer normalization.

[0134] ;

[0135] In the formula, For location index, For dimensional indexing, For the position encoding matrix, , This is the weight matrix. , For bias vectors, Let l be the intermediate dimension of the feedforward network, l be the index of the coding layer, and L be the total number of coding layers. This is the output of the l-th layer, and the input of the first layer is... The output of each of the remaining layers is from the previous layer. As input.

[0136] 5. Output Projection Layer: Consists of a linear layer that maps the encoded features of the final layer to multi-site load prediction results; specifically,

[0137] The final layer encoded features are mapped to multi-site future charging load prediction results through the output projection layer:

[0138] ;

[0139] In the formula, Transpose represents the transpose of the tensor's time dimension and feature dimension. The time-dimensional projection matrix, The output projection matrix is ​​responsible for mapping the last layer's encoded features to the predicted output in terms of sequence length and feature dimension. This is the bias vector.

[0140] During the training of the multi-site charging load prediction network, the obtained historical charging power input sequence will be used to construct the network. The network is divided into training, testing, and validation sets. The training set is used for network training. It should be noted that the historical charging power input sequence is used as the training set. The historical charging power is collected at a time interval of 15 minutes. That is, the original charging transaction records are aggregated into a demand time series matrix at a granularity of 15 minutes. This is because when the sampling interval is less than 5 minutes, the power curve of a single charging station is mainly driven by the access and disconnection events of individual vehicles, showing strong random step jumps. The proportion of noise components in the total variance increases significantly, and the load correlation between stations is submerged by random fluctuations. When the sampling interval is greater than 1 hour, the ramping process of the morning and evening peaks and short-term power spikes are averaged and smoothed during the aggregation process, and the peak-valley transition characteristics of charging demand are weakened. The granularity of 15 minutes can not only retain the ramping rate and short-term spike shape of the morning and evening peaks, but also filter out the random jumps caused by the access and disconnection of individual vehicles. At the same time, it is consistent with the minimum billing period of time-of-use electricity price. Moreover, the electricity price shows a clear step change at this granularity. Therefore, this invention adopts a time interval of 15 minutes.

[0141] Drawing on the dimensionless expression of price elasticity, this paper constructs a local sensitivity index of the network to price input by automatically differentiating the network to predict the partial derivative of demand with respect to electricity price, thereby characterizing the internal response of the data-driven model.

[0142] ;

[0143] In the formula, For site i, the local diagnostic price elasticity at a single moment; The predicted demand is obtained by backpropagation of the automatic differential pair with respect to the input electricity price. and All values ​​are the original values ​​after standardization. A value of 0.01 is used to avoid abnormal elasticity dominating losses during a few near-zero demand moments. The site-level elasticity for each site is obtained by averaging the local diagnostic elasticity of that site across all moments, and is used for subsequent analysis and statistics. The reference target value is constructed based on the comprehensive literature value, superimposed with a bounded range given by the site's functional attributes:

[0144] ;

[0145] In the formula, the combined value of global household electricity demand price elasticity and vehicle fuel price elasticity, -0.45, is used as the reference benchmark; Let i be the reference target elasticity; the pruning interval is tightened to [−0.8,−0.1] to maintain the economic constraint of negative elasticity; Take 0.15; Based on the prior knowledge of site functional areas: commercial areas and public fast charging stations have higher price sensitivity, so a value of -0.15 is assigned; residential and office areas have lower slow charging stations, so a value of +0.15 is assigned; and other sites have a value of 0. Site heterogeneity is preserved based on the global prior. The deviation between the diagnostic price sensitivity index and the reference target value is used as the regularization loss term. Directional regularization is applied to the network's local response in the price dimension under the batch average meaning.

[0146] ;

[0147] In the formula, Obtained through automatic differentiation: Gradient tracking is enabled on the price input during forward propagation, and automatic differentiation is performed using the formula in reverse mode. , For the predicted demand of station i at time t in batch n, For the corresponding electricity price, take the electricity price at the end of the input window. As a price benchmark, the loss term applies directional regularization to the network's local response in the price dimension in the sense of batch averaging, while temporal fluctuations at specific moments are preserved by data-driven learning.

[0148] The weights of the regularization term are dynamically adjusted according to the training phase. It provides strong directional guidance in the early stages and maintains a weak regularization effect in the later stages.

[0149] ;

[0150] In the formula, The training rounds are designed to provide directional guidance for parameter updates in the price dimension during the phase with a weight of 1.0; data-driven improvement dominates when the weight is reduced to 0.1; and weak regularization is retained during the phase with a weight of 0.01, ultimately achieving a balance between prediction accuracy and training process stability.

