Charging pile order quantity prediction method and system based on evolutionary graph and attention mechanism

Abstract

The application discloses a charging pile order quantity prediction method and system based on an evolution graph and an attention mechanism, and the method comprises the following steps: performing sliding window processing on a space-time sequence matrix constituted by order quantities of each charging site, generating a sample pair composed of a historical window sequence and a future window sequence, and constructing a sample data set; performing instance normalization, time sequence information injection and channel expansion on the historical window sequence to obtain a high-dimensional feature graph; constructing a prediction model comprising a channel attention module, a multi-scale time feature extraction module, an evolution graph module and an output layer; performing supervised training on the prediction model by using the high-dimensional feature graph to obtain a trained prediction model; and predicting the input charging pile order quantity by using the trained prediction model and outputting a future order quantity. The application improves the model prediction precision and generalization ability.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of charging pile operation management, and particularly relates to a charging pile order quantity prediction method and system based on an evolution graph and an attention mechanism. BACKGROUND

[0002] With the rapid development of the new energy vehicle industry, charging piles, as key supporting facilities, face increasingly severe challenges in terms of layout optimization and operational efficiency. The number of charging pile orders shows significant spatiotemporal differences, such as significant differences in demand between weekdays and holidays, and between urban centers and suburbs. If there is an excess of charging piles, it will lead to idle resources and cost waste; if there is an insufficient number of charging piles, it will easily cause users to queue up and wait, reducing user experience and thus restricting the promotion and popularization of new energy vehicles.

[0003] In the field of charging pile order prediction, traditional time series prediction methods (such as ARIMA, exponential smoothing method, decision tree model, etc.) can capture the trend and periodicity of data to some extent, but they perform poorly when dealing with charging pile order data with complex dynamic spatiotemporal correlations. Such methods usually cannot effectively model the spatial dependence between different charging sites, nor can they adapt to the dynamic changes in spatial structure over time, resulting in limited prediction accuracy.

[0004] In recent years, deep learning-based time series prediction methods have made progress in various fields, to some extent improving prediction performance. However, existing methods still have the following outstanding problems:

[0005] (1) Insufficient use of spatial information: Most models focus only on feature extraction in the time dimension, ignoring the geographical correlation and interactive influence between charging sites, resulting in prediction results deviating from the actual spatial distribution pattern;

[0006] (2) Insufficient modeling ability of long sequence dependency: In the face of long time span order sequences, models are prone to gradient disappearance or overfitting, making it difficult to capture deep dependencies across time periods;

[0007] (3) Poor adaptation to non-stationary sequences: Charging pile order data is influenced by multiple factors such as holidays, weather, and user behavior, exhibiting obvious non-stationary characteristics. Existing methods have poor fitting ability and generalization performance for such data, with a low correlation coefficient (CORR) between predicted values and true values, making it difficult to meet actual scheduling needs.

[0008] Therefore, there is an urgent need to develop a charging pile order quantity prediction method that can integrate spatiotemporal features, adaptively learn dynamic spatial dependencies, and have strong generalization ability, to support intelligent scheduling and resource optimization of charging facilities. SUMMARY

[0009] The application aims to provide an order quantity prediction method and system for charging piles based on an evolutionary graph and an attention mechanism, so as to solve the problem of low prediction accuracy and poor generalization ability caused by ignoring dynamic spatial dependence and weak modeling ability for non-stationary time series in the prior art.

[0010] The application solves the above technical problems through the following technical scheme: an order quantity prediction method for charging piles based on an evolutionary graph and an attention mechanism, comprising:

[0011] Performing sliding window processing on a space-time sequence matrix composed of order quantities of each charging site to generate a sample pair composed of a historical window sequence and a future window sequence, so as to construct a sample data set;

[0012] Performing instance normalization, time sequence information injection and channel expansion on the historical window sequence to obtain a high-dimensional feature map;

[0013] Constructing a prediction model comprising a channel attention module, a multi-scale time feature extraction module, an evolutionary graph module and an output layer; performing supervised training on the prediction model using the high-dimensional feature map to obtain a trained prediction model;

[0014] The channel attention module is configured to perform adaptive channel weighting on the high-dimensional feature map.

[0015] The multi-scale time feature extraction module is configured to extract and fuse multi-scale time features from the weighted feature map.

[0016] The evolutionary graph module comprises:

[0017] A node extraction unit configured to extract static node features from the space-time sequence matrix and map them to a dynamic graph embedding space;

[0018] An evolutionary graph learning unit configured to take the output of the node extraction unit as an initial state, construct spatial dependence relationships between charging sites based on multi-scale time features, and generate a dynamic graph structure;

[0019] A graph convolution unit configured to perform feature propagation and aggregation on the dynamic graph structure, realize deep fusion of space-time features, and obtain a space-time feature tensor;

[0020] The output layer is configured to map the space-time feature tensor to a single-channel feature tensor, perform inverse normalization processing, and output a future order quantity prediction result for each charging site.

[0021] The trained prediction model is used to predict the order quantity of input charging piles, and a future order quantity is output.

[0022] The application adopts a reversible instance normalization (RevIN) technology (including instance normalization and inverse normalization), effectively eliminates the non-stationarity in the charging pile order quantity, makes the model training process more stable and converges faster, thereby directly improving the accuracy of the prediction result. At the same time, the channel attention module and the multi-scale time feature extraction module work together to ensure that the model can adaptively focus on key features and capture complex time dependence from short-term to long-term, further reducing the prediction bias.

[0023] The core innovation of the application is to dynamically construct a spatial correlation graph between charging stations by using an evolutionary graph learning unit, overcoming the inherent defect that traditional static graph models cannot reflect the dynamic evolution of space-time relationship; and combining a graph convolution unit for feature fusion, so that the model can accurately learn and quantify complex spatial patterns, fundamentally solving the technical problem that existing methods ignore or cannot effectively model spatial dependence.

[0024] Because the application specially processes the non-stationarity of data (RevIN) and deeply enhances and fuses the space-time features, the model has stronger adaptability to data of different distributions (such as data of different regions and different time periods). This means that the model can still maintain reliable prediction performance when encountering unseen data patterns, showing excellent generalization ability and being suitable for large-scale and complex charging pile network prediction tasks in complex environments.

[0025] The application restores the final single-channel feature tensor to the original order quantity dimension by performing inverse normalization processing at the output layer. This makes the prediction result no longer an abstract numerical value, but an order quantity with direct physical meaning, which can seamlessly connect to practical application systems such as charging pile intelligent scheduling and resource optimization configuration, greatly improving the practical value of the method.

