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

Abstract

The invention discloses a charging pile order quantity prediction method and system based on an evolutionary graph and an attention mechanism, and the method comprises the steps: carrying out the sliding window processing of a space-time sequence matrix composed of the order quantity of each charging station, and generating 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; performing supervised training on the prediction model by using the high-dimensional feature map to obtain a trained prediction model; and predicting the number of input charging pile orders by using the trained prediction model, and outputting the number of future orders. According to the invention, the model prediction precision and generalization ability are improved.

Description

Technical Field

[0001] This invention belongs to the field of charging pile operation and management technology, and in particular relates to a method and system for predicting the number of charging pile orders based on evolutionary graphs and attention mechanisms. Background Technology

[0002] With the rapid development of the new energy vehicle industry, charging piles, as a key supporting facility, face increasingly severe challenges in terms of layout optimization and operational efficiency. The number of charging pile orders exhibits significant spatial and temporal differences, such as noticeable differences in demand between weekdays and holidays, and between urban centers and suburbs. An oversupply of charging piles will lead to resource idleness and cost waste; an undersupply will easily cause users to queue, reducing user experience and thus hindering the promotion and popularization of new energy vehicles.

[0003] In the field of charging pile order forecasting, traditional time series forecasting methods (such as ARIMA, exponential smoothing, and decision tree models) can capture the trend and periodic characteristics of data to some extent, but they perform poorly when dealing with charging pile order data with complex dynamic spatiotemporal relationships. These methods usually struggle to effectively model the spatial dependencies between different charging stations and cannot adapt to the dynamic spatial structure that changes over time, resulting in limited forecast accuracy.

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

[0005] (1) Insufficient utilization of spatial information: Most models only focus on feature extraction in the time dimension, ignoring the geographical connections and interactions between charging stations, which leads to prediction results deviating from the actual spatial distribution pattern;

[0006] (2) The model is not good at modeling long-term dependencies: When faced with order sequences with a long time span, the model is prone to gradient vanishing or overfitting, and it is difficult to capture deep dependencies across time periods.

[0007] (3) Poor adaptation of non-stationary sequences: The charging pile order data is affected by multiple factors such as holidays, weather, and user behavior, and exhibits obvious non-stationary characteristics. Existing methods have poor fitting ability and generalization performance for this type of data, and the correlation coefficient (CORR) between the prediction results and the actual values ​​is low, making it difficult to meet actual scheduling needs.

[0008] Therefore, there is an urgent need to develop a method for predicting the number of charging pile orders that can integrate spatiotemporal features, adaptively learn dynamic spatial dependencies, and has strong generalization ability, so as to support the intelligent scheduling and resource optimization of charging facilities. Summary of the Invention

[0009] The purpose of this invention is to provide a method and system for predicting the number of charging pile orders based on evolutionary graphs and attention mechanisms, so as to solve the problems of low prediction accuracy and poor generalization ability caused by ignoring dynamic spatial dependence and weak ability to model non-stationary time series in the prior art.

[0010] This invention solves the above-mentioned technical problems through the following technical solution: a method for predicting the number of charging pile orders based on evolutionary graphs and attention mechanisms, comprising:

[0011] A sliding window process is applied to the spatiotemporal sequence matrix consisting of the number of orders at each charging station to generate sample pairs consisting of historical window sequences and future window sequences, in order to construct a sample dataset.

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

[0013] A prediction model is constructed, comprising a channel attention module, a multi-scale temporal feature extraction module, an evolutionary graph module, and an output layer; the prediction model is then trained in a supervised manner using the high-dimensional feature map to obtain the trained prediction model.

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

[0015] The multi-scale temporal feature extraction module is used to extract and fuse multi-scale temporal features from the weighted feature map;

[0016] The evolutionary graph module includes:

[0017] A node extraction unit is used to extract static node features from the spatiotemporal sequence matrix and map them to a dynamic graph embedding space.

[0018] The evolutionary graph learning unit is used to construct the spatial dependencies between charging stations based on multi-scale temporal features, using the output of the node extraction unit as the initial state, in order to generate a dynamic graph structure.

[0019] Graph convolutional units are used to perform feature propagation and aggregation on the dynamic graph structure to achieve deep fusion of spatiotemporal features and obtain spatiotemporal feature tensors.

[0020] The output layer is used to map the spatiotemporal feature tensor into a single-channel feature tensor, perform inverse normalization processing, and output the prediction results of the future order quantity for each charging station.

[0021] The trained prediction model is used to predict the number of input charging pile orders and output the future order quantity.

[0022] This invention employs Reversible Instance Normalization (RevIN) technology (including instance normalization and denormalization) to effectively eliminate the non-stationarity in the number of charging pile orders, making the model training process more stable and converging faster, thereby directly improving the accuracy of the prediction results. Simultaneously, the channel attention module and the multi-scale temporal feature extraction module work together to ensure that the model can adaptively focus on key features and capture complex temporal dependencies from the short to the long term, further reducing prediction bias.

[0023] The core innovation of this invention lies in using evolutionary graph learning units to dynamically construct spatial relationship graphs between charging stations, overcoming the inherent defect that traditional static graph models cannot reflect the dynamic evolution of spatiotemporal relationships; and combining graph convolution units for feature fusion, enabling the model to accurately learn and quantify complex spatial patterns, fundamentally solving the technical problem that existing methods ignore or cannot effectively model spatial dependencies.

[0024] Because this invention specifically addresses the non-stationarity of data (RevIN) and deeply enhances and fuses spatiotemporal features, the model possesses stronger adaptability to data with different distributions (such as data from different regions and time periods). This means that the model can maintain reliable predictive performance even when encountering unseen data patterns, demonstrating excellent generalization ability and making it suitable for large-scale, complex charging pile network prediction tasks.

[0025] This invention restores the final single-channel feature tensor to its 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, enabling seamless integration with practical application systems such as intelligent scheduling of charging piles and optimized resource allocation, greatly enhancing the practical value of the method.

[0026] Furthermore, instance normalization is performed on the historical window sequence, including:

[0027] Calculate the statistics for the time series corresponding to each charging station in the historical window sequence;

[0028] The corresponding time series is normalized based on the aforementioned statistics.