[0151] The total error loss is calculated by weighting the mean squared error loss with the price elasticity regularization term, and the network is then trained end-to-end.

[0152] ;

[0153] In the formula, For mean square error loss, For the total loss, This represents the actual needs of station i at time t in batch n.

[0154] Training with total loss As an optimization objective, the maximum number of training epochs is set to 50. After each training epoch, the mean squared error is calculated on the validation set. When the mean squared error on the validation set no longer decreases for 8 consecutive epochs, the training is terminated early, and the network parameters with the lowest mean squared error on the validation set are rolled back and saved as the final model.

[0155] When using a trained multi-site charging load prediction network to predict charging demand for the next 24 hours, historical charging power data for the seven days preceding the prediction period is obtained. This data is processed, including missing value imputation, outlier site removal, and normalization preprocessing, to construct a historical charging power input sequence. This historical charging power input sequence is then input into the trained multi-site charging load prediction network to obtain the charging demand for the next 24 hours. , To predict the sequence length.

[0156] Example 2

[0157] In this embodiment, based on the method provided in Embodiment 1, as an example, data on the electric vehicle charging network in Shenzhen, Guangdong Province, China, is collected, including 247 charging stations and approximately 611,600 original charging transaction records. These records are aggregated into a demand time series matrix at 15-minute intervals. Abnormal stations with a zero or missing value ratio exceeding 50% are deleted. For the remaining stations, forward padding and mean padding are used to process the missing data, and the power values ​​are truncated to a reasonable range of [0, 500] kW to eliminate the influence of extreme outliers. After removing abnormal stations, 187 stations are retained for modeling. The dataset is divided into a training set, a validation set, and a test set in a 7:1:2 ratio.

[0158] A multi-site charging load prediction network is constructed, which can perform the following operations:

[0159] Temporal feature processing: Four sets of one-dimensional convolutions with kernel sizes of 1, 3, 5, and 7 are deployed to capture instantaneous fluctuations, short-term cycles, and long-term trend features, respectively. After concatenation and linear fusion, convolutional attention gating and temporal self-attention are applied to obtain temporal feature representations.

[0160] Spatial feature processing: A geographic adjacency matrix is ​​constructed using Gaussian kernel bandwidth and connection threshold; a functional adjacency matrix is ​​constructed using cosine similarity threshold; geographic weights and functional weights are fused into a hybrid adjacency matrix, which is then normalized and processed by two layers of graph convolution and multi-head spatial attention to obtain spatial feature representations;

[0161] Price feature processing and price elasticity regularization: Price representation is obtained by embedding three features, namely electricity price, service price, and price change, through a two-layer MLP; a reference elasticity interval is constructed for each station with a reference benchmark of -0.45, a bounded correction range of 0.15, and a pruning interval of [-0.8, -0.1]; price sensitivity is diagnosed by automatic differentiation and a regularization loss term is constructed.

[0162] Cross-attention and fusion: The three features are projected onto a unified dimension, and cross-attention is performed according to the spatiotemporal backbone being queried by price and space being queried by time. Then, the fused features are obtained by adaptive association weights and gating fusion.

[0163] Periodic modeling and encoding output: Fast Fourier transform is performed on the fused features and the first four main periodic components are extracted. Multi-head attention is applied and gating fusion is performed at the short-term, medium-term and long-term scales. After superimposing the location encoding, the multi-site charging load prediction results for the next 24 hours are output through the Transformer encoding layer.

[0164] Joint training and evaluation: The prediction network is trained end-to-end by using the mean squared error loss plus a dynamically weighted price elasticity regularization term to form the total loss; and the trained prediction network is used to predict the charging demand of charging stations.

[0165] Figure 2 The figure shows the prediction results of the proposed method on the first six charging stations in the Shenzhen dataset. The red dashed line represents the predicted value, the blue solid line represents the actual value, and the horizontal axis represents the continuous time steps of the test set. The rolling prediction results and actual values ​​during the test period are compared, and the mean absolute error (MAE) of the corresponding stations is marked. It can be seen that the predicted curve and the actual curve generally match well in terms of peak and valley positions and fluctuation amplitude. Although the Shenzhen dataset has strong demand heterogeneity and complex inter-station dependencies, which significantly increases the prediction difficulty, the model can still accurately characterize the temporal evolution of charging load at each station in a very large-scale heterogeneous scenario, verifying the good generalization ability and robustness of the proposed method.

