Charging pile occupation information prediction method and system

Abstract

The invention discloses a charging pile occupation information prediction method and system, and relates to the technical field of electric vehicles, and the method comprises the steps: obtaining the attribute data, historical operation data and an adjacent matrix of a charging station in a target area; the attribute data comprises a time attribute, a geographic attribute, a weather attribute and a site attribute; based on historical occupation information in the historical operation data, screening out key attributes in the attribute data; the key attribute is one or more of a time attribute, a geographic attribute, a weather attribute and a site attribute; updating the adjacency matrix based on the key attribute to obtain a new adjacency matrix; and based on the historical operation data, the key attributes and the new adjacency matrix, utilizing a pre-trained prediction model to obtain a prediction result of the charging pile occupation information. According to the method provided by the invention, accurate prediction of charging pile occupation information can be realized, and data support is provided for optimizing power grid load scheduling and charging pile layout management.

Description

Technical Field

[0001] The present invention relates to the technical field of electric vehicles, and in particular to a method and system for predicting charging pile occupancy information. Background Art

[0002] Electric vehicles (EVs) have gained widespread popularity due to their clean and efficient nature. However, disordered EV charging can impact power system stability. The randomness and uncertainty of EV charging behavior significantly complicates grid load scheduling, especially when charging demand overlaps with peak grid loads, significantly increasing peak loads and increasing operational pressure. Charging stations are the primary gateway for EVs to connect to the grid. Predicting real-time changes in charging pile occupancy at charging stations can effectively mitigate grid risks.

[0003] In addition, from the perspective of the matching of supply and demand of charging piles, the current distribution of charging piles is not balanced with user needs. The utilization rate of charging piles in some areas is low, while queues may occur in some hot areas.

[0004] Therefore, in order to ensure the balance of the power grid and avoid the risk of peak overload, as well as to achieve the rational allocation of charging pile resources, it is particularly important to accurately predict the occupancy information of charging piles. Summary of the Invention

[0005] To solve the problems existing in the related art, the present invention provides a charging pile occupancy information prediction method and system, which can achieve accurate prediction of charging pile occupancy information and provide data support for optimizing power grid load scheduling and charging pile layout management.

[0006] To achieve the above object, the present invention provides a method for predicting charging pile occupancy information, comprising the following steps: Obtaining attribute data, historical operating data, and an adjacency matrix for charging stations in a target area; the attribute data includes time attributes, geographic attributes, weather attributes, and station attributes; and the adjacency matrix is used to characterize the spatial associations between charging stations. Based on the historical occupancy information in the historical operation data, key attributes in the attribute data are screened out; the key attributes are one or more of the time attribute, geographic attribute, weather attribute, and site attribute; Updating the adjacency matrix based on the key attributes to obtain a new adjacency matrix; Based on the historical operation data, key attributes and the new adjacency matrix, a pre-trained prediction model is used to obtain a prediction result of charging pile occupancy information.

[0007] Optionally, updating the adjacency matrix based on the key attribute to obtain a new adjacency matrix includes: Perform embedding learning on the regional topological relationships in the adjacency matrix to generate node embedding vectors; Construct node attribute vectors based on the key attributes; Fuse the node embedding vectors and the node attribute vectors to obtain node feature vectors; Calculate the similarity between every two node feature vectors to obtain a similarity matrix; Multiply the similarity matrix and the adjacency matrix to obtain a new adjacency matrix.

[0008] Optionally, the prediction model includes: Residual Time-Gated Convolution (Res-GTC) network, Lightweight Graph Convolutional Attention (Sim-GCA) network, Long Short-Term Memory (LSTM) network, and Time-Aware Attention (TPA) network; based on the historical operation data, key attributes, and new adjacency matrix, use the pre-trained prediction model to obtain the prediction result of the charging pile occupancy information, including: Use the historical operation data and key attributes as time-series input data; Extract the time features in the time-series input data through the Res-GTC network to obtain high-order time features; Based on the node features of the high-order time features, use the Sim-GCA network to aggregate the multi-hop spatial information of the new adjacency matrix to generate high-order spatial features; Fuse the high-order time features and high-order spatial features to obtain spatio-temporal fusion features; Use the LSTM network to capture the long-term dependencies in the spatio-temporal fusion features to generate a hidden state sequence; Use the TPA network to perform dynamic attention weight assignment and feature aggregation on the hidden state sequence, and then generate the prediction result of the charging pile occupancy information through linear transformation.

[0009] Optionally, when the Res-GTC network extracts the time features in the time-series input data to obtain high-order time features, it includes: Perform one-dimensional causal convolution on the time-series input data to obtain a causal convolution result; Dynamically adjust the time features of the causal convolution result through a gating mechanism to obtain the key time features at each time step; Stack the key time features at each time step through residual connections to obtain high-order time features.