[0029] Furthermore, the time sequence information injection is achieved 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; , These represent the position encoding vectors at... , The value at position 2i represents the even-numbered dimension index in the position encoding vector, and 2i+1 represents the odd-numbered dimension index in the position encoding vector. This represents the index of the time step within the normalized sequence. This represents the total dimension of the positional encoding vector.

[0033] Furthermore, the channel attention module is SENet, and the high-dimensional feature map is adaptively weighted using SENet, including:

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

[0035] The channel global scalar is used to extract features by a multilayer perceptron with a bottleneck structure to obtain the channel attention weight vector;

[0036] The high-dimensional feature map is multiplied channel by channel by channel attention weight vector, and then residually connected to the high-dimensional feature map to obtain the weighted feature map.

[0037] Furthermore, the multi-scale temporal feature extraction module is a multi-branch dilated convolution module; the multi-branch dilated convolution module includes a multi-scale temporal convolution module and a Transformer encoder. The multi-scale temporal convolution module adopts a parallel dual-path gating structure, and both paths contain multi-scale dilated convolution sub-modules. The multi-scale dilated convolution sub-modules are implemented through multiple parallel branches, and each branch uses convolutional layers with different kernel widths but the same dilation rate to extract temporal features at different scales.

[0038] A multi-branch dilated convolution module is used to extract multi-scale temporal features from the weighted feature map, including:

[0039] Multi-scale local temporal features are extracted in parallel from the weighted feature map using a multi-scale dilated convolutional submodule. The multi-scale local temporal features are then adaptively fused using a gating mechanism to obtain a preliminary fused feature representation.

[0040] The Transformer encoder is used to encode the preliminary fused feature representation to capture global long-term dependencies, resulting in the final multi-scale temporal features.

[0041] Furthermore, the node extraction unit includes:

[0042] A CNN-based feature extractor is used to extract static node features from the spatiotemporal sequence matrix.

[0043] A linear transformation layer is used to map the static node features to the dynamic graph embedding space.

[0044] Furthermore, the evolutionary graph learning unit is used to construct spatial dependencies between charging stations based on multi-scale temporal features, using the output of the node extraction unit as the initial state, to generate a dynamic graph structure, including:

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

[0046] The node embedding state is updated using a gated loop unit based on the segmented aggregation features and the hidden state of the previous time step.

[0047] A dynamic graph structure is generated based on the updated node embedding using a dual-channel MLP.

[0048] Furthermore, the graph convolutional unit is a hybrid graph convolutional network, which is used to perform feature propagation and aggregation on the dynamic graph structure to achieve deep fusion of spatiotemporal features, including:

[0049] Graph convolution operations are performed on the dynamic graph structure to achieve deep spatiotemporal feature fusion by aggregating feature information from multi-level neighborhoods and fusing it with the node's own feature information.

[0050] Based on the same concept, the present invention 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, wherein the processor executes the computer program or instructions to implement the charging pile order quantity prediction method as described above.

[0051] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0052] This invention effectively eliminates the non-stationarity in the number of charging pile orders by introducing reversible instance normalization technology, thereby improving the stability and convergence speed of model training. By combining channel attention mechanism and multi-scale time feature extraction, the model can simultaneously and accurately capture local details, periodic patterns and global long-term dependencies in the time series, thereby improving the model's prediction accuracy.

[0053] The evolutionary graph module of this invention can dynamically learn and update the correlation strength between charging stations based on time-series data, generating a dynamic graph structure that changes over time. This enables the model to adaptively capture complex dynamic spatial interactions (such as demand overflow during peak hours, competition effects between adjacent stations, etc.), fundamentally improving the model's ability to characterize real-world scenarios.

[0054] The collaborative work of the evolutionary graph learning unit and the graph convolutional unit enables deep spatiotemporal feature fusion. The evolutionary graph learning unit is responsible for perceiving the dynamic spatial structure, while the graph convolutional unit performs multi-level neighborhood information propagation and aggregation on the dynamic spatial structure. This architecture, which separates structure learning from feature learning yet tightly integrates them, allows the model to fully explore the intrinsic correlations in spatiotemporal data, thereby making more accurate predictions. Attached Figure Description

[0055] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only one embodiment of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1 This is a flowchart of the charging pile order quantity prediction method in an embodiment of the present invention;

[0057] Figure 2 This refers to the total number of orders placed at all charging stations within a given day in each time period, as described in this embodiment of the invention.

[0058] Figure 3 This refers to the total number of orders placed by all charging stations each day within a given week, as described in this embodiment of the invention.

[0059] Figure 4 This is a diagram of the SENet architecture in an embodiment of the present invention;

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

[0061] Figure 6 This is an architectural diagram of the multi-scale dilated convolutional submodule in an embodiment of the present invention;

[0062] Figure 7 This is a diagram of the CNN-based feature extractor architecture in an embodiment of the present invention;

[0063] Figure 8 This is an architecture diagram of the evolutionary graph learning unit in an embodiment of the present invention. Detailed Implementation

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

[0065] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0066] Example 1

[0067] Figure 1 A flowchart of the charging pile order quantity prediction method provided by the present invention is shown. Figure 1 As shown, the charging pile order quantity prediction method provided by the present invention includes the following steps:

[0068] Step 1: Construct the sample dataset.

[0069] Obtain the raw dataset containing order number, charging pile number, site information, order start time, and order end time from the charging pile operation management platform. The order start time and order end time refer to the start and end times when a user uses the charging pile to charge their electric vehicle. Partial information from the raw dataset is shown in Table 1.

[0070] Table 1 shows some information about the original dataset:

[0071] As can be seen from the original dataset in Table 1, its excessively complex information content is not suitable for model training. Therefore, the original dataset is preprocessed to obtain a dataset suitable for model training. The preprocessing in this invention involves: removing incomplete or erroneous noise data from the original dataset; obtaining the temporal distribution information of the original dataset; and counting the number of orders for each charging station at time intervals (e.g., 1 hour). Finally, the statistical results are summarized to obtain a spatiotemporal sequence matrix. The columns of the spatiotemporal sequence matrix represent the spatial numbers of each charging station, and the rows represent the number of charging pile orders for a certain charging station within a certain time period, as shown in Table 2. Each charging station in Table 2 contains multiple charging piles, and the number of orders for a certain charging station refers to the sum of the number of orders for all charging piles within that charging station.