[0026] Further, the historical window sequence is normalized, including:

[0027] The statistical quantity of each charging station corresponding to the time sequence in the historical window sequence is calculated;

[0028] The corresponding time sequence is normalized based on the statistical quantity.

[0029] Further, the time sequence information injection is realized by adding the position encoding vector to the normalized sequence; the position encoding vector is generated according to the following formula:

[0030] ;

[0031] ;

[0032] Where PE represents the position encoding vector. , respectively represent the values of the position encoding vector at , 2i represents an even dimension index in the position encoding vector, and 2i+1 represents an odd dimension index in the position encoding vector; t represents the position index of the time step in the normalized sequence; N represents the total dimension of the position encoding vector.

[0033] Further, the channel attention module is SENet, and the SENet is used for adaptive channel weighting of the high-dimensional feature map, which comprises:

[0034] performing global average pooling on each channel of the high-dimensional feature map in the space-time dimension to obtain a channel global scalar;

[0035] extracting features of the channel global scalar through a multilayer perceptron with a bottleneck structure to obtain a channel attention weight vector;

[0036] performing channel-by-channel multiplication on the high-dimensional feature map and the channel attention weight vector, and then performing residual connection with the high-dimensional feature map to obtain a weighted feature map.

[0037] Further, the multi-scale time feature extraction module is a multi-branch dilated convolution module; the multi-branch dilated convolution module comprises a multi-scale time convolution module and a Transformer encoder, the multi-scale time convolution module adopts a parallel double-path gating structure and both paths comprise a multi-scale dilated convolution submodule, the multi-scale dilated convolution submodule is implemented through a plurality of parallel branches, and each branch uses a convolution layer with different convolution kernel widths but the same dilation rate to extract time features of different scales;

[0038] extracting multi-scale time features from the weighted feature map using the multi-branch dilated convolution module, which comprises:

[0039] extracting multi-scale local time features from the weighted feature map in parallel using the multi-scale dilated convolution submodule, and adaptively fusing the multi-scale local time features through a gating mechanism to obtain a preliminary fused feature representation;

[0040] encoding the preliminary fused feature representation using the Transformer encoder to capture global long-term dependencies, thereby obtaining final multi-scale time features.

[0041] Further, the node extraction unit comprises:

[0042] a CNN-based feature extractor configured to extract static node features from the spatio-temporal sequence matrix;

[0043] a linear transformation layer configured to map the static node features to a dynamic graph embedding space.

[0044] Further, the evolutionary graph learning unit is configured to take the output of the node extraction unit as an initial state, and construct a spatial dependency relationship between charging stations based on the multi-scale time features to generate a dynamic graph structure, including:

[0045] The multi-scale time features are divided into a plurality of sub-segments according to time intervals, and each sub-segment is aggregated by an aggregation operator to obtain segmented aggregated features;

[0046] The node embedding state is updated by a gated recurrent unit according to the segmented aggregated features and a previous time hidden state;

[0047] The dynamic graph structure is generated by a dual-channel MLP according to the updated node embedding.

[0048] Further, the graph convolution unit is a hybrid graph convolution network configured to perform feature propagation and aggregation on the dynamic graph structure to realize deep fusion of spatio-temporal features, including:

[0049] The graph convolution operation is performed on the dynamic graph structure to realize deep fusion of spatio-temporal features by aggregating feature information from multi-order neighborhoods and fusing with the node feature information.

[0050] Based on the same concept, the application also provides a charging pile order quantity prediction system based on evolutionary graph and attention mechanism, including a memory, a processor and a computer program or instructions stored in the memory, the processor executes the computer program or instructions to realize the charging pile order quantity prediction method as described above.

[0051] Compared with the prior art, the application has the following advantages:

[0052] The application effectively eliminates the non-stationarity in the charging pile order quantity by introducing reversible instance normalization technology, improves the stability and convergence speed of model training, and combines channel attention mechanism and multi-scale time feature extraction, so that the model can accurately capture local details, periodic rules and global long-time dependency in time series at the same time, and improve the prediction accuracy of the model.

[0053] The evolutionary graph module of the application can dynamically learn and update the correlation strength between charging stations according to time series data to generate a dynamic graph structure that changes over time. This enables the model to adaptively capture complex dynamic spatial interactions (such as peak demand overflow, adjacent station competition effect, etc.), fundamentally improving the model's ability to depict real-world scenarios.

[0054] The synergistic work of the evolution graph learning unit and the graph convolution unit realizes deep spatio-temporal feature fusion. The evolution graph learning unit is responsible for perceiving dynamic spatial structure, and the graph convolution unit performs multi-order neighborhood information propagation and aggregation on the dynamic spatial structure. This structure learning and feature learning separation and close combination architecture enables the model to fully mine the internal correlation in spatio-temporal data, thereby making more accurate predictions. BRIEF DESCRIPTION OF DRAWINGS

[0055] In order to more clearly illustrate the technical solutions of the present application, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings described below are only one embodiment of the present application, and other drawings can also be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0056] Figure 1 is the flow chart of the charging pile order quantity prediction method in the embodiment of the present application;

[0057] Figure 2 is the total number of orders of all charging sites in each time period in a day in the embodiment of the present application;

[0058] Figure 3 is the total number of orders of all charging sites in each day in a week in the embodiment of the present application;

[0059] Figure 4 is the architecture diagram of the SENet in the embodiment of the present application;

[0060] Figure 5 is the architecture diagram of the multi-branch dilated convolution module in the embodiment of the present application;

[0061] Figure 6 is the architecture diagram of the multi-scale dilated convolution submodule in the embodiment of the present application;

[0062] Figure 7 is the architecture diagram of the CNN-based feature extractor in the embodiment of the present application;

[0063] Figure 8 is the architecture diagram of the evolution graph learning unit in the embodiment of the present application. DETAILED DESCRIPTION

[0064] The technical solutions in the present application will be described clearly and completely below in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0065] The technical solutions of the present application are described in detail below with specific examples. The following specific examples can be combined with each other, and the same or similar concepts or processes may not be described in detail in some examples.

[0066] Example 1

[0067] Figure 1 A flow chart of the charging pile order quantity prediction method provided by the present application is shown. As shown in Figure 1 The charging pile order quantity prediction method provided by the present application includes the following steps:

[0068] Step 1: Construct a sample data set.