[0010] Optionally, after obtaining the high-order spatial features, the method further includes: Perform three-dimensional attention weight assignment on the high-order spatial features through the energy function in the Sim-GCA network to obtain refined high-order spatial features; Then, fuse the high-order temporal features and high-order spatial features to obtain spatio-temporal fusion features, specifically: Fuse the high-order temporal features and the refined high-order spatial features to obtain spatio-temporal fusion features.

[0011] Optionally, fusing the high-order temporal features and high-order spatial features to obtain spatio-temporal fusion features includes: Align the dimensional structures of the high-order temporal features and the high-order spatial features through a reshaping operation; Concatenate the high-order temporal features and high-order spatial features with aligned dimensional structures to obtain spatio-temporal fusion features.

[0012] Optionally, based on the historical occupancy information in the historical operation data, filter out the key attributes in the attribute data, including: Use the Pearson correlation coefficient to perform a correlation analysis on the attribute data and the historical occupancy information in the historical operation data, and filter out the key attributes in the attribute data.

[0013] The present invention also provides a charging pile occupancy information prediction system, including: A data acquisition unit for acquiring the attribute data, historical operation data, and adjacency matrix of charging stations in a target area; the attribute data includes: time attribute, geographical attribute, weather attribute, and station attribute; the adjacency matrix is used to represent the spatial association relationship between each charging station; A data processing unit for: Based on the historical occupancy information in the historical operation data, filter out the key attributes in the attribute data; the key attributes are one or more of the time attribute, geographical attribute, weather attribute, and station attribute; Update the adjacency matrix based on the key attributes to obtain a new adjacency matrix; A prediction unit for obtaining a prediction result of the charging pile occupancy information by using a pre-trained prediction model based on the historical operation data, key attributes, and new adjacency matrix.

[0014] According to the specific embodiments provided by the present invention, the following technical effects of the present invention are disclosed: For the problem of predicting the occupancy information of charging piles in a complex environment with spatio-temporal feature correlations, the method and system for predicting the occupancy information of charging piles provided by the present invention fully consider the influence of factors such as weather conditions, charging prices, and the location of charging stations, integrate and analyze data in different dimensions, and screen out key attributes (such as weather, price) that affect the occupancy of charging piles; further update the adjacency matrix through the key attributes, so that the adjacency matrix originally representing static spatial relationships can dynamically reflect the spatial correlation changes under different conditions. For example, when the weather is bad, since users are more inclined to choose nearby charging stations, the correlation between some charging stations may increase; based on this, the present invention can accurately capture the periodic fluctuations of users' charging demands in the time dimension (such as morning and evening rush hours, seasonal changes) and the regional interaction relationships in the spatial dimension (such as load complementarity between commercial areas and residential areas), and realize the accurate prediction of the occupancy information of charging piles.

[0015] The present invention not only realizes the accurate prediction of the occupancy information of charging piles, but also provides effective support for optimizing the power grid load balance and the allocation of charging pile resources: on the one hand, it can actively avoid the risk of power grid overload according to the occupancy information of charging piles, prevent systematic safety hazards such as voltage fluctuations and line congestion caused by local load surges, and ensure the stable operation of the power system during peak hours; on the other hand, based on the occupancy information of charging piles, it can also predict the demand of hot spots, so as to optimize the layout management of charging piles, improve the overall utilization rate of charging piles and alleviate the regional supply-demand imbalance, providing technical support for building a flexible power system and a smart charging network. Brief Description of the Drawings

[0016] By describing the exemplary embodiments of the present invention in more detail in conjunction with the drawings, the above and other objects, features, and advantages of the present invention will become more obvious. Among them, in the exemplary embodiments of the present invention, the same reference numerals generally represent the same components.

[0017] Figure 1 It is a schematic flowchart of the method for predicting the occupancy information of charging piles shown in the embodiments of the present invention; Figure 2 It is a schematic diagram of the prediction model structure shown in the embodiments of the present invention; Figure 3 It is a schematic diagram of the Sim-GCA network structure shown in the embodiments of the present invention; Figure 4 It is a comparison chart of the prediction trends of the 139th charging area shown in the embodiments of the present invention; Figure 5 It is a comparison chart of the prediction trends of the 18th charging area shown in the embodiments of the present invention; Figure 6 It is a comparison chart of the prediction trends of the 78th charging area shown in the embodiments of the present invention; Figure 7 It is a comparison chart of the predicted trend of the charging area No. 113 shown in the embodiments of the present invention; Figure 8 It is a schematic diagram of the module structure of the charging pile occupancy information prediction system shown in the embodiments of the present invention. Detailed implementation manners

[0018] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0019] The purpose of the present invention is to provide a method and system for predicting charging pile occupancy information to accurately predict the occupancy information of charging piles and provide data support for optimizing the power grid load balance and charging pile resource allocation.

[0020] Please refer to Figure 1 , Figure 1 It is a schematic diagram of the method flow of the charging pile occupancy information prediction method. The above prediction method includes: Step 101: Obtain the attribute data, historical operation data, and adjacency matrix of the charging stations in the target area.