[0072] To understand the spatiotemporal distribution and periodic trends of charging pile order volume, a spatiotemporal sequence matrix is ​​analyzed. To more clearly demonstrate the time series characteristics of the spatiotemporal sequence matrix, a time trend analysis is first performed. Figure 2 This displays the total number of orders placed at all charging stations within a given day, across different time periods. Figure 2 It can be seen that the peak time for orders each day is from 11:00 PM to 5:00 AM the next day, indicating that most new energy vehicle users' electricity needs are concentrated during their rest time at night. Figure 3 This shows the total number of orders placed at all charging stations each day during a given week. Figure 3It can be seen that the total number of orders from Monday to Sunday shows a slight downward trend with little difference between days, indicating that the number of orders changes relatively steadily within a week. However, the number of orders on weekdays is greater than the number of orders 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 time-dependent nature of the data distribution trend.

[0073] Analysis reveals that the spatiotemporal sequence matrix exhibits clear patterns and trends. By accurately capturing these time dependencies, the model can achieve more accurate predictions.

[0074] Table 2 shows a partial sequence of the spatiotemporal sequence matrix:

[0075] A sliding window process is applied to the spatiotemporal sequence matrix X based on the time step (or window width) to generate sample pairs consisting of historical window sequences and future window sequences. All sample pairs constitute the sample dataset. Let the number of charging stations be N, and the total time length be T. all If the time step is T, then the spatiotemporal sequence matrix X is a matrix with dimension N×T. all The historical window sequence is an N×T matrix, and the future window sequence is an N×T matrix. pred The matrix, T pred This indicates the prediction step size.

[0076] To facilitate the reading and processing of data from different dimensions, the sample dataset in CSV format was converted to H5 format. H5 (Hierarchical Data Format version 5), as a hierarchical file structure, performs well in tasks requiring efficient data compression and storage. When processing large-scale multivariate time series data, H5 format can significantly reduce storage space usage while accelerating data read and write speeds. This is crucial for model training and prediction processes that require frequent access to and processing of large amounts of data, improving the efficiency of the entire process. Furthermore, using H5 format allows models to easily read and process data from different dimensions without the need for complex data conversion.

[0077] Step 2: Perform instance normalization, time sequence information injection, and channel expansion on the historical window sequence to obtain a high-dimensional feature map.

[0078] To address the weakness in modeling non-stationary time series, this invention employs Reversible Instance Normalization (RevIN) to eliminate the non-stationarity of the input time series, thereby improving 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 model input and perform precise inverse normalization at the model output, thus ensuring the physical meaning of the prediction results.

[0079] Reversible instance normalization includes an instance normalization phase and an inverse normalization phase. Before inputting samples into the model, instance normalization is performed on the historical window sequence, including:

[0080] Step 2.1: Calculate the mean and standard deviation of the time series corresponding to each charging station in the historical window sequence;

[0081] Step 2.2: Normalize the corresponding time series based on the mean and standard deviation:

[0082] (1)

[0083] in, This represents the time series of the nth charging station within the b-th historical window sequence of a certain batch B; This represents the normalized time series; Representing time series The mean, , Representing time series The number of orders at time point t; Representing time series standard deviation , To represent a constant and prevent division by zero, this embodiment... Set to 1e -5 .

[0084] The model outputs a single-channel feature tensor. Then, the calculated mean and standard deviation are used for inverse normalization to restore the original dimensions, thus obtaining the predicted order quantity results for each site. The specific calculation formula is as follows:

[0085] (2)

[0086] In a specific embodiment of this invention, temporal sequence information injection is achieved by adding the positional encoding vector to the normalized sequence. The positional encoding vector aims to provide the model with positional information of the input data, enhancing the model's expressive power. This invention uses sine and cosine functions to generate positional codes. Since sine and cosine functions are periodic, combinations of sine and cosine functions of different frequencies can generate a unique encoding vector for each position. The design of the positional encoding ensures that it not only contains absolute position but also reflects relative positional information: that is, the relationship between any two positional encoding vectors can reflect their relative positions in the sequence. The positional encoding vector is generated according to the following formula:

[0087] (3)

[0088] (4)

[0089] Where PE represents the position encoding vector; , These represent the position encoding vectors at... , The value at position 2i represents the even-numbered dimension index in the position encoding vector, and 2i+1 represents the odd-numbered dimension index in the position encoding vector. This represents the index of the time step within the normalized sequence. This represents the total dimension of the positional encoding vector.

[0090] Using sine and cosine functions, codes with periodicity and relative positional information can be generated, effectively capturing positional relationships within a sequence and improving the model's predictive accuracy. Furthermore, different coding methods are applicable to different input data and tasks, enhancing the model's generalization ability.

[0091] To adapt to the subsequent multi-scale temporal feature extraction module based on 2D convolution, the sequence with injected temporal information is expanded and transformed in the channel dimension to obtain a high-dimensional feature map. The sequence with injected temporal information is treated as a special 2D feature map: first, a dimension is added using `torch.unsqueeze()`, expanding the sequence dimension to [B, 1, N, T], where B represents the batch size and 1 represents the initial single channel; then, an initial convolutional layer (`self.start_conv`) expands the number of channels from 1 to C (e.g., 16), resulting in the high-dimensional feature map. , This represents the set of all real-valued matrices of shape [B,C,N,T]. Structurally, the sequence is transformed into a two-dimensional image format with B batches, C channels, N pixel heights (number of charging stations), and T pixel widths (time steps), but its physical meaning remains a spatiotemporal sequence matrix of B samples, N charging stations, and T time steps.

[0092] Step 3: Construct a prediction model. The prediction model includes a channel attention module, a multi-scale temporal 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.