[0166] Figure 3 The distribution of the posterior mean diagnostic price sensitivity for each station after training is presented. Based on the converged elasticity values, the stations are divided into three categories: highly elastic (60 stations, ...). <−0.55), a priori anchoring type (75 stations, −0.55≤ ≤−0.3) and weakly elastic type (52 stations, >−0.3). The economic reference value of −0.45 and the overall mean of −0.437 are close, indicating that the overall response level of the model is consistent with the economic prior. The distribution of the three types of sites shows a certain deviation and broadening relative to their respective target reference values, indicating that the specific elasticity values ​​are mainly formed by data-driven losses, preserving the heterogeneity between sites while maintaining the overall economic rationality.

[0167] Figures 4(a) and 4(b) track the evolution of the diagnostic price sensitivity coefficient over 30 training rounds. Figure 4(a) shows that in the early strong regularization stage, the means of the three types of sites are distributed around the economic reference value of -0.45, with large within-group variances. After entering the weak regularization stage, the trajectories of the three types of means diverge significantly. The strong elasticity type gradually converges to about -0.67, the weak elasticity type converges to about -0.20, and the prior anchoring type remains around -0.42. The three sets of trajectories maintain a smooth and monotonous evolution without significant oscillations or regressions, indicating that the proposed regularization method has good gradient stability throughout the training process and maintains good compatibility with data-driven loss. Figure 4(b) shows the round-by-round tracking of the elasticity coefficients of six representative sites. The evolution trajectory of all sites remains within a reasonable range of [-0.8, -0.1], indicating that while the regularization term provides global directional constraints, it allows each site to form heterogeneous price responses based on its own data.

[0168] Example 3

[0169] This embodiment provides a multi-site charging load prediction system that considers diverse heterogeneous characteristics, including:

[0170] The historical charging power input sequence acquisition module is used to acquire historical charging demand data, geographical data, and price data from multiple sites, process the data, and construct a historical charging power input sequence.

[0171] A prediction network construction module is used to construct a multi-site charging load prediction network; the prediction network includes a cascaded heterogeneous feature three-branch encoder, a cross-attention collaborative expression module, a multi-scale periodic modeling module, a Transformer encoding module, and an output projection layer;

[0172] The prediction network training module is used to train the multi-site charging load prediction network using the obtained historical charging power input sequence to obtain the trained prediction network.

[0173] The charging demand prediction module is used to predict the charging demand of multiple charging stations for the next 24 hours using a trained prediction network.

[0174] A non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the multi-site charging load prediction method considering diverse heterogeneous characteristics as described above.

[0175] Furthermore, the present invention adopts the following technical solution:

[0176] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the multi-site charging load prediction method considering diverse heterogeneous characteristics as described above.

[0177] From the above description of the embodiments, those skilled in the art will clearly understand that the facilities of the present invention can be implemented using software plus necessary general-purpose hardware platforms. Embodiments of the present invention can be implemented using existing processors, or by dedicated processors used for this or other purposes for suitable systems, or by hardwired systems. Embodiments of the present invention also include non-transitory computer-readable storage media, comprising machine-readable media for carrying or having machine-executable instructions or data structures stored thereon; such machine-readable media can be any available medium accessible by a general-purpose or special-purpose computer or other machine with a processor. For example, such machine-readable media can include RAM, ROM, EPROM, EEPROM, CD-ROM or other optical disc storage, disk storage or other magnetic storage devices, or any other medium that can be used to carry or store the required program code in the form of machine-executable instructions or data structures and is accessible by a general-purpose or special-purpose computer or other machine with a processor. When information is transmitted or provided to a machine via a network or other communication connection (hardwired, wireless, or a combination of hardwired and wireless), that connection is also considered a machine-readable medium.

[0178] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after such changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A multi-site charging load prediction method considering diverse heterogeneous characteristics, characterized in that, Includes the following steps: Acquire historical charging demand data, geographical data, and price data from multiple sites, process the data, and construct a historical charging power input sequence; A multi-site charging load prediction network is constructed; the prediction network includes a cascaded heterogeneous feature three-branch encoder, a cross-attention collaborative expression module, a multi-scale periodic modeling module, a Transformer encoding module, and an output projection layer; The multi-site charging load prediction network is trained using the obtained historical charging power input sequence to obtain the trained prediction network. The trained prediction network is used to predict the charging demand of multiple charging stations for the next 24 hours.