[0021] Among them, the adjacency matrix is used to characterize the spatial association relationship between each charging station. The attribute data includes: time attribute, geographical attribute, weather attribute, and station attribute.

[0022] When obtaining the relevant data of the charging stations in the target area, multi-dimensional information needs to be collected to support subsequent spatio-temporal feature analysis and prediction. The obtained attribute data of the charging stations includes four-dimensional information of time, space, environment, and station itself characteristics; specifically, the time attribute covers date, time point, and season information related to the periodicity of charging behavior, such as peak hours in the morning and evening, differences between weekdays and holidays, etc.; the geographical attribute includes spatial location characteristics such as whether the area where the charging station is located is a business district and the surrounding traffic density, which can reflect the impact of regional functions on charging demand; the weather attribute includes meteorological conditions such as temperature, weather type (sunny, rainy, foggy, etc.), humidity, and wind speed, which can be used to analyze the fluctuation effect of weather changes on charging behavior; the station attribute includes information such as charging station number, station capacity, and charging pricing, which can reflect the service ability of the station and the impact of charging price on user selection.

[0023] Historical operation data reflects the actual operating status of charging stations over a past period of time, including dynamic information such as the real-time occupancy rate, start and stop times, and charging power of charging piles; such data records the continuous changes in the usage of charging piles in the form of time series, and can reveal the regular and sudden characteristics of user charging behaviors, such as queuing phenomena during peak hours and idle states during low-utilization periods, providing historical law support for predicting future charging pile occupancy.

[0024] The adjacency matrix characterizes the spatial correlation relationship between each charging station; the adjacency matrix can store the adjacency status or distance information between each charging station in a two-dimensional structure. For example, if two charging stations are in the same community or the service radii overlap, they are marked as adjacent nodes; this spatial correlation can reflect the conduction effect of charging demand. For example, when a certain station is at full load, users may turn to neighboring stations, thus forming a demand coupling relationship between regions. The construction of the adjacency matrix is usually based on the geographical location coordinates of charging stations, and the connection strength between nodes is determined through spatial distance calculation or regional administrative division.

[0025] In the application, relevant data of charging stations in the target area can be obtained through data crawling technologies (such as Python crawlers), and preprocessing operations can be performed after obtaining the corresponding original data; among them, preprocessing can include operations such as abnormal data cleaning and normalization processing.

[0026] Step 102: Based on the historical occupancy information in the historical operation data, filter out the key attributes in the attribute data.

[0027] In the application, correlation analysis can be performed on the historical occupancy information and the attribute data to filter out the key attributes in the attribute data. When performing correlation analysis, methods such as Spearman rank correlation, chi-square test, mutual information method, and random forest feature importance assessment can be used; for example, the Pearson correlation coefficient can be used to perform correlation analysis on the attribute data and the historical occupancy information in the historical operation data to filter out the key attributes in the attribute data.

[0028] Among them, the key attributes are one or more of the time attribute, geographical attribute, weather attribute, and station attribute; it can be understood that as previously mentioned, the time attribute, geographical attribute, weather attribute, and station attribute each contain a variety of information. For example, the weather attribute specifically includes information such as temperature, rainfall, and wind speed. When the key attribute includes the weather attribute, specifically, it can be one or more of the information contained in the weather attribute (such as the key attribute includes temperature and rainfall); similarly, the key attribute can also specifically be one or more of the information contained in the time attribute, geographical attribute, or station attribute.

[0029] Step 103: Update the adjacency matrix based on the key attributes to obtain a new adjacency matrix.

[0030] When constructing the spatial correlation relationship among charging stations, the present invention breaks through the limitation of the existing adjacency matrix that only depends on geographical distance, and realizes the dynamic update of the adjacency matrix by integrating the key attributes of charging stations. The key attributes can reflect the potential correlation between charging stations. For example, there may be a correlation in demand fluctuations between two business district charging stations with similar functions but non-adjacent geographical locations.

[0031] In one embodiment, updating the adjacency matrix based on the key attributes to obtain a new adjacency matrix specifically includes the following steps: Perform embedding learning on the regional topological relationship in the adjacency matrix to generate node embedding vectors; Construct node attribute vectors based on the key attributes; Fuse the node embedding vectors and node attribute vectors to obtain node feature vectors; Calculate the similarity between pairwise node feature vectors to obtain a similarity matrix; Multiply the similarity matrix and the adjacency matrix to obtain a new adjacency matrix.

[0032] In application, updating the adjacency matrix mainly lies in constructing an attribute-enhanced similarity measure. First, the Node2Vec algorithm can be used to perform topological semantic mining on the original adjacency matrix to generate node embedding vectors containing multi-hop connection relationships, capturing the indirect spatial correlation between charging stations. Subsequently, the key attributes (such as whether it is located in a business district) are mapped to the same vector space through a fully connected network and linearly superimposed with the topological embedding vectors to form an enhanced representation vector (node feature vector) that combines functional attributes and spatial structures; further, the cosine similarity algorithm can be used to calculate the similarity between the node feature vectors of each charging station to generate a similarity matrix representing the multi-dimensional attribute correlation; finally, the similarity matrix can be combined with the original adjacency matrix through the Hadamard product (element-wise multiplication), so that the obtained new adjacency matrix not only retains the original spatial adjacency relationship but also superimposes the similarity weight of functional attributes.