[0093] The channel attention module is used to adaptively weight the high-dimensional feature map U; the multi-scale temporal feature extraction module is used to extract and fuse multi-scale temporal features from the weighted feature map; the node extraction unit is used to extract static node features from the spatiotemporal sequence matrix X and map them to the dynamic graph embedding space; the evolutionary graph learning unit is used to construct the spatial dependency relationship between charging stations based on the multi-scale temporal features with the output of the node extraction unit as the initial state, so as to generate a dynamic graph structure; the graph convolution unit is used to perform feature propagation and aggregation on the dynamic graph structure to achieve deep fusion of spatiotemporal features and obtain spatiotemporal feature tensors; the output layer is used to map the spatiotemporal feature tensors to single-channel feature tensors and perform inverse normalization processing (as in formula (2)) to output the prediction results of the future order quantity of each charging station.

[0094] In a specific embodiment of the present invention, the channel attention module is SENet (e.g., Figure 4 As shown), SENet is used to adaptively weight the high-dimensional feature map U, including:

[0095] The squeeze operation performs global average pooling on each channel of the high-dimensional feature map in the space-time dimension (N×T) to obtain the channel's global scalar Z. The squeeze operation captures the global contextual information of each channel across the entire space-time domain. For the c-th channel, the output of the squeeze operation... for:

[0096] (5)

[0097] in, This 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 dimensions of the channel's global scalar Z are B×C×1×1, which change to B×C after adjusting the view.

[0098] Excitation operation: This involves learning the nonlinear interactions between channels using a multilayer perceptron (MLP) with a bottleneck structure, and generating a channel attention weight vector S. For example... Figure 4 As shown, the multilayer perceptron (MLP) with a bottleneck structure has two fully connected layers (FC). The first fully connected layer reduces the number of channels from C to C / a (where a is the reduction ratio) and uses the ReLU activation function. The second fully connected layer restores the number of channels to C and uses the Sigmoid activation function to generate a weight between 0 and 1 for each channel. This indicates the importance of the channel:

[0099] (6)

[0100] Where sigmoid represents the sigmoid activation function, ReLU represents the ReLU activation function, W1 and W2 are the weight matrices of the two fully connected layers, and b1 and b2 are the bias vectors of the two fully connected layers. Represents the first global scalar Z in the channel. A scalar of a channel, Indicates the first The weights of each channel, the channel attention weight vector The weights of all channels The structure is modified, and the dimensions change from B×C to B×C×1×1, which facilitates subsequent operations.

[0101] Reweighting and Residual Connections: For High-Dimensional Feature Maps With channel attention weight vector Channel-wise multiplication is performed to enhance important features and suppress minor features. Then, a shortcut connection is introduced to add the feature map after channel-wise multiplication to the high-dimensional feature map U to alleviate gradient vanishing and ensure the stability of model training. The weighted features are:

[0102] (7)

[0103] in, This represents the weighted feature, with dimensions B×C×N×T; This represents the weight of the c-th channel. This represents the feature of the c-th channel in the high-dimensional feature map U.

[0104] Through the channel attention mechanism, the model can adaptively focus on the feature channels that are most effective for the task of predicting the number of charging orders, thereby improving the feature representation capability. The addition of residual connections further ensures the effective training of the deep network and provides a direct propagation path for the gradient, allowing the gradient to propagate back more smoothly, thereby alleviating the gradient vanishing problem and helping the network to train and converge.

[0105] In a specific embodiment of the present invention, the multi-scale temporal feature extraction module is a multi-branch dilated convolution module. For example... Figure 5 As shown, the multi-branch dilated convolution module includes a multi-scale temporal convolution module (TConv) and a Transformer encoder. The multi-scale temporal convolution module adopts a parallel dual-path gating structure, and both paths contain multi-scale dilated convolution sub-modules. The multi-scale dilated convolution sub-modules are implemented through four parallel branches. Each branch uses convolutional layers with different kernel widths but the same dilation rate to extract temporal features at different scales, such as... Figure 6 As shown, the multi-branch dilated convolution module captures short-term and long-term temporal dependencies by using dilated convolution kernels of different scales in parallel, and combines a gating mechanism to finely regulate the information flow, providing a rich feature base for prediction.

[0106] like Figure 5 As shown, both paths of the multi-scale temporal convolution module consist of a multi-scale dilated convolution submodule (Dilated_Inception), serving as a filter and a gate, respectively. The core design of the multi-scale dilated convolution submodule is to explicitly capture multi-scale temporal context information using parallel convolution paths and dilated convolution techniques. For example... Figure 6 As shown, the multi-scale dilated convolutional submodule in this embodiment contains four branches, each employing an independent two-dimensional convolutional layer with a kernel size of (1, kernel). Here, 'in' represents the number of input channels and 'out' represents the number of output channels, where 'kernel' is a predefined value in kernel_set={2,3,6,7}. All four branches share the same dilation rate, dilation=(1,2). The kernel height is fixed at 1, ensuring that operations do not mix across different spatial charging stations, maintaining the feature independence of each charging station. The varying kernel widths directly determine the basic scale of the receptive field for each branch. The introduction of dilation increases the effective receptive field of each branch. For a convolutional layer with a kernel width of kernel and a dilation rate of dilation, its effective receptive field size is... It can be calculated using the following formula:

[0107] (8)

[0108] For example, when , hour, This enables the model to efficiently capture rich temporal features, ranging from local details to long-term trends. Since the convolutional kernel sizes of each branch differ, the length of their output sequences may vary slightly. A simultaneous pruning strategy is employed to uniformly prune the outputs of all branches to the shortest sequence length, ensuring consistency in subsequent concatenation operations. Finally, the outputs of each branch are concatenated along the channel dimension (torch.cat) to form features that integrate multi-scale temporal contextual information.

[0109] To dynamically adjust the multi-scale feature flow and enhance nonlinear expressive power, the outputs of two multi-scale dilated convolutional sub-modules (referred to as filter_conv and gate_conv, respectively) are fused. The output of filter_conv is transformed to the [-1, 1] interval using the tanh activation function, serving as the information content. The output of gate_conv is transformed to the [0, 1] interval using the sigmoid activation function, serving as the gating signal controlling the information flow at each scale. Finally, the two results are multiplied element-wise (Hadamard Product) to obtain the output of the multi-scale temporal convolutional module. :

[0110] (9)

[0111] Where sigmoid represents the sigmoid activation function. and These represent the outputs of filter_conv and gate_conv, respectively. The Hadamard product represents element-wise multiplication.

[0112] Gating mechanisms can adaptively assess the importance of features at different scales and perform weighted fusion accordingly, thereby significantly improving the model's ability to model complex temporal dynamics and its robustness.