[0069] An original data set containing order number, charging pile number, site information, order start time, and order end time is obtained from a charging pile operation management platform. The order start time and order end time refer to the start time and end time of the user using the charging pile to charge the electric vehicle. Part of the information of the original data set is shown in Table 1.

[0070] Table 1 shows part of the information of the original data set:

[0071]

[0072] From the original data set in Table 1, it can be seen that the excessive amount of information is not suitable for model training, so the original data set is preprocessed to obtain a data set that can be used for model training. The preprocessing of the present application is: eliminating noise data with incomplete or incorrect information in the original data set; obtaining the time distribution information of the original data set, and statistically counting the order quantity of each charging site at intervals of time period (e.g. 1h), and finally summarizing the statistical results to obtain a space-time sequence matrix. The column of the space-time sequence matrix represents the space number of each charging site, and the row represents the charging pile order quantity of a certain charging site within a certain time period, as shown in Table 2. Each charging site in Table 2 contains multiple charging piles, and the order quantity of a certain charging site refers to the sum of the order quantities of all charging piles in the charging site.

[0073] In order to understand the characteristics of the space-time distribution and periodic trend of the charging pile order quantity, data analysis is performed on the space-time sequence matrix. In order to more clearly show the time series characteristics of the space-time sequence matrix, time trend analysis is performed first. Figure 2 The total order quantity of all charging sites in each time period within a day is shown. As Figure 2 It can be seen that the order quantity is relatively high from 23:00 to 5:00 the next day every day, indicating that most new energy vehicle users' electricity demand is concentrated in the evening rest time. Figure 3 The total order quantity of all charging sites in each day within a week is shown. As Figure 3It can be seen that the total order quantity from Monday to Sunday in a week shows a slight downward trend, and the difference between each day is not large, indicating that the order quantity changes relatively smoothly within a week, but the order quantity on weekdays is larger than that on weekends, indicating that most new energy vehicle users have greater demand for vehicle use and charging on weekdays than on weekends, which also reflects the characteristics of the data distribution trend having certain time dependence.

[0074] Through analysis, it is known that the space-time sequence matrix has obvious rules and trends, and the model can achieve more accurate prediction by accurately capturing such time dependence.

[0075] Table 2 is part of the space-time sequence matrix:

[0076]

[0077] According to the time step (or window width), the space-time sequence matrix X is processed by sliding window to generate sample pairs composed of historical window sequences and future window sequences, and all sample pairs constitute a sample dataset. Let the number of charging stations be N, the total length of time be T all , and the time step be T, then the space-time sequence matrix X is a matrix of dimension N×T all , the historical window sequence is a matrix of dimension N×T, the future window sequence is a matrix of dimension N×T pred , and T pred represents the prediction step.

[0078] In order to facilitate reading and processing of data of different dimensions, the sample dataset in CSV format is converted to H5 format. H5 file format (Hierarchical Data Format version 5) as a hierarchical file structure, it has good performance in efficient data compression and storage tasks, when processing large-scale multivariate time series data, H5 format can significantly reduce the storage space occupation, at the same time, speed up the data reading and writing speed, which is very important for model training and prediction process that need to frequently access and process a large amount of data, can improve the efficiency of the whole process, and after using H5 format, the model can easily read and process data of different dimensions without complex data conversion.

[0079] Step 2: instance normalization, time sequence information injection and channel expansion are performed on the historical window sequence to obtain a high-dimensional feature map.

[0080] In order to solve the problem of weak non-stationary time series modeling ability, the application adopts reversible instance normalization (RevIN) technology to eliminate the non-stationarity of input time series and improve the stability and convergence speed of model training. The core idea of RevIN is to normalize each sample (i.e. each historical window sequence) at the input end of the model and perform accurate de-normalization at the output end of the model, thereby ensuring the physical meaning of the prediction results.

[0081] The reversible instance normalization includes an instance normalization stage and a de-normalization stage. Before the sample is input into the model, the historical window sequence is normalized, including:

[0082] Step 2.1: Calculate the mean and standard deviation of the time series corresponding to each charging station in the historical window sequence;

[0083] Step 2.2: Normalize the corresponding time series based on the mean and standard deviation:

[0084] (1)

[0085] wherein, represents the time series of the nth charging station in the bth historical window sequence in batch B; represents the normalized time series; represents the mean of the time series , , represents the order quantity at the tth time point in the time series ; represents the standard deviation of the time series , , represents a constant to prevent division by zero, and in the embodiment is set to 1e -5 .

[0086] After the model outputs a single-channel feature tensor , the calculated mean and standard deviation are used for de-normalization to restore the original dimension to obtain the future order quantity prediction results of each station . The specific calculation formula is:

[0087] (2)

[0088] In the specific embodiments of the present application, the time sequence information injection is realized by adding the position encoding vector to the normalized sequence. The position encoding vector aims to provide the model with the position information of the input data, enhancing the model expression capability. The present application uses sine and cosine functions to generate the position encoding, and the sine and cosine functions have periodicity, and the combination of sine and cosine functions of different frequencies can generate a unique encoding vector for each position. The design of the position encoding ensures that it not only contains absolute position, but also reflects relative position information: that is, the relationship between any two position encoding vectors can reflect their relative positions in the sequence. The position encoding vector is generated according to the following formula:

[0089] (3)

[0090] (4)

[0091] wherein PE represents the position encoding vector; , respectively represent the values of the position encoding vector at , , 2i represents the even dimension index in the position encoding vector, and 2i+1 represents the odd dimension index in the position encoding vector; represents the position index of the time step in the normalized sequence; represents the total dimension of the position encoding vector.

[0092] The sine and cosine functions can generate encoding with periodicity and relative position information, which can effectively capture the position relationship in the sequence and improve the accuracy of the model in prediction. In addition, different encoding methods are also applicable to different input data and tasks, improving the generalization ability of the model.

[0093] In order to adapt to the subsequent two-dimensional convolution-based multi-scale time feature extraction module, the sequence injected with time sequence information is expanded and transformed in the channel dimension to obtain a high-dimensional feature map. The sequence injected with time sequence information is regarded as a special two-dimensional feature map: first, the dimension is expanded to [B, 1, N, T] by torch. unsqueeze() operation, wherein B represents the batch size, and 1 represents the initial single channel; then, an initial convolution layer (self. start_conv) is used to expand the channel number from 1 to C (such as 16), to obtain a high-dimensional feature map , denotes the set of all real matrices with shape [B, C, N, T]. The sequence is structurally converted into a two-dimensional image format with B batches, C channels, N pixel height (number of charging stations), and T pixel width (time steps), but its physical meaning is still a spatio-temporal sequence matrix of B samples, N charging stations, and T time steps.