[0033] Specifically, for the original adjacency matrix A, the Node2Vec algorithm can be used to perform embedding learning on the regional topological relationship and encode it into a node embedding matrix that preserves the graph structure information , and the row vectors of the node embedding matrix i represent the topological semantic information of the regional node ; Among them, and are hyperparameters that control the random walk strategy, respectively adjusting the breadth-first and depth-first search tendencies; R represents the set of real numbers, used to define the numerical type as real, Dis the preset embedding dimension, N Indicates the number of regional nodes.

[0034] In this way, the discrete graph structure can be transformed into a continuous vector space, explicitly enhancing the model's ability to model complex spatial relationships such as indirect connections and cross-region paths.

[0035] Furthermore, in order to integrate the semantics of functional types into the topological representation, regional attribute information is introduced. For example, the regional functional attribute in the key attributes, such as whether it is a Central Business District (CBD), can be constructed as a node attribute vector. ( Indicates the number of features contained in the node. If it only contains the single attribute CBD, then =1) and projected into the topological embedding space through a fully connected network. The process is expressed as: ; Where, and The first and second learnable parameters are output. is the attribute semantic vector; ReLU is the activation function.

[0036] By further fusing attribute semantics (node attribute vector) and topology embedding (node embedding vector) by element-by-element addition, the enhanced node feature vector is obtained as follows: ; Where, It also encodes regional functional attributes and topological structure information.

[0037] Then for each pair of nodes and The node feature vector of and Calculate the cosine similarity and get the similarity matrix , and multiply it with the adjacency matrix A to obtain the new feature adjacency matrix A new , the expression is as follows: ; .

[0038] The spatial information enhancement method proposed in the present invention, which integrates graph structure embedding and regional attribute features, can improve the model's ability to model complex spatial dependencies by reconstructing the adjacency matrix. The new adjacency matrix obtained after the update has dual characterization capabilities: maintaining the physical connection characteristics of the geographical proximity of charging stations at the base layer, and characterizing the possibility of cross-regional demand transmission through similarity weights at the enhancement layer. This update mechanism enables the establishment of virtual connection channels between charging stations that are far apart but have complementary functions. For example, when a charging station in a commercial area is fully loaded, the demand weight of the charging station in the adjacent residential area will automatically increase, thereby more accurately reflecting the behavioral patterns of users choosing alternative stations in actual operations.

[0039] It should be noted that in the process of updating the adjacency matrix, the corresponding node attribute vector can be constructed based on one or more information in the key attributes; for example, the key attributes specifically include two attribute information: charging pricing and whether it is located in a commercial area. When constructing the node attribute vector, it can be constructed only based on the attribute information of whether it is located in a commercial area.

[0040] Step 104: Based on the historical operation data, key attributes and the new adjacency matrix, a pre-trained prediction model is used to obtain a prediction result of the charging pile occupancy information.

[0041] Through multi-dimensional data fusion and spatiotemporal joint modeling, the present invention can accurately predict the occupancy information of charging piles; based on the obtained prediction results, high-load sites and idle areas can be known in advance, thereby supporting the grid side to adjust the power supply strategy in real time to balance the peak-valley differences, reduce the impact of disordered charging on grid stability, and avoid the risk of local overload; the operation side can optimize the layout of charging piles based on the prediction results, ensure the balance between supply and demand while improving the overall utilization rate of charging piles; in addition, the operation side can also implement differentiated pricing strategies based on the prediction results, dynamically guide user charging behavior, alleviate queuing pressure in hot areas and enhance the resource value of low-utilization sites.

[0042] See Figure 2 , Figure 2 Figure 2 is a schematic diagram of the prediction model structure. In one embodiment, the prediction model includes: a residual temporal gated convolution (Res-GTC) network, a lightweight graph convolutional attention (Sim-GCA) network, a long short-term memory (LSTM) network, and a time-aware attention (TPA) network.

[0043] Based on historical operational data, key attributes, and the new adjacency matrix, the above-mentioned prediction results of charging pile occupancy information are obtained using a pre-trained prediction model, including: Use historical operational data and key attributes as time series input data; The Res-GTC network is used to extract the time features in the time series input data and obtain high-order time series features; Based on the node features of high-order temporal features, use the Sim-GCA network to aggregate the multi-hop spatial information of the new adjacency matrix to generate high-order spatial features; Fuse the high-order temporal features and high-order spatial features to obtain spatio-temporal fusion features; Use the LSTM network to capture the long-term dependencies in the spatio-temporal fusion features and generate a hidden state sequence; Use the TPA network to perform dynamic attention weight assignment and feature aggregation on the hidden state sequence, and then generate the prediction result of the charging pile occupancy information through a linear transformation.