[0113] Due to the features output by the multi-scale temporal convolution module Therefore, the features output by the multi-scale temporal convolution module cannot be directly input into the Transformer encoder, which only accepts 3D tensors. After serialization adaptation mapping, we get This is to facilitate input into the Transformer encoder. The specific implementation of the serialization adapter mapping is as follows:

[0114] (10)

[0115] (11)

[0116] (12)

[0117] Formula (10) represents rearranging the dimensions, placing the time dimension T at the front of the sequence axis; Formula (11) represents adjusting the tensor shape, merging the 3rd / 4th dimension (channel × node) into a single feature dimension, with each time step corresponding to a length of The vector; Formula (12) represents the linear mapping of the vector at each time step to the encoder embedding dimension d, ultimately obtaining , Let represent the set of all real matrices of shape [B,T,d]. This is the standard input shape of the transformer encoder (batch size × sequence length × embedding dimension). Here, `permute` represents reordering the dimensions (PyTorch's `permute` operation), rearranging the dimensions from [B,C,N,T] to [B,T,C,N]; `reshape` represents shape transformation (PyTorch's `reshape` or `view`), transforming the dimensions from [B,T,C,N] to [B,T,C×N]. Represents the projection matrix, Projecting 1D spatial-channel features onto d-dimensional space; d represents the bias vector, and d represents the encoder embedding dimension.

[0118] Will The input is fed into the transformer encoder, and the output of the transformer encoder is... The process is as follows:

[0119] (13)

[0120] (14)

[0121] (15)

[0122] in, For the number of attention heads, For each attention head, the embedding dimension is given by d, where d represents the encoder embedding dimension. These are the corresponding projections and output weights; , and These represent the query vector, key vector, and value vector, respectively. Indicates intermediate quantity; For the query / key / value of the h-th attention head, the superscript in formula (14) This represents the matrix transpose operation; To normalize the time dimension along the bond, For scaling items, To splice the pieces in the "head" dimension, For layer normalization, It is a linear forward propagation layer. This represents the residual connection between the sequence representation processed by the multi-head attention mechanism and the original input sequence, followed by the normalized intermediate representation obtained through layer normalization.

[0123] The output of the transformer encoder is then serialized and adapted. Rearrangement as multi-scale temporal features , which serves as the input in the evolutionary graph learning unit.

[0124] By cascading and cooperating with a multi-scale temporal convolution module and a Transformer encoder, this invention can simultaneously capture rich features from local details to global trends in time series and model extremely long-range dependencies, significantly improving the predictive ability for complex spatiotemporal series data such as the number of charging pile orders.

[0125] In a specific embodiment of the present invention, the node extraction unit includes a CNN-based feature extractor and a linear transformation layer. The CNN-based feature extractor is used to extract static node features from the spatiotemporal sequence matrix X. Figure 7 As shown, the CNN-based feature extractor employs hierarchical convolutional feature extraction, comprising two one-dimensional convolutional layers (conv1d), three batch normalization layers (BatchNorm1d), and finally a fully connected layer (FC) to output a fixed-dimensional feature vector. When the input spatiotemporal sequence matrix X passes through the CNN-based feature extractor, features of different lengths are extracted and converted into a feature vector of uniform dimension, facilitating subsequent processing.

[0126] (16)

[0127] in, Let T represent the input spatiotemporal sequence matrix, where N represents the number of stations, and T represents the number of stations. all Indicates all time lengths; Two-layer one-dimensional convolution operation; This indicates the flattening operation, which rearranges a multidimensional feature tensor into a two-dimensional tensor. Here, represents the weight matrix of the fully connected layer in the feature extractor, ReLU represents the ReLU activation function, and BN represents batch normalization. The output static node features have a shape of [N,D] s Static node characteristics Used for subsequent learning of dynamic graph structures and convolution operations.

[0128] The linear transformation layer maps the static node features extracted by the feature extractor to the dynamic graph embedding space through linear transformation, resulting in... As the initial state of the Gated Recurrent Unit (GRU) in the evolutionary graph learning unit:

[0129] (17)

[0130] in, This is the weight matrix of the fully connected layer in the linear transformation layer. This is a static node feature.

[0131] The evolutionary graph learning unit adopts a GRU-based architecture, such as... Figure 8 As shown. In the evolutionary graph learning unit, the multi-scale temporal feature Q is divided into multiple sub-segments according to the time interval, and the t-th sub-segment... The time length is r, and each sub-segment is aggregated using the aggregation operator AGG to obtain segmented aggregated features, which are then used as input to GRU:

[0132] (18)

[0133] in, represents the segmentation aggregation feature of the t-th sub-segment; r represents the time length of the sub-segment; This represents the feature of the t-th segment at the k-th time step; AGG is the average pooling over the time dimension, i.e., the aggregation operator.

[0134] The segmented aggregated features of each sub-segment, along with the hidden state from the previous time step, are input into the GRU to obtain new node embeddings. These new node embeddings are then input into a dual-channel MLP (MLP). e and MLP m The support matrix and mask matrix of the graph are generated separately. Then, the mask matrix is ​​activated by a sigmoid function. Finally, the support matrix and the activated mask matrix are multiplied by a Hadamard product to obtain the dynamic graph structure. That is, the adjacency matrix of a dynamic graph.

[0135] MLP e MLP is used to learn an "original graph structure candidate," which is equivalent to predicting which nodes should have edges and the strength of those edges. The output is a support matrix, reflecting the potential connection strength between nodes. m Used to output the mask matrix, it performs "gated filtering" on the support matrix, using the sigmoid activation function to determine which edges are preserved or weakened.

[0136] In a specific embodiment of the present invention, the graph convolutional unit is a hybrid graph convolutional network (MixProp). The MixProp network is used to perform feature propagation and aggregation on a dynamic graph structure, achieving deep fusion of spatiotemporal features to obtain a spatiotemporal feature tensor. Specifically, the MixProp network performs graph convolution operations on a dynamic graph structure, aggregating feature information from multi-order neighborhoods and fusing it with the node's own feature information to achieve deep fusion of spatiotemporal features. The deeply fused features corresponding to each sub-segment are concatenated in the time dimension to obtain the spatiotemporal feature tensor. The spatiotemporal feature tensor is mapped to a single-channel feature tensor and inversely normalized to output the predicted future order quantity for each charging station.