2. The multi-site charging load prediction method considering multi-variable heterogeneous characteristics according to claim 1, characterized in that: The heterogeneous feature three-branch encoder includes a temporal branch encoder, a spatial branch encoder, and a price branch encoder. The temporal branch encoder includes a one-dimensional convolution with four sets of kernels, a convolutional attention gating unit, and a temporal self-attention layer. The spatial branch encoder includes a hybrid adjacency matrix construction module, a two-layer graph convolutional network, and a multi-head spatial attention layer. The price branch encoder includes a two-layer MLP price embedding network. The temporal branch encoder, spatial branch encoder, and price branch encoder each output three-branch features. , , .

3. The multi-site charging load prediction method considering multi-variable heterogeneous characteristics according to claim 1, characterized in that: The time branch encoder, spatial branch encoder, and price branch encoder each output three-branch features. , , The method is as follows: One-dimensional convolutions with four sets of kernels (1, 3, 5, and 7) are used to capture the features of charging demand at different time scales, including immediate fluctuations, short-term cycles, and long-term trends. A convolutional attention enhancement mechanism is introduced to generate feature-level attention weights through one-dimensional convolutions and adaptively gate the fused features. ; In the formula, Let B be the feature at the k-th scale, and B be the batch size. The total dimension of time features This is a one-dimensional convolution with a kernel size of k, and ReLU is the activation function. For bias vectors, To fuse the weight matrix, For bias vectors, The merged time features are represented by the concatenation operation, where Concat represents the concatenation operation. The enhanced temporal features; H1, H3, H5, and H7 represent the values ​​of k when k is 1, 3, 5, and 7, respectively. ; By modeling long-range dependencies within a sequence using a temporal self-attention mechanism, a temporal representation with both multi-scale local perception and global temporal modeling capabilities is generated. ; In the formula, , , These are query, key, and value matrices, respectively, all derived from... Linear projection yields, The dimension of the key vector. Here is the attention weight matrix. For self-attention output, Softmax is a normalized exponential function; Establish geographical adjacency relationships based on the geographical distance between sites: ; In the formula, The geographical distance between site i and site j For Gaussian kernel bandwidth, For connection threshold, This is the element in the i-th row and j-th column of the geographic adjacency matrix; The cosine similarity of power time series between sites is calculated, and a hybrid adjacency matrix is ​​obtained by fusing geographical and functional adjacency relationships. The adjacency matrix is ​​then normalized. ; In the formula: , Let i be the historical power time-series vector of stations i and j. Similarity scores are calculated only in the training set interval, and the functional adjacency matrix constructed accordingly remains fixed in the validation and test sets. For similarity threshold, The element in the i-th row and j-th column of the functional adjacency matrix. For geographical weight, For functional weights, It is a hybrid adjacency matrix. It is a diagonal matrix. It is the identity matrix. This is the normalized adjacency matrix; It is a geographic adjacency matrix; It is a functional adjacency matrix; Charging station topology encoding is achieved using a two-layer graph convolutional network. ; In the formula: The site node feature matrix is ​​obtained by linearly mapping the historical power sequences of each site along the time dimension. The convolution weights for the first layer graph. For the convolution weights of the second layer graph, and For bias, , These are the spatial features output by the first and second layer graph convolutions, respectively. For site node feature dimensions, , Output dimensions for two-layer graph convolution; A multi-head spatial attention mechanism is deployed on top of graph convolutional features to calculate dynamic interaction weights between sites, enabling cross-site information aggregation. Through neighborhood aggregation and feature transformation, a deep nonlinear transformation is performed on the node features to obtain the final spatial feature output. ; In the formula: For the number of attention heads, , , Let m be the query, key, and value projection matrix of the attention head. To output the projection matrix, For the output of the m-th attention head, This is the structure-aware projection matrix. For the final spatial feature output, For a single attention head dimension, For the aggregated output of multi-head spatial attention, For structure-aware transformation features; A price feature tensor is constructed using three characteristics: electricity price, service price, and price changes. : ; In the formula, Let be the standardized price feature vector of station i at time t. The standardized electricity price at station i at time t. The standardized service price for station i at time t. This represents the total price change; Price representation obtained through two-layer MLP embedding Output as the price branch: ; In the formula, For the first layer of embedded weights, For the second layer of embedded weights, , It is a bias vector. For price embedding features, For the price MLP hidden layer dimension, Price is embedded as a dimension.

4. The multi-site charging load prediction method considering multi-variable heterogeneous characteristics according to claim 3, characterized in that: The cross-attention collaborative representation module includes three linear projection layers and three cross-attention layers, projecting the three-branch features to a unified dimension. Then, the cross-attention mechanism is used for interaction, and the final fused feature is obtained by fusing the adaptive association weight network and the gating network.