[0044] It should be noted that the information contained in the aforementioned key attributes is one or more. When using historical operation data and key attributes as temporal input data, one or more pieces of information in the key attributes can be selected; for example, the key attributes include two attribute information: charging pricing and whether it is located in a business district. Specifically, the charging pricing in the historical operation data and key attributes can be used as the temporal input data.

[0045] In the application, when training the prediction model, to achieve the generalization ability of the model, a time series segmentation strategy is adopted to divide the data set: the historical data is divided into a training set, a validation set, and a test set in chronological order, and the ratio is set to 6:2:2. To prevent time information leakage, ensure that the time windows of the validation set and the test set are strictly after the training set. For the model hyperparameters (such as learning rate, number of graph convolutional layers, number of attention heads), joint optimization is performed through grid search and the early stopping mechanism of the validation set; among them, grid search traverses the key parameter combinations within the preset parameter space; the early stopping mechanism is to terminate the training when the validation set loss does not decrease for 5 consecutive epochs and retain the optimal weights.

[0046] To evaluate the adaptability of the prediction model to dynamic spatio-temporal patterns, rolling time window cross-validation is adopted. For example, the data with a total duration of 30 days can be divided into multiple overlapping subsequences, and each subsequence contains a historical window (12 time intervals, 60 minutes) and a prediction window (15 - 60 minutes); training-validation sample pairs are generated by sliding the window to ensure that the patterns in all time periods are fully learned.

[0047] In one embodiment, the above Res-GTC network extracts the time features in the temporal input data to obtain high-order temporal features, including: Perform one-dimensional causal convolution on the temporal input data to obtain the causal convolution result; Dynamically adjust the time features of the causal convolution result through a gating mechanism to obtain the key time features of each time step; Stack the key time features of each time step through residual connection to obtain high-order temporal features.

[0048] In the application, in view of the complexity of the long-term and short-term dependencies intertwined in the timing characteristics of the charging demand, a residual time gated convolutional network Res-GTC is designed in the prediction model (G-STAN). Through the co-design of the gating mechanism and residual connection, the time dynamic characteristics in the charging demand are captured. Among them, the gated convolutional structure is composed of one-dimensional causal convolution and gated linear unit GLU. For the input of the th layer ( , , representing the initial time series features formed by splicing the occupancy rate and price, where is the time sliding window) the output is as follows: ; In the formula, and represent the main linear projection and the gating signal in GLU in turn; The symbol represents the element-wise Hadamard product. sigmoid represents the activation function, and the

[0049] function is used to adjust the input of the current state to identify the key structures and dynamic changes in the time series. In addition, the non-linear gating mechanism helps to capture global features by stacking time layers. Figure 3 , Figure 3 See for the schematic diagram of the Sim-GCA network structure; the present invention designs a lightweight graph convolutional attention network Sim-GCA in the prediction model to capture broader spatial dependencies; a multi-layer GCN graph convolutional structure is adopted, so that the node information of the new adjacency matrix can be propagated among multiple neighbors; at the same time, through a weighted fusion strategy, the output of the time convolutional layer (Res-GTC network) is connected with the graph convolution result, so that both the depth of the time series is retained and the model's understanding of the spatial structure is enhanced; the connection results are stacked to generate the final result of the GCN convolution ; ; ; In the formula, is the output of the th layer of GCN; the fusion connection between layers is expressed as ; when , and are the time series feature sequences (high-order time series features) output by the Res-GTC network; the parameter For controlling the degree of fusion, "stack" represents the stacking operation.

[0050] In one embodiment, after obtaining the high-order spatial features, the method further includes: Performing three-dimensional attention weight assignment on the high-order spatial features through the energy function in the Sim-GCA network to obtain refined high-order spatial features.

[0051] In applications, an attention component SimAM can also be designed in the Sim-GCA network. As a lightweight attention mechanism based on convolutional neural networks, compared with existing channel attention modules and spatial attention modules, it infers the three-dimensional (node, time, and feature dimensions) attention weights of each neuron in the feature map without increasing the original network parameters, thus significantly reducing the model complexity. The core of SimAM is an energy function that generates three-dimensional attention weights by mining the information richness of neurons, and these weights are used to refine the feature map and improve the accuracy of charging demand prediction. Among them, the energy function is defined as: ; In the formula, is the target neuron; is a constant; and are the mean and variance of the target neuron removed from the channel in sequence. The SimAM attention component further refines the input graph signal based on the concept of the energy function as: ; In the formula, X " represents the refined high-order spatial features, represents the energy matrix, aggregates all in the channel and spatial dimensions. By using the function to limit the value, SimAM can avoid excessive weight values while maintaining the relative importance of neurons, simplify the tuning of the model structure, and reduce the need for operator selection.