[0137] Using the learned dynamic graph adjacency matrix For the input fragment Perform graph convolution propagation:

[0138] (19)

[0139] in, The graph convolution outputs features (i.e., features after deep fusion) for the t-th segment. For the t-th sub-segment; is the dynamic graph adjacency matrix learned by GRU, MLPe, and MLPm, and MixProp is a hybrid graph convolutional network.

[0140] The graph convolution output features of all sub-segments are concatenated in the time dimension to obtain the final result:

[0141] (20)

[0142] in, The final output spatiotemporal feature tensor contains the convolution results of all sub-segments; The graph convolution output features for the t-th segment. To stitch along the time dimension, the original batch dimension, channel dimension, and node dimension remain unchanged; The number of channels is represented by the superscript m, which indicates the number of sub-segments.

[0143] Spatiotemporal feature tensor The input is fed into the channel mapping module. This module consists of a nonlinear activation function and a 1×1 convolutional layer stacked sequentially. First, the feature representation capability is enhanced through nonlinear transformation, and then pointwise convolution is used to gradually reduce the number of feature channels from the intermediate channel dimension C (e.g., 32) to the final required single-channel output. Specifically, the tensor shape H of the input to the channel mapping module is (B,C,N,T), and the tensor shape Y1 of the output result is (B,1,N,T), where B is the batch size, C is the number of channels, N represents the number of charging stations, and T represents the time step. Then, inverse normalization is performed using formula (2) to obtain the result. Finally, a truncation operation is performed on the time dimension, meaning only the first T is extracted. pred Time is the final prediction result:

[0144] (twenty one)

[0145] in, Indicates the prediction step size; This represents the feature tensor after channel mapping output and inverse normalization. This indicates the final prediction result.

[0146] Through a series of evolutionary and graph structure operations, the original time-series features are transformed into high-dimensional feature sequences rich in spatiotemporal context information. Driven by the evolutionary graph learning unit, the model can accurately capture dynamically changing spatiotemporal patterns and deeply explore the complex interactions between nodes. For example, when the order volume of surrounding charging stations increases, the order volume of the charging station itself may decrease due to demand diversion or competition effects; the data distribution in different time periods may also vary significantly. The evolutionary graph learning unit can dynamically identify such relationships and then effectively model these nonlinear spatiotemporal associations through graph convolution operations, significantly improving prediction accuracy.

[0147] The core functionality of the MixProp hybrid graph convolutional network lies in its ability to efficiently propagate and aggregate features between nodes based on dynamically generated graph structures, thereby enhancing the model's capacity to represent complex dependencies. Specifically, MixProp utilizes a multi-level neighborhood information aggregation mechanism to fuse features from neighboring nodes with different hop counts onto a dynamically changing graph structure, thus providing a more comprehensive characterization of the local and global context of nodes. Furthermore, MixProp employs residual or gating mechanisms to adaptively weight and fuse the original node features with the aggregation results, preserving the node's inherent characteristics while incorporating neighborhood information, thus improving the model's adaptability and stability to dynamic graph structures.

[0148] The evolutionary graph learning unit and MixProp work closely together and complement each other: the former, as a "graph structure generator," is responsible for dynamically inferring the dependencies between nodes from the data and outputting support and mask matrices to construct a time-varying graph topology; the latter, as a "feature processor," relies on the graph structure provided by the former to perform multi-level feature propagation and aggregation, achieving deep spatiotemporal feature fusion. Together, they form a complete "structure learning-feature learning" closed loop, effectively improving the model's ability to model dynamic spatiotemporal data and its predictive performance.

[0149] Step 4: Use high-dimensional feature maps to perform supervised training on the prediction model to obtain the trained prediction model.

[0150] In a specific embodiment of the present invention, supervised training of the prediction model is performed using high-dimensional feature maps, including:

[0151] Step 4.1: Perform adaptive channel weighting on the high-dimensional feature map;

[0152] Step 4.2: Extract and fuse multi-scale temporal features from the weighted feature map;

[0153] Step 4.3: Extract static node features from the spatiotemporal sequence matrix and map them to the dynamic graph embedding space;

[0154] Step 4.4: Using the output of Step 4.3 as the initial state, construct the spatial dependency relationship between charging stations based on multi-scale temporal features to generate a dynamic graph structure;

[0155] Step 4.5: Perform feature propagation and aggregation on the dynamic graph structure to achieve deep fusion of spatiotemporal features and obtain spatiotemporal feature tensors;

[0156] Step 4.6: Map the spatiotemporal feature tensor to a single-channel feature tensor, perform inverse normalization, and output the prediction results of the future order quantity for each charging station;

[0157] Step 4.7: Calculate the loss function based on the predicted future order quantity of each charging station and the corresponding future window sequence, and adjust the parameters of the prediction model (including weight parameters and biases) according to the loss function to achieve supervised training.

[0158] Step 5: Use the trained prediction model to predict the number of input charging pile orders and output the future order quantity.

[0159] The number of input charging pile orders is counted to obtain the spatiotemporal sequence matrix to be predicted; the spatiotemporal sequence matrix to be predicted is normalized by instance, injected with time sequence information and expanded by channel to obtain the high-dimensional feature map to be predicted; the high-dimensional feature map to be predicted is input into the trained prediction model to obtain the single-channel feature tensor; the single-channel feature tensor is denormalized to obtain the prediction result of the number of future orders.

[0160] To verify the effectiveness of the method of this invention (SE-ESG for short), a comparative experiment was conducted on the historical order dataset of a charging pile operation and management platform. The sample dataset was divided into training, validation, and test sets in a 6:2:2 ratio. The evaluation metrics used were root relative square error (RSE), relative absolute error (RAE), and Pearson correlation coefficient (CORR).

[0161] The performance of each model in predicting future order volume is shown in Tables 3 and 4.