[0094] Step 3: Constructing a prediction model, the prediction model includes a channel attention module, a multi-scale time feature extraction module, an evolutionary graph module, and an output layer, the evolutionary graph module includes a node extraction unit, an evolutionary graph learning unit, and a graph convolution unit.

[0095] The channel attention module is used for adaptive channel weighting of the high-dimensional feature map U; the multi-scale time feature extraction module is used for extracting and fusing multi-scale time features from the weighted feature map; the node extraction unit is used for extracting static node features from the spatio-temporal sequence matrix X and mapping to a dynamic graph embedding space; the evolutionary graph learning unit is used to construct a spatial dependency relationship between charging stations based on multi-scale time features with the output of the node extraction unit as the initial state, to generate a dynamic graph structure; the graph convolution unit is used for feature propagation and aggregation on the dynamic graph structure, realizing deep fusion of spatio-temporal features, to obtain a spatio-temporal feature tensor; the output layer is used to map the spatio-temporal feature tensor to a single-channel feature tensor and perform inverse normalization (such as formula (2)), and output the future order quantity prediction results of each charging station.

[0096] In the specific embodiments of the present application, the channel attention module is an SE-Net (as shown in Figure 4 ), which uses SE-Net to adaptively weight the high-dimensional feature map U, including:

[0097] Squeeze operation: global average pooling is performed on each channel of the high-dimensional feature map in the space-time dimension (N×T) to obtain a channel global scalar Z. The squeeze operation captures the global context information of each channel in the entire spatio-temporal range. For the c-th channel, the output of the squeeze operation is :

[0098] (5)

[0099] wherein represents the feature value of the c-th channel in the high-dimensional feature map at the i-th charging station and the j-th time step, . The size of the channel global scalar Z is B×C×1×1, which is adjusted to B×C after view adjustment.

[0100] Excitation operation: a multi-layer perceptron (MLP) with a bottleneck structure is used to learn the nonlinear interaction relationship between channels and generate a channel attention weight vector S. As shown inFigure 4 As shown, the multi-layer perception (MLP) with a bottleneck structure has two fully connected layers (FC), the first fully connected layer reduces the channels from C to C / a (a is the reduction ratio) and uses the ReLU activation function; the second fully connected layer restores the channels to C and uses the Sigmoid activation function to generate a weight between 0 and 1 for each channel , representing the importance of the channel:

[0101] (6)

[0102] wherein sigmoid represents the sigmoid activation function, ReLU represents the ReLU activation function, W1 and W2 are weight matrices of the two fully connected layers respectively, b1 and b2 are bias vectors of the two fully connected layers respectively, represents the scalar of the i-th channel in the global scalar Z of the channel, represents the weight of the i-th channel, The channel attention weight vector is composed of the weights of all channels , and the size changes from BxC to BxCx1x1, which is convenient for subsequent operations.

[0103] Reweight and residual connection: the high-dimensional feature map is multiplied by the channel attention weight vector channel by channel, which strengthens important features and suppresses secondary features; then, a residual connection (shortcut connection) is introduced, which adds the feature map multiplied by the channel to the high-dimensional feature map U to alleviate the gradient vanishing problem and ensure the stability of the model training. The weighted feature is:

[0104] (7)

[0105] wherein represents the weighted feature, and the size is BxCxNXT; represents the weight of the c-th channel, represents the feature of the c-th channel in the high-dimensional feature map U.

[0106] Through the channel attention mechanism, the model can adaptively focus on those feature channels that are most effective for the charging order quantity prediction task, thereby improving the ability of feature representation. The addition of the residual connection further guarantees the effective training of the deep network, provides a direct propagation path for the gradient, so that the gradient can be more smoothly backpropagated, thereby alleviating the gradient vanishing problem and helping the training and convergence of the network.

[0107] ​In the specific embodiments of the present application, the multi-scale time feature extraction module is a multi-branch dilated convolution module. As shown in Figure 5 , the multi-branch dilated convolution module includes a multi-scale time convolution module (TConv) and a Transformer encoder, the multi-scale time convolution module adopts a parallel double-path gating structure and both paths contain a multi-scale dilated convolution sub-module, the multi-scale dilated convolution sub-module is implemented through 4 parallel branches, each branch uses a convolution layer with different kernel width but the same dilation rate to extract time features of different scales, as shown in Figure 6 . The multi-branch dilated convolution module captures short-term and long-term temporal dependencies by using different scales of dilated convolution kernels in parallel, and combines the gating mechanism to fine-tune the information flow, providing a rich feature basis for prediction.

[0108] As shown in Figure 5 , both paths of the multi-scale time convolution module are composed of a multi-scale dilated convolution sub-module (Dilated_Inception), which are used as filters (Filter) and gates (Gate), respectively. The core of the design of the multi-scale dilated convolution sub-module is to use parallel convolution paths and dilated convolution technology to explicitly capture multi-scale temporal context information. As shown in Figure 6 , the multi-scale dilated convolution sub-module of the present embodiment contains 4 branches, each branch uses an independent two-dimensional convolution layer, the size of the convolution kernel is (1, kernel), in represents the input channel number of the convolution, out represents the output channel number of the convolution, where kernel is a predefined different value in kernel_set={2, 3, 6, 7}, and the 4 branches share the same dilation rate dilation=(1, 2). The convolution kernel height is fixed at 1, which ensures that the operation does not mix across spatial charging sites, maintaining the feature independence of each charging site. The convolution kernel width kernel is different for each branch, which directly determines the basic scale of the receptive field of each branch. The introduction of the dilation rate increases the effective receptive field of each branch. For a convolution layer with a kernel width of kernel and a dilation rate of dilation, the effective receptive field size can be calculated by the following formula:

[0109] (8)

[0110] For example, when , , This enables the model to efficiently capture rich temporal features from local details to long-term trends. Due to the different convolution kernel sizes of each branch, the output sequence length may differ slightly. A synchronous cropping strategy is adopted to unify the output of all branches to the length of the shortest sequence to ensure consistency for subsequent concatenation operations. Finally, the outputs of each branch are concatenated in the channel dimension (torch.cat) to form features that integrate multi-scale temporal context information.