[0052] Furthermore, the present invention combines the LSTM network and the time pattern attention mechanism TPA network to capture the time dynamic characteristics in the charging demand. First, the high-order time features and high-order spatial features extracted by spatio-temporal convolution are processed and reconstructed to generate spatio-temporal fusion features suitable for subsequent input formats ; The spatio-temporal fusion features contain the historical time series information of all regions. To capture the independent time-dependent patterns of each region, can be split by region dimension intoN ( N is a positive integer) independent sequences , each single-region spatio-temporal sequence is input into the LSTM network for encoding to generate the corresponding hidden state .

[0053] In one embodiment, the above-mentioned fusion of high-order temporal features and high-order spatial features to obtain spatio-temporal fusion features includes: Align the dimensional structures of the high-order temporal features and high-order spatial features through a reshaping operation; Concatenate the high-order temporal features and high-order spatial features with aligned dimensional structures to obtain spatio-temporal fusion features.

[0054] In the application, the main role of the LSTM network is to capture the long-term dependencies of the charging demand. By dynamically updating the hidden state through its memory unit, it can remember long-term features such as morning and evening rush hours while ignoring irrelevant short-term noises. The features first pass through a long short-term memory layer, and the calculation process of the hidden state is as follows: ; ; ; ; ; ; In the formula, , , and respectively represent the input feature weights of the input gate, forget gate, output gate, and candidate memory unit; , , and respectively represent the hidden state weights of the input gate, forget gate, output gate, and candidate memory unit; , , and respectively represent the bias terms of the input gate, forget gate, output gate, and candidate memory unit; is the hidden state of the LSTM network at time or the initial hidden state at time 0; respectively represent the input, forget, and output gates of the long short-term memory layer; represents the candidate memory unit, c tRepresents the memory cell state; tanh represents the hyperbolic tangent activation function. Through this mechanism, a hidden state sequence reflecting the long-term trend can be extracted from the sequence data and used as the input for the subsequent TPA network.

[0055] The TPA network is mainly used to dynamically allocate the importance weights of the time dimension and highlight the influence of key time points (such as peak periods or sudden demands). Specifically, on the hidden state sequence deploy CNN filters of length to perform convolution on features, generating a convolutional feature matrix containing local features , where represents the convolution result of the th row vector of the hidden state sequence and the th filter; then, through a scoring function, the row vector is compared with the current hidden state and normalized to obtain the attention score for each row for weighted summation of the row vectors to obtain the aggregated time features. The expression is as follows: ; ; In the formula, represents the attention weight matrix, sigmoid is the activation function, represents the th row convolutional feature vector, represents the aggregated time feature vector; by adopting a weighted mechanism, the interference of non-critical time points can be effectively reduced, enabling the model to capture both macroscopic trends and microscopic changes simultaneously.

[0056] Finally, three linear transformation layers , and (time feature projection matrix, hidden state projection matrix, and joint projection matrix) act as fully connected layers to fuse the information output by the TPA network and output the final prediction result Y as: .

[0057] Furthermore, to verify the prediction accuracy of the prediction method provided by the present invention, taking the target area as Shenzhen urban area as an example, the historical data of 247 traffic cells publicly available within the scope of Shenzhen is used as the data set, and the information in the data set is shown in Table 1 below.

[0058] Table 1

[0059] These traffic zones cover a total of 18,061 charging piles, with a time span of 30 days and a data sampling frequency of 5 minutes, resulting in a total of 8,640 records. The features in the dataset include timestamp, charging price, adjacency matrix, regional distance, charging pile occupancy rate, and node attributes (such as the number of piles and functional types).

[0060] For this dataset, it can be divided into a training set, a validation set, and a test set in chronological order, with a ratio of 6:2:2; rolling window cross-validation is used to optimize the spatio-temporal feature extraction ability; the optimal hyperparameter combination is determined through grid search, and overfitting is prevented based on the early stopping mechanism. The prediction task covers multi-scale intervals from 15 to 60 minutes, and the features input into the model include historical occupancy rate, charging price, and adjacency matrix; it should be noted that the adjacency matrix input into the prediction model provided by the present invention is a new adjacency matrix.

[0061] When evaluating the accuracy of the prediction results, the root mean square error RMSE, mean absolute error MAE, and mean absolute percentage error MAPE are selected as evaluation indicators to reflect the accuracy of the model in predicting the occupancy number of charging piles in the charging station; the expressions for each indicator are: ; ; .

[0062] In the formula, k represents a certain regional node, n is the total number of regional nodes, represents the regional node k 's true value, represents the predicted value of the regional node k .

[0063] Refer to Table 2 below. Table 2 shows the average values of the comparative experiments of the prediction model G-STAN of the present invention and other existing prediction methods and prediction models on the same dataset. The root mean square error of the prediction model (G-STAN) provided by the present invention is 4.56%, the mean absolute error is 2.49%, and the mean absolute percentage error is 15.47%. Compared with mathematical prediction, machine learning methods, and existing prediction models, it can be found that the prediction accuracy of the prediction model adopted by the present invention is significantly higher.