5. The multi-site charging load prediction method considering multi-variable heterogeneous characteristics according to claim 4, characterized in that: The method for obtaining the final fusion feature is as follows: Projecting time, space, and price characteristics onto a unified dimension : ; in, , , These are projection matrices representing temporal, spatial, and price characteristics, respectively. , , Features after projection; The cross-attention mechanism is used to achieve collaborative representation among heterogeneous features. Spatiotemporal features, as the predictive backbone, are actively queried by price features to inject price response information, while space, as a structural representation, is queried by time to supplement structural information. ; in, For temporal-spatial cross-attention features, For price-time cross-attention features, Price-space cross-attention features; The adaptive association weights of the three-way cross features are learned and then gated to obtain the final fused features: ; In the formula, For adaptive association weight vector, Control vector, For element-wise product, Features after splicing This is the final fusion feature.

6. The multi-site charging load prediction method considering multi-variable heterogeneous characteristics according to claim 1, characterized in that: The multi-scale periodic modeling module includes a fast Fourier transform, an inverse Fourier transform, and a multi-head attention and gating fusion layer at three scales. The multi-scale periodic modeling module performs frequency domain analysis on the fused features and extracts the Top-K main periodic frequency components, retaining the first four main frequency components; the main periodic components are weighted with the original fused features through inverse Fourier transform to obtain periodic enhancement features, and multi-head attention is applied at the short-term, medium-term and long-term scales respectively. Multi-head attention at all three scales uses the complete sequence as the query and the corresponding time slice as the key and value, with a uniform output length. By learning the fusion weights of three time scales through a gating mechanism, hierarchical attention fusion features are obtained.

7. The multi-site charging load prediction method considering multi-variable heterogeneous characteristics according to claim 1, characterized in that: When training the multi-site charging load prediction network, the mean squared error loss and the price elasticity regularization term are weighted to form the total error loss, and the network is trained end-to-end: ; In the formula, For mean square error loss, For the total loss, For price elasticity regularization, As the weight of the regularization term, For the actual needs of station i at time t in batch n, Let B represent the predicted demand for station i at time t in batch n, where B is the batch size and N is the number of charging stations. To predict the sequence length.

8. The multi-site charging load prediction method considering multi-variable heterogeneous characteristics according to claim 7, characterized in that: The method for obtaining the price elasticity regularization term is as follows: By using automatic differential calculations to predict the partial derivative of demand with respect to electricity prices, a local sensitivity index of the network to price input is constructed: ; In the formula, For site i, the local diagnostic price elasticity at a single moment; The predicted demand is obtained by backpropagation of the automatic differential pair with respect to the input electricity price. and All values ​​are the original values ​​after standardization. Take 0.01; The reference target value is constructed by superimposing the comprehensive literature value on a bounded range given by the site's functional attributes: ; In the formula, The reference target elasticity for site i; Take 0.15; Based on the prior knowledge of the site's functional areas: Using the deviation between the diagnostic price sensitivity index and the reference target value as the regularization loss term, directional regularization is applied to the network's local response in the price dimension under the batch average meaning: ; In the formula, Obtained through automatic differentiation techniques: For the predicted demand of station i at time t in batch n, For the corresponding electricity price, take the electricity price at the end of the input window. As a price benchmark.

9. The multi-site charging load prediction method considering multi-variable heterogeneous characteristics according to claim 7, characterized in that: The regularization weights are dynamically adjusted according to the training phase. It provides strong directional guidance in the early stages and maintains a weak regularization effect in the later stages. ; In the formula, The training rounds are designed to provide directional guidance for parameter updates in the price dimension during the phase with a weight of 1.0; data-driven improvement dominates when the weight is reduced to 0.1; and weak regularization is retained during the phase with a weight of 0.01, ultimately achieving a balance between prediction accuracy and training process stability.

10. A multi-site charging load prediction system considering diverse heterogeneous characteristics, characterized in that, include: The historical charging power input sequence acquisition module is used to acquire historical charging demand data, geographical data, and price data from multiple sites, process the data, and construct a historical charging power input sequence. A prediction network construction module is used to construct a multi-site charging load prediction network; the prediction network includes a cascaded heterogeneous feature three-branch encoder, a cross-attention collaborative expression module, a multi-scale periodic modeling module, a Transformer encoding module, and an output projection layer; The prediction network training module is used to train the multi-site charging load prediction network using the obtained historical charging power input sequence to obtain the trained prediction network. The charging demand prediction module is used to predict the charging demand of multiple charging stations for the next 24 hours using a trained prediction network.