[0162] Table 3 shows the experimental results for predicted lengths of 3h / 6h / 9h:

[0163] In Tables 3 and 4, the following representative baseline models were selected for comparison: TSMixer [KDD2023] is a time-series prediction model based on MLP; MSGNet [AAAI2024] is a multi-scale graph neural network; Autoformer [NeurIPS2021] is a Transformer variant based on autocorrelation mechanism; DLinear [AAAI2023] is a simple linear decomposition model; and ESG [KDD2022] is an evolutionary scale adaptive graph network.

[0164] Table 4 shows the experimental results for predicted lengths of 12h / 24h:

[0165] As shown in Tables 3 and 4, the SE-ESG model of this invention outperforms traditional time series prediction models in most accuracy metrics. Compared to TSMixer and MSGNet, its RSE is reduced by more than 60% at all time steps, and compared to Autoformer and DLinear, its RSE is reduced by more than 35% at all time steps. The reduction in RAE at all time steps is even greater. In terms of CORR, the SE-ESG model of this invention significantly outperforms all traditional time series prediction models. Except for slightly lower CORR at some time steps, the correlation coefficient of SE-ESG is more than twice that of other traditional models, demonstrating that the model has good predictive performance at all time steps.

[0166] Compared to other baseline models, the charging pile order dataset achieved the best performance on the SE-ESG model. That is, SE-ESG reduced the prediction error while maintaining high correlation, achieving a balance between prediction accuracy and inference efficiency, and proving the feasibility of the model in the field of charging pile order quantity prediction.

[0167] To gain a deeper understanding of the impact of each key component in the SE-ESG model of this invention on the final performance, a series of ablation experiments were conducted for verification. Ablation experiments, as a scientific research method, are mainly used to determine the degree of influence of a condition or parameter on the result. Similar to the "controlled variable method," they involve controlling conditions or parameters one by one to observe changes in the results, thereby determining their degree of influence.

[0168] The key components of the SE-ESG model of this invention are SENet, RevIN, and positional encoding. These three parts play an important role in the accuracy of the model's prediction results and the stability of the model's training process. Therefore, ablation experiments were designed for these three key components.

[0169] The comparison of the training results of each ablation experiment is shown in Tables 5 to 7.

[0170] Table 5 shows a comparison of the SENet ablation experiment training results:

[0171] SENet ablation experiments: SENet aims to better capture temporal dependencies at different scales, reduce the risk of model overfitting, and thus improve model training performance. This embodiment designs a comparative experiment: on a charging pile order dataset, models with and without the SENet module are trained respectively. The training results are then evaluated and analyzed to clarify the specific impact of SENet on model performance.

[0172] As shown in Table 5, removing SENet significantly increases all error values ​​of the model. The change in RSE is particularly pronounced, and these two evaluation metrics primarily measure the error between predicted and true values. Therefore, SENet plays a crucial role in enabling the model to better capture the dependencies between time series data and obtain predictions closer to the true values; models with SENet added will have more accurate prediction results.

[0173] Ablation Experiment of Location Encoding: Location encoding mainly adds temporal sequence information to the input data, enabling the model to better capture data features and thus perform well when faced with various types of data, improving the model's generalization ability. This embodiment designs a comparative experiment: on a charging pile order dataset, models with and without location encoding are trained respectively. The training results are then evaluated and analyzed to clarify the specific impact of location encoding on model performance.

[0174] Table 6 shows that all error values ​​of the model increased after removing location encoding. The change in RSE was particularly significant; in the comparison models where only location encoding was removed, the RSE increased by an average of approximately 0.8%. Furthermore, the correlation coefficient also showed a downward trend, indicating that the correlation between the model's predictions and the actual values ​​decreased after removing location encoding. This demonstrates that location encoding plays a crucial role in enabling the model to better capture the spatiotemporal characteristics of the data and thus grasp its development trends, which is beneficial for improving the model's predictive relevance and accuracy.

[0175] Table 6 shows a comparison of the experimental results of location-coded ablation:

[0176] RevIN ablation experiment: RevIN aims to normalize and denormalize the input and output data respectively, thereby improving the stability of the training data and avoiding the significant impact of certain abnormal discrete data points on the model training process. This embodiment designs a comparative experiment: on the charging pile order dataset, models with and without RevIN embedding are trained respectively, and then the training results are evaluated and analyzed to clarify the specific impact of RevIN on model performance.

[0177] Table 7 shows that some accuracy metrics of the model decrease after removing RevIN. Ablation experiments on SENet and positional encoding show that SENet and positional encoding contribute significantly to improving prediction relevance and reducing prediction error. For the RevIN ablation experiment, it can be observed that the RSE metric changes significantly after removing and adding RevIN; the RAE of the complete model with RevIN is significantly lower than that of the comparison model without RevIN. This indicates that adding RevIN effectively stabilizes the input data, avoiding the impact of abnormal discrete data on the model training process, thereby improving the model's prediction accuracy.

[0178] Table 7 shows a comparison of RevIN ablation experiment training results:

[0179] To verify the generalization ability of the model of this invention, this embodiment selected two public datasets and conducted a comparative experiment with the original ESG model. The results show that the model of this invention maintains excellent performance on heterogeneous data.

[0180] The publicly available datasets in this embodiment are Traffic (traffic flow information) and Wind (wind energy information). A series of comparative experiments were conducted at time steps of 3h, 6h, 12h, and 24h, and the training results corresponding to each time step were collected, as shown in Tables 8 and 9.

[0181] Table 8 shows the results of the comparative experiment on the public dataset Traffic:

[0182] Table 9 shows the results of the comparative experiment on the publicly available Wind dataset:

[0183] As shown in Tables 8 and 9, the SE-ESG model of this invention still performs well on the public datasets Traffic and Wind, and even outperforms the original ESG model at certain time steps. The evaluation metrics listed in Tables 8 and 9 also verify the excellent predictive performance of the SE-ESG model on these two public datasets, indirectly confirming the model's good generalization ability.

[0184] Analysis of these experimental results confirms that the model of this invention not only has excellent predictive performance in the field of charging pile order quantity prediction, but also has strong generalization ability when facing different spatiotemporal series data (such as Traffic and Wind), and can cope with the test of complex data conditions under various different circumstances.

[0185] Example 2

[0186] This invention provides a charging pile order quantity prediction system based on evolutionary graph and attention mechanism. The prediction system includes a memory, a processor, and a computer program or instructions stored in the memory. The processor executes the computer program or instructions to implement the charging pile order quantity prediction method in Embodiment 1 of this invention.