[0064] Table 2

[0065] The existing prediction models participating in the comparative display in Table 2 are Vector Autoregression (VAR), K-Nearest Neighbor Algorithm (KNN), Graph Convolutional Network (GCN), Graph Attention Network (GAT), Long Short-Term Memory Network (LSTM), Graph Convolutional Long Short-Term Memory Network (GCN-LSTM), Spatio-Temporal Graph Convolutional Network (STGCN), Diffusion Convolutional Recurrent Neural Network (DCRNN), and Spatio-Temporal Residual Network (ST-ResNet).

[0066] As can be seen from Table 2, as a recurrent neural network, LSTM has better prediction accuracy than the existing statistical models VAR and KNN. Although GCN has advantages in processing graph data, it performs poorly in three metrics. While GAT is stronger than GCN in the flexibility of weight assignment, its overall effect is still not ideal. In contrast, models that combine recurrent networks and graph convolutional networks (such as GCN-LSTM, DCRNN, and STGCN) and the ST-ResNet model with a residual structure are significantly superior to the single LSTM model in prediction performance. This illustrates the importance of combining graph structure knowledge and time series patterns for improving model performance.

[0067] By observing the results of various time length prediction tasks, it can be seen that in the short-term prediction of 15 minutes, the RMSE index of ST-ResNet is slightly lower than that of the G-STAN model proposed in the present invention. This may be because the former's prediction in a short time is more dependent on historical data, while the latter pays more attention to the capture of spatio-temporal dynamic features, so its advantage has not been reflected in short-term prediction tasks. As the prediction time length increases, the RMSE, MAE, and MAPE indexes of each model increase to varying degrees. This indicates that longer time span prediction tasks pose higher requirements for the spatio-temporal capture ability of the model. In contrast, the G-STAN model shows superior performance in all indexes. Its average RMSE is 4.56%, MAE is 2.49%, and MAPE is 15.47%. Compared with the existing VAR and LSTM models, the G-STAN model performs excellently in prediction tasks of different time scales. Compared with the spatio-temporal prediction models DCRNN and ST-ResNet, G-STAN reduces the RMSE by 21.92% and 5.0% respectively, reduces the MAE by 36.96% and 4.79% respectively, and reduces the MAPE by 16.92% and 7.48% respectively. These significant improvements indicate that the prediction method and model provided by the present invention can effectively capture complex spatio-temporal dependence relationships and achieve accurate prediction of charging pile occupancy information.

[0068] Furthermore, reference can also be made to Figures 4 to 7 , where Figure 4 is the comparison chart of the prediction trends in charging area No. 139, Figure 5 is the comparison chart of the prediction trends in charging area No. 18, Figure 6It is a comparison chart of the predicted trends for charging area No. 78. Figure 7 It is a comparison chart of the predicted trends for charging area No. 113. Figures 4 to 7 It shows the specific comparison of the predicted trends of multiple charging stations. Specifically, the G-STAN model of the present invention is compared with the existing graph convolutional long short-term memory network GCN-LSTM, spatio-temporal graph convolutional network STGCN, and spatio-temporal residual network ST-ResNet. It is not difficult to see that the prediction results of the present invention are more in line with the actual true value Actual, and the prediction accuracy is significantly better than the existing prediction methods and models.

[0069] Corresponding to the foregoing method embodiment for implementing the application function, the present invention also provides a charging pile occupancy information prediction system and corresponding embodiments.

[0070] Please refer to Figure 8 , Figure 8 It is a schematic diagram of the module structure of the charging pile occupancy information prediction system. The prediction system includes: A data acquisition unit 81, configured to acquire the attribute data, historical operation data, and adjacency matrix of the charging stations in the target area; the attribute data includes: time attribute, geographical attribute, weather attribute, and station attribute; the adjacency matrix is used to represent the spatial association relationship between the charging stations; A data processing unit 82, configured to: Based on the historical occupancy information in the historical operation data, screen out the key attributes in the attribute data; the key attributes are one or more of the time attribute, geographical attribute, weather attribute, and station attribute; Update the adjacency matrix based on the key attributes to obtain a new adjacency matrix; A prediction unit 83, configured to obtain the prediction result of the charging pile occupancy information based on the historical operation data, key attributes, and new adjacency matrix by using a pre-trained prediction model.

[0071] In one embodiment, in terms of screening out the key attributes in the attribute data based on the historical occupancy information in the historical operation data, the above data processing unit 82 is specifically configured to: Perform a correlation analysis on the attribute data and the historical occupancy information in the historical operation data by using the Pearson correlation coefficient, and screen out the key attributes in the attribute data.

[0072] In one embodiment, in terms of updating the adjacency matrix based on the key attributes to obtain a new adjacency matrix, the above data processing unit 82 is specifically configured to: Perform embedding learning on the regional topological relationship in the adjacency matrix to generate node embedding vectors; Based on the key attributes, construct node attribute vectors; Fuse the node embedding vectors and node attribute vectors to obtain node feature vectors; Calculate the similarity between pairwise node feature vectors to obtain a similarity matrix; Multiply the similarity matrix and the adjacency matrix to obtain a new adjacency matrix.