[0111] To dynamically adjust the multi-scale feature flow and enhance the non-linear expression ability, the outputs of two multi-scale dilated convolution sub-modules (referred to as filter_conv and gate_conv, respectively) are fused. The output of filter_conv is transformed to the interval [-1, 1] by the tanh activation function, serving as the information content. The output of gate_conv is transformed to the interval [0, 1] by the sigmoid activation function, serving as the control signal for the flow of information at each scale. Finally, the above two results are multiplied element by element (Hadamard Product) to obtain the output of the multi-scale temporal convolution module :

[0112] (9)

[0113] where sigmoid denotes the sigmoid activation function, and denote the outputs of filter_conv and gate_conv, respectively; is the Hadamard product, representing element-wise multiplication.

[0114] The gating mechanism can adaptively evaluate the importance of features at different scales and perform weighted fusion accordingly, significantly improving the model's ability to model complex temporal dynamics and robustness.

[0115] Since the features output by the multi-scale temporal convolution module cannot be directly input into the Transformer encoder, which only accepts three-dimensional tensors, the features output by the multi-scale temporal convolution module are mapped after serialization adaptation to to facilitate input into the Transformer encoder. The specific implementation of the serialization adaptation mapping is as follows:

[0116] (10)

[0117] (11)

[0118] (12)

[0119] Equation (10) represents the rearrangement of dimensions, putting the time dimension T in front of the sequence axis; Equation (11) represents the adjustment of the tensor shape, merging the 3rd / 4th dimensions (channel x node) into a single feature dimension, each time step corresponding to a vector of length Equation (12) represents the linear mapping of each time step’s vector to the encoder embedding dimension d, resulting in , is a set of all real matrices of shape [B, T, d]. is the standard input shape of the transformer encoder (batch size x sequence length x embedding dimension). permute represents the rearrangement of dimensions (PyTorch’s permute operation), rearranging the dimensions from [B, C, N, T] to [B, T, C, N]; reshape represents the shape transformation (PyTorch’s reshape or view), transforming the dimensions from [B, T, C, N] to [B, T, C x N]; is the projection matrix, projecting the space-channel feature of dimension to dimension d; is the bias vector, and d represents the encoder embedding dimension.

[0120] Inputting into the transformer encoder, the output of the transformer encoder is

[0121] (13)

[0122] (14)

[0123] (15)

[0124] wherein K is the number of attention heads, is the embedding dimension of each attention head, and d represents the encoder embedding dimension; are the corresponding projection and output weights; , , and represent the query vector, key vector, and value vector, respectively; represents the intermediate quantity; is the query / key / value of the hth attention head, and the superscript in Equation (14) represents the transpose operation of the matrix; is the normalization along the time dimension of the key, is the scaling item, is the concatenation in the “head” dimension,​ is a layer normalization, is a linear forward propagation layer. represents a residual connection between the sequence representation after multi-head attention mechanism processing and the original input sequence, and then a normalized intermediate representation is obtained through layer normalization.

[0125] The output result of the transformer encoder is then adapted through serialization to be rearranged into multi-scale time features as input in the evolutionary graph learning unit.

[0126] Through the cascade and collaborative work of the multi-scale time convolution module and the transformer encoder, the application can simultaneously capture rich features from local details to global trends in time series and model extremely long-range dependencies, significantly improving the prediction ability of the complex spatiotemporal sequence data of the charging pile order quantity.

[0127] In the specific embodiments of the application, the node extraction unit includes a CNN-based feature extractor and a linear transformation layer, and the CNN-based feature extractor is used to extract static node features from the spatiotemporal sequence matrix X. As shown in Figure 7 The CNN-based feature extractor adopts a hierarchical convolution feature extraction measurement, which includes two one-dimensional convolution layers conv1d and three batch normalization layers BatchNorm1d, and finally outputs a fixed-dimensional feature vector through a fully connected layer FC. The input spatiotemporal sequence matrix X will be feature-extracted and converted into a uniform-dimensional feature vector after passing through the CNN-based feature extractor, facilitating subsequent processing.

[0128] (16)

[0129] wherein, represents the input spatiotemporal sequence matrix, N represents the number of stations, T all represents all time lengths; two one-dimensional convolution operations; represents a flattening operation, i.e., rearranging a multi-dimensional feature tensor into a two-dimensional tensor; is a fully connected layer weight matrix in the feature extractor, ReLU represents a ReLU activation function, and BN represents batch normalization, is the output static node feature, with a shape of [N, D s ]. The static node feature is used for subsequent learning and convolution operations of the dynamic graph structure.

[0130] The linear transformation layer maps the static node features extracted by the feature extractor to a dynamic graph embedding space through linear transformation, and obtains As the initial state of the gated recurrent unit GRU in the evolutionary graph learning unit:

[0131] (17)

[0132] wherein, is a full connection layer weight matrix in the linear transformation layer, is a static node feature.

[0133] The evolutionary graph learning unit adopts a GRU-based architecture, as shown in Figure 8 In the evolutionary graph learning unit, the multi-scale time features Q are divided into multiple sub-segments according to time intervals, and the time length of the tth sub-segment is r, and each sub-segment is aggregated by an aggregation operator AGG to obtain a segmented aggregated feature, and the segmented aggregated feature is taken as the input of the GRU:

[0134] (18)

[0135] wherein, denotes the segmented aggregated feature of the tth sub-segment; r denotes the time length of the sub-segment; denotes the feature of the kth time step of the tth segment; AGG is an average pooling in the time dimension, i.e., an aggregation operator.

[0136] The segmented aggregated feature of each sub-segment is input into the GRU together with the hidden state of the previous moment to obtain a new node embedding. The new node embedding is input into a double-channel MLP (MLP e and MLP m ), respectively, to generate a support matrix and a mask matrix of the graph, and then the mask matrix is activated by a sigmoid function, and finally the support matrix and the activated mask matrix are Hadamard multiplied to obtain a dynamic graph structure , i.e., a dynamic graph adjacency matrix.

[0137] MLP e is used to learn an “original graph structure candidate”, which is equivalent to predicting which nodes should have edges and the strength of the edges, and outputs a support matrix reflecting the potential connection strength between nodes. MLP m is used to output a mask matrix to “gate filter” the support matrix, and determine which edges are retained or weakened through a sigmoid activation function.