[0187] Although not shown, the prediction system includes a processor that performs various appropriate operations and processes based on programs and / or data stored in read-only memory (ROM) or loaded from a storage portion into random access memory (RAM). The processor can be a multi-core processor or may contain multiple processors. In some embodiments, the processor may include a general-purpose main processor and one or more specialized coprocessors, such as a central processing unit, graphics processing unit (GPU), neural network processor (NPU), digital signal processor (DSP), etc. Various programs and data required for system operation are also stored in RAM. The processor, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.

[0188] The processor and memory described above are used together to execute programs / instructions stored in the memory. When the program / instructions are executed by the computer, they can implement the methods, steps, or functions described in the above embodiments.

[0189] The above description only discloses specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or modifications that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting the number of charging pile orders based on evolutionary graphs and attention mechanisms, characterized in that, The prediction method includes: A sliding window process is applied to the spatiotemporal sequence matrix consisting of the number of orders at each charging station to generate sample pairs consisting of historical window sequences and future window sequences, in order to construct a sample dataset. The historical window sequence is subjected to instance normalization, time sequence information injection, and channel expansion to obtain a high-dimensional feature map; A prediction model is constructed, comprising a channel attention module, a multi-scale temporal feature extraction module, an evolutionary graph module, and an output layer; the prediction model is then trained in a supervised manner using the high-dimensional feature map to obtain the trained prediction model. The channel attention module is used to perform adaptive channel weighting on the high-dimensional feature map. The multi-scale temporal feature extraction module is used to extract and fuse multi-scale temporal features from the weighted feature map; The evolutionary graph module includes: A node extraction unit is used to extract static node features from the spatiotemporal sequence matrix and map them to a dynamic graph embedding space. The evolutionary graph learning unit is used to construct the spatial dependencies between charging stations based on multi-scale temporal features, using the output of the node extraction unit as the initial state, in order to generate a dynamic graph structure. Graph convolutional units are used to perform feature propagation and aggregation on the dynamic graph structure to achieve deep fusion of spatiotemporal features and obtain spatiotemporal feature tensors. The output layer is used to map the spatiotemporal feature tensor into a single-channel feature tensor, perform inverse normalization processing, and output the future order quantity prediction results for each charging station. The trained prediction model is used to predict the number of input charging pile orders and output the future order quantity.

2. The method for predicting the number of charging pile orders based on evolutionary graphs and attention mechanisms according to claim 1, characterized in that, Instance normalization of the historical window sequence includes: Calculate the statistics for the time series corresponding to each charging station in the historical window sequence; The corresponding time series is normalized based on the aforementioned statistics.

3. The method for predicting the number of charging pile orders based on evolutionary graphs and attention mechanisms according to claim 1, characterized in that, The time sequence information injection is achieved by adding the position encoding vector to the normalized sequence; the position encoding vector is generated according to the following formula: ； ； Where PE represents the position encoding vector; , These represent the position encoding vectors at... , The value at position 2i represents the even-numbered dimension index in the position encoding vector, and 2i+1 represents the odd-numbered dimension index in the position encoding vector. This represents the index of the time step within the normalized sequence. This represents the total dimension of the positional encoding vector.

4. The method for predicting the number of charging pile orders based on evolutionary graphs and attention mechanisms according to claim 1, characterized in that, The channel attention module is SENet, and the high-dimensional feature map is adaptively weighted using SENet, including: Global average pooling is performed on each channel of the high-dimensional feature map in the space-time dimension to obtain a channel global scalar; The channel global scalar is used to extract features by a multilayer perceptron with a bottleneck structure to obtain the channel attention weight vector; The high-dimensional feature map is multiplied channel by channel by channel attention weight vector, and then residually connected to the high-dimensional feature map to obtain the weighted feature map.

5. The method for predicting the number of charging pile orders based on evolutionary graphs and attention mechanisms according to claim 1, characterized in that, The multi-scale temporal feature extraction module is a multi-branch dilated convolution module; the multi-branch dilated convolution module includes a multi-scale temporal convolution module and a Transformer encoder. The multi-scale temporal convolution module adopts a parallel dual-path gating structure, and both paths contain multi-scale dilated convolution sub-modules. The multi-scale dilated convolution sub-modules are implemented through multiple parallel branches. Each branch uses convolutional layers with different kernel widths but the same dilation rate to extract temporal features at different scales. A multi-branch dilated convolution module is used to extract multi-scale temporal features from the weighted feature map, including: Multi-scale local temporal features are extracted in parallel from the weighted feature map using a multi-scale dilated convolutional submodule. The multi-scale local temporal features are then adaptively fused using a gating mechanism to obtain a preliminary fused feature representation. The Transformer encoder is used to encode the initially fused feature representation to capture global long-term dependencies, resulting in the final multi-scale temporal features.

6. The method for predicting the number of charging pile orders based on evolutionary graphs and attention mechanisms according to claim 1, characterized in that, 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 the dynamic graph embedding space.

7. The method for predicting the number of charging pile orders based on evolutionary graphs and attention mechanisms according to claim 1, characterized in that, The evolutionary graph learning unit is used to construct spatial dependencies between charging stations based on multi-scale temporal features, using the output of the node extraction unit as the initial state, to generate a dynamic graph structure, including: The multi-scale temporal features are divided into multiple sub-segments according to time intervals, and each sub-segment is aggregated using an aggregation operator to obtain segmented aggregated features; The node embedding state is updated using a gated loop unit based on the segmented aggregation features and the hidden state of the previous time step. A dynamic graph structure is generated based on the updated node embedding using a dual-channel MLP.

8. The method for predicting the number of charging pile orders based on evolutionary graphs and attention mechanisms according to any one of claims 1 to 7, characterized in that, The graph convolutional unit is a hybrid graph convolutional network, which is used to perform feature propagation and aggregation on the dynamic graph structure to achieve deep fusion of spatiotemporal features, including: Graph convolution operations are performed on the dynamic graph structure to achieve deep spatiotemporal feature fusion by aggregating feature information from multi-level neighborhoods and fusing it with the node's own feature information.

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

Images