[0073] In one embodiment, the prediction model includes: a Residual Time-Gated Convolution (Res-GTC) network, a Lightweight Graph Convolutional Attention (Sim-GCA) network, a Long Short-Term Memory (LSTM) network, and a Time-Aware Attention (TPA) network; in terms of obtaining the prediction result of the charging pile occupancy information by using the pre-trained prediction model based on the historical operation data, key attributes, and the new adjacency matrix, the above prediction unit 83 is specifically configured to: Use the historical operation data and key attributes as the sequential input data; Extract the time features in the sequential input data through the Res-GTC network to obtain high-order sequential features; Based on the node features of the high-order sequential features, use the Sim-GCA network to aggregate the multi-hop spatial information of the new adjacency matrix to generate high-order spatial features; Fuse the high-order sequential features and the high-order spatial features to obtain spatio-temporal fusion features; Use the LSTM network to capture the long-term dependencies in the spatio-temporal fusion features to generate a hidden state sequence; Use the TPA network to perform dynamic attention weight assignment and feature aggregation on the hidden state sequence, and then generate the prediction result of the charging pile occupancy information through a linear transformation.

[0074] In one embodiment, in terms of extracting the time features in the sequential input data through the Res-GTC network to obtain high-order sequential features, the above prediction unit 83 is specifically configured to: Perform one-dimensional causal convolution on the sequential input data to obtain a causal convolution result; Dynamically adjust the time features of the causal convolution result through a gating mechanism to obtain the key time features of each time step; Stack the key time features of each time step through a residual connection to obtain high-order sequential features.

[0075] In one embodiment, after obtaining the high-order spatial features, the above prediction unit 83 is further configured to: Perform three-dimensional attention weight assignment on the high-order spatial features through the energy function in the Sim-GCA network to obtain refined high-order spatial features; Then, fuse the high-order sequential features and the high-order spatial features to obtain spatio-temporal fusion features, specifically: Fuse the high-order sequential features and the refined high-order spatial features to obtain spatio-temporal fusion features.

[0076] In one embodiment, in terms of fusing high-order temporal features and high-order spatial features to obtain spatio-temporal fusion features, the above-mentioned prediction unit 83 is specifically configured to: Align the dimensional structures of the high-order temporal features and the high-order spatial features through a reshaping operation; Concatenate the high-order temporal features and the high-order spatial features with aligned dimensional structures to obtain spatio-temporal fusion features.

[0077] Regarding the prediction system in the above embodiment, the specific manners in which each unit performs operations have been described in detail in the embodiment related to the prediction method, and will not be elaborated herein. In addition, the prediction system and the above prediction method utilize the same prediction model (G-STAN). For the introduction of the prediction model, refer to the foregoing content and will not be repeated here.

[0078] The embodiments of the present invention have been described above. The above description is exemplary and not exhaustive, and is not limited to the disclosed embodiments. Many modifications and variations are obvious to those of ordinary skill in the art in the technical field without departing from the scope and spirit of the described embodiments. The selection of the terms used herein is intended to best explain the principles of the embodiments, the practical application, or the improvement of the technology in the market, or to enable other ordinary skill in the art in the technical field to understand the disclosed embodiments.

Claims

1. A method for predicting the occupancy information of a charging pile, characterized in that, Including: Obtain the attribute data, historical operation data, and adjacency matrix of charging stations in the target area; The attribute data includes: time attribute, geographical attribute, weather attribute, and station attribute; the adjacency matrix is used to characterize the spatial association relationship between charging stations; Based on the historical occupancy information in the historical operation data, filter out the key attributes in the attribute data; the key attributes are one or more of the time attribute, geographical attribute, weather attribute, and station attribute; Update the adjacency matrix based on the key attributes to obtain a new adjacency matrix; Based on the historical operation data, key attributes, and new adjacency matrix, use a pre-trained prediction model to obtain the prediction result of the charging pile occupancy information.

2. The method for predicting the occupancy information of a charging pile according to claim 1, wherein Updating the adjacency matrix based on the key attributes to obtain a new adjacency matrix includes: Perform embedding learning on the regional topological relationship in the adjacency matrix to generate node embedding vectors; Based on the key attributes, construct node attribute vectors; Fuse the node embedding vectors and the node attribute vectors to obtain node feature vectors; Calculate the similarity between every two node feature vectors to obtain a similarity matrix; Multiply the similarity matrix and the adjacency matrix to obtain a new adjacency matrix.

3. The method for predicting the occupancy information of a charging pile according to claim 1, wherein The prediction model includes: Residual Time-Gated Convolution Res-GTC network, Lightweight Graph Convolutional Attention Sim-GCA network, Long Short-Term Memory LSTM network, and Time-Aware Attention TPA network; based on the historical operation data, key attributes, and new adjacency matrix, using a pre-trained prediction model to obtain the prediction result of the charging pile occupancy information includes: Use the historical operation data and key attributes as time-series input data; Extract the time features in the time-series input data through the Res-GTC network