[0138] In the specific embodiments of the present application, the graph convolution unit is a hybrid graph convolution network, and feature propagation and aggregation are performed on a dynamic graph structure using the hybrid graph convolution network MixProp to realize deep fusion of spatio-temporal features and obtain a spatio-temporal feature tensor. Specifically, the graph convolution operation is performed on the dynamic graph structure using the hybrid graph convolution network MixProp, deep fusion of spatio-temporal features is realized by aggregating feature information from multiple orders of neighborhoods and fusing with the node itself feature information, the deep fused features corresponding to each sub-fragment are spliced in the time dimension to obtain a spatio-temporal feature tensor, the spatio-temporal feature tensor is mapped to a single-channel feature tensor and is subjected to inverse normalization processing, and the future order quantity prediction results of each charging station are output.

[0139] Using learned dynamic graph adjacency matrix , for the input fragment Graph convolution propagation:

[0140] (19)

[0141] wherein, is the graph convolution output feature (i.e., the deep fused feature) of the tth fragment; is the tth sub-fragment; is the learned dynamic graph adjacency matrix of GRU, MLPe and MLPm, and MixProp is a hybrid graph convolution network.

[0142] The graph convolution output features of all sub-fragments are spliced in the time dimension to obtain the final result:

[0143] (20)

[0144] wherein, is the final output spatio-temporal feature tensor containing the convolution results of all sub-fragments; is the graph convolution output feature of the tth fragment, is spliced along the time dimension, and the original batch dimension, channel dimension and node dimension are kept unchanged; is the number of channels; the superscript m represents the number of sub-fragments.

[0145] Spatio-temporal feature tensor The input is mapped to a channel mapping module. The module is stacked by a nonlinear activation function and a 1x1 convolution layer in sequence. The feature expression capability is enhanced by nonlinear transformation first, and then the feature channel number is gradually reduced from the intermediate channel dimension C (for example, 32) to the final single-channel output by using point-by-point convolution. Specifically, the tensor shape H of the input to the channel mapping module is (B, C, N, T), and the tensor Y1 shape of the output result is (B, 1, N, T), where B is the batch size, C is the channel number, N represents the number of charging stations, and T represents the time step. Then, the inverse normalization is performed using formula (2) to obtain , and finally, the interception operation is performed on the time dimension, that is, only the previous T pred time is taken as the final prediction result:

[0146] (21)

[0147] wherein, represents the prediction step; represents the feature tensor after channel mapping output and inverse normalization, represents the final prediction result.

[0148] After a series of evolution and graph structure operations, the original time series features are transformed into high-dimensional feature sequences rich in spatio-temporal context information. Under the driving of the evolutionary graph learning unit, the model can accurately capture the dynamic spatio-temporal rules and deeply mine the complex interaction relationships between nodes. For example, when the order quantity of a charging station increases, it may decrease due to demand diversion or competition effect; the data distribution of different time periods may also have significant differences. The evolutionary graph learning unit can dynamically identify such relationships, and then effectively model such nonlinear spatio-temporal correlations through graph convolution operations, significantly improving the prediction accuracy.

[0149] The core function of the mixed graph convolution network MixProp is to efficiently propagate and aggregate features between nodes based on dynamically generated graph structures, thereby enhancing the model's representation ability for complex dependency relationships. Specifically, MixProp uses a multi-order neighborhood information aggregation mechanism to integrate features of neighbor nodes of different hops on dynamically changing graph structures, thereby more comprehensively describing the local and global context of nodes. In addition, MixProp uses a residual or gating mechanism to adaptively weight and fuse the original node features and aggregation results, preserving the node's own characteristics while introducing neighborhood information, improving the model's adaptability and stability to dynamic graph structures.

[0150] The evolutionary graph learning unit and MixProp are closely coordinated and complementary in function: the former serves as a "graph structure generator" and is responsible for dynamically inferring the dependency relationship between nodes from data to output a support matrix and a mask matrix to construct a time-varying graph topology; the latter serves as a "feature processor" and relies on the graph structure provided by the former to perform multi-level feature propagation and aggregation, thereby realizing deep spatio-temporal feature fusion. The two together form a complete "structure learning-feature learning" closed loop, effectively improving the modeling capability and prediction performance of the model on dynamic spatio-temporal data.

[0151] Step 4: Supervised training of the prediction model using the high-dimensional feature map to obtain a trained prediction model.

[0152] In the specific embodiments of the present application, the supervised training of the prediction model using the high-dimensional feature map includes:

[0153] Step 4.1: Adaptive channel weighting of the high-dimensional feature map;

[0154] Step 4.2: Extraction and fusion of multi-scale time features from the weighted feature map;

[0155] Step 4.3: Extraction of static node features from the spatio-temporal sequence matrix and mapping to a dynamic graph embedding space;

[0156] Step 4.4: Construction of spatial dependency relationships between charging sites based on multi-scale time features to generate a dynamic graph structure using the output of step 4.3 as the initial state;

[0157] Step 4.5: Feature propagation and aggregation on the dynamic graph structure to realize deep fusion of spatio-temporal features and obtain a spatio-temporal feature tensor;

[0158] Step 4.6: Mapping of the spatio-temporal feature tensor to a single-channel feature tensor and performing inverse normalization to output the future order quantity prediction results for each charging site;

[0159] Step 4.7: Calculation of a loss function based on the future order quantity prediction results for each charging site and the corresponding future window sequence, adjustment of the parameters (including weight parameters and biases) of the prediction model based on the loss function, and realization of supervised training.

[0160] Step 5: Prediction of the input charging pile order quantity using the trained prediction model to output the future order quantity.

[0161] The input charging pile order quantity is counted to obtain a to-be-predicted spatio-temporal sequence matrix; the to-be-predicted spatio-temporal sequence matrix is subjected to instance normalization, time sequence information injection and channel expansion to obtain a to-be-predicted high-dimensional feature map; the to-be-predicted high-dimensional feature map is input into the trained prediction model to obtain a single-channel feature tensor; and the single-channel feature tensor is subjected to inverse normalization processing to obtain a prediction result of the future order quantity.

[0162] To verify the effectiveness of the method (SE-ESG for short) of the application, comparative experiments were performed on a historical order data set of a certain charging pile operation management platform. The sample data set was divided into a training set, a validation set and a test set according to a ratio of 6:2:2. The evaluation indexes were root relative squared error (RSE), relative absolute error (RAE) and Pearson correlation coefficient (CORR).

[0163] The performance of each model on the task of predicting future order quantity is shown in Tables 3 and 4.

[0164] Table 3 shows experimental results for a prediction length of 3h / 6h / 9h:

[0165]

[0166] In Tables 3 and 4, the following representative baseline models were selected for comparison: TSMixer[KDD2023] is a kind of time series prediction model based on MLP; MSGNet[AAAI2024] is a kind of multi-scale graph neural network; Autoformer[NeurIPS2021] is a kind of Transformer variant based on autocorrelation mechanism; DLinear[AAAI2023] is a kind of simple linear decomposition model; and ESG[KDD2022] is a kind of evolutionary scale adaptive graph network.

[0167] Table 4 shows experimental results for a prediction length of 12h / 24h:

[0168]

[0169] From Table 3 and Table 4, it can be seen that the accuracy indicators of the SE-ESG model of the application are mostly better than those of the conventional time series prediction model, the RSE of the SE-ESG model is reduced by more than about 60% at each time step compared with TSMixer and MSGNet, the RSE of the SE-ESG model is reduced by more than about 35% at each time step compared with Autoformer and DLinear, and the training result of the RAE at each time step is reduced more. In terms of the indicator CORR, the SE-ESG model of the application is much higher than each conventional time series prediction model, except that it is slightly lower than the ESG model at some time steps, the correlation coefficient of the SE-ESG is more than twice the CORR of the remaining conventional models, proving that the model has good prediction performance at each time step.

[0170] Compared with other baseline models, the charging pile order data set achieves the best performance on the SE-ESG model, that is, the SE-ESG realizes the reduction of prediction error while maintaining high correlation, realizes the unification of prediction accuracy and inference performance, and proves the feasibility of the model in the charging pile order quantity prediction field.

[0171] In order to further understand the influence of each key component in the SE-ESG model of the application on the final performance, a series of ablation experiments are set to verify. Ablation experiment is a scientific research method mainly used to determine the degree of influence of a condition or parameter on the result, which is similar to the "control variable method". By controlling conditions or parameters one by one to observe the change of the result, the influence of them on the result is determined.

[0172] The key components of the SE-ESG model of the application mainly include SE-Net, RevIN and position encoding. These three parts have important effects on the accuracy of the model prediction result and the stability of the model training process. Therefore, ablation experiments are designed for these three key components.

[0173] The training results of each ablation experiment are shown in Tables 5 to 7.

[0174] Table 5 is a comparison of the training results of the SE-Net ablation experiment:

[0175]

[0176] The ablation experiment of SE-Net: SE-Net aims to better capture time-dependent relationships of different scales and reduce the risk of model overfitting, thereby improving the model training effect. In this embodiment, a comparison experiment is designed: on the charging pile order data set, the model with and without embedding the SE-Net module is trained respectively, and then the training results are evaluated and analyzed to determine the specific influence of SE-Net on the model performance.

[0177] From Table 5, it can be seen that the error values of the model all rise significantly without SENet. Especially, the change of RSE is most obvious, and the two evaluation indexes mainly measure the error comparison between the predicted value and the true value. Therefore, SENet plays an important role in enabling the model to better capture the dependency between time series and obtain a predicted value closer to the true value. The model with added SENet will have more accurate prediction results.

[0178] Ablation experiment of position encoding: Position encoding mainly adds time sequence information to the input data, so that the model can better capture data features, and thus the model can have good performance when facing various different data, thereby improving the generalization ability of the model. In this embodiment, a comparative experiment is designed: on the charging pile order data set, the model with embedded position encoding and the model without embedded position encoding are trained respectively, and then the training results are evaluated and analyzed to determine the specific influence of position encoding on the performance of the model.

[0179] From Table 6, it can be seen that the error values of the model all rise after removing the position encoding. Among them, the change of RSE is more obvious, and in the comparative model after removing the position encoding, the RSE increases by about 0.8% on average. In addition, the correlation coefficient also shows a downward trend, that is, after removing the position encoding, the correlation between the prediction results of the model and the true value will decrease. It can be seen that position encoding plays an important role in enabling the model to better capture the spatiotemporal relationship features in the data and thus master its development trend, and it is beneficial to improving the prediction correlation and prediction accuracy of the model.

[0180] Table 6 is a comparison of the results of the position encoding ablation experiment:

[0181]

[0182] Ablation experiment of RevIN: RevIN aims to normalize and denormalize the input and output data respectively, so as to improve the stationarity of the training data and avoid the influence of some abnormal discrete data points on the training process of the model. In this embodiment, a comparative experiment is designed: on the charging pile order data set, the model with embedded RevIN and the model without embedded RevIN are trained respectively, and then the training results are evaluated and analyzed to determine the specific influence of RevIN on the performance of the model.

[0183] As can be seen from Table 7, after removing RevIN, the partial precision indicators of the model have a downward trend. As can be known from the ablation experiments on SENet and position encoding, SENet and position encoding have a greater contribution to the model in improving the prediction correlation and reducing the prediction error. As for the ablation experiment of RevIN, it can be found that the RSE indicators of the model change more obviously after removing and adding RevIN, and the RAE of the complete model added with RevIN is significantly lower than that of the comparative model without adding RevIN. Therefore, after adding RevIN, the model effectively smooths the input data, avoids the influence of abnormal discrete data on the model training process, and thus improves the prediction accuracy of the model.

[0184] Table 7 is a comparison of the training results of the RevIN ablation experiment:

[0185]

[0186] In order to verify the generalization of the model of the present application, two public data sets are selected for comparison experiments with the original ESG model. The results show that the model of the present application still maintains excellent performance on heterogeneous data.

[0187] The disclosed data set of this embodiment is Traffic (traffic flow information) and Wind (wind energy information), and a series of comparative experiments are carried out on the time steps of 3h, 6h, 12h and 24h, and the training results corresponding to each time step are collected, as shown in Tables 8 and 9.

[0188] Table 8 is the comparative experiment results on the public data set Traffic:

[0189]

[0190] Table 9 is the comparative experiment results on the public data set Wind:

[0191]

[0192] As can be seen from Tables 8 and 9, on the public data sets Traffic and Wind, the SE-ESG of the present application still has good performance effect, and even in some time steps, the effect is overall superior to that of the original ESG model. The excellent prediction performance of the SE-ESG model on the two public data sets is verified by the evaluation indicators listed in Tables 8 and 9, which indirectly proves the good generalization ability of the model.

[0193] Through analysis of these experimental results, it is confirmed that the model of the application not only has excellent prediction performance in the field of charging pile order quantity prediction, but also has strong generalization ability when facing different spatio-temporal sequence data (such as Traffic and Wind), and can cope with the test of complex data conditions under various conditions.

[0194] Embodiment two

[0195] The embodiment of the application provides a charging pile order quantity prediction system based on an evolution graph and an attention mechanism, the prediction system comprising a memory, a processor and a computer program or instructions stored on the memory, and the processor executes the computer program or instructions to realize the charging pile order quantity prediction method in the embodiment one of the application.

[0196] Although not shown, the prediction system comprises a processor, which can perform various appropriate operations and processes according to programs and / or data stored in a read-only memory (ROM) or programs and / or data loaded from a storage part into a random access memory (RAM). The processor can be a multi-core processor, or can comprise a plurality of processors. In some embodiments, the processor can comprise a general-purpose main processor and one or more special-purpose coprocessors, such as a central processing unit, a graphics processing unit (GPU), a neural network processing unit (NPU), a digital signal processor (DSP), etc. In the RAM, various programs and data required for system operation are also stored. The processor, the ROM and the RAM are connected to each other through a bus. An input / output (I / O) interface is also connected to the bus.

[0197] The above processor and memory are used together to execute programs / instructions stored in the memory, and the programs / instructions can realize the methods, steps or functions described in the above embodiments when executed by a computer.

[0198] The above disclosure is only a specific embodiment of the application, but the protection scope of the application is not limited thereto, and any person skilled in the art can easily think of changes or modifications within the technical scope disclosed by the application, which should be covered within the protection scope of the application.

Claims

1. An evolution graph and attention mechanism based charging pile order quantity prediction method, characterized in that, The prediction method comprises: performing sliding window processing on a space-time sequence matrix composed of order quantities of each charging site to generate a sample pair composed of a historical window sequence and a future window sequence, so as to construct a sample data set; performing instance normalization, time sequence information injection and channel expansion on the historical window sequence to obtain a high-dimensional feature map; constructing a prediction model comprising a channel attention module, a multi-scale time feature extraction module, an evolutionary graph module and an output layer; and performing supervised training on the prediction model by using the high-dimensional feature map to obtain a trained prediction model; The channel attention module is configured to perform adaptive channel weighting on the high-dimensional feature map. The multi-scale time feature extraction module is configured to extract and fuse multi-scale time features from the weighted feature map. The evolutionary graph module comprises: a node extraction unit configured to extract static node features from the space-time sequence matrix and map the static node features to a dynamic graph embedding space; an evolutionary graph learning unit configured to take the output of the node extraction unit as an initial state, construct a spatial dependency relationship between charging sites based on the multi-scale time features, and generate a dynamic graph structure; a graph convolution unit configured to perform feature propagation and aggregation on the dynamic graph structure, realize deep fusion of space-time features, and obtain a space-time feature tensor; The output layer is configured to map the space-time feature tensor to a single-channel feature tensor, perform inverse normalization processing, and output a future order quantity prediction result of each charging site. The trained prediction model is used to predict the order quantity of the input charging pile, and a future order quantity is output.

2. The method of claim 1, wherein the method is based on an evolutionary graph and an attention mechanism. The instance normalization of the historical window sequence comprises: calculating the statistics of the time sequence corresponding to each charging site in the historical window sequence; performing normalization processing on the corresponding time sequence based on the statistics.

3. The method of claim 1, wherein the method is based on an evolutionary graph and an attention mechanism. The time sequence information injection is realized by adding a position encoding vector to the normalized sequence; the position encoding vector is generated according to the following formula: ； ； where PE denotes a position encoding vector; , denote the values of the position encoding vector at , 2i denotes an even dimension index in the position encoding vector, and 2i+1 denotes an odd dimension index in the position encoding vector; denotes the position index of the time step in the normalized sequence; denotes the total dimension of the position encoding vector.

4. The method of claim 1, wherein the method is based on an evolutionary graph and an attention mechanism. The channel attention module is SE-Net, and the SE-Net is used to perform adaptive channel weighting on the high-dimensional feature map, which comprises: performing global average pooling on each channel of the high-dimensional feature map in the space-time dimension to obtain a channel global scalar; performing feature extraction on the channel global scalar by using a multilayer perceptron with a bottleneck structure to obtain a channel attention weight vector; performing channel-by-channel multiplication on the high-dimensional feature map and the channel attention weight vector, and then performing residual connection with the high-dimensional feature map to obtain a weighted feature map.

5. The method of claim 1, wherein the method further comprises: The multi-scale time feature extraction module is a multi-branch dilated convolution module; the multi-branch dilated convolution module comprises a multi-scale time convolution module and a Transformer encoder, the multi-scale time convolution module adopts a parallel double-path gating structure, and both paths comprise a multi-scale dilated convolution submodule, the multi-scale dilated convolution submodule is implemented by a plurality of parallel branches, and each branch uses a convolution layer with different convolution kernel widths but the same dilation rate to extract time features of different scales; The multi-branch dilated convolution module is used to extract multi-scale time features from the weighted feature map, including: A multi-scale dilated convolution submodule is used to extract multi-scale local time features from the weighted feature map in parallel, and a gating mechanism is used to adaptively fuse the multi-scale local time features to obtain a preliminary fused feature representation; The preliminary fused feature representation is encoded by the Transformer encoder to capture global long-term dependencies, and the final multi-scale time features are obtained.

6. The method of claim 1, wherein the method further comprises: The node extraction unit includes: A CNN-based feature extractor is used to extract static node features from the spatiotemporal sequence matrix; A linear transformation layer is used to map the static node features to a dynamic graph embedding space.

7. The method of claim 1, wherein the method further comprises: The evolutionary graph learning unit is used to construct spatial dependencies between charging sites based on the multi-scale time features to generate a dynamic graph structure, including: The multi-scale time features are divided into multiple sub-segments according to the time interval, and each sub-segment is aggregated by an aggregation operator to obtain a segmented aggregated feature; A gated recurrent unit is used to update the node embedding state based on the segmented aggregated feature and the previous time hidden state; A double-channel MLP is used to generate a dynamic graph structure based on the updated node embedding.

8. The evolution graph and attention mechanism based charging pile order quantity prediction method according to any one of claims 1-7, characterized in that, The graph convolution unit is a hybrid graph convolution network that is used to propagate and aggregate features on the dynamic graph structure to achieve deep spatiotemporal feature fusion, including: A graph convolution operation is performed on the dynamic graph structure to aggregate feature information from multiple orders of neighborhood and fuse it with the node's own feature information to achieve deep spatiotemporal feature fusion.

9. An evolution graph and attention mechanism based charging pile order quantity prediction system, comprising a memory, a processor and a computer program or instructions stored on the memory, characterized in that, The processor executes the computer program or instructions to implement the charging pile order quantity prediction method according to any one of claims 1-8.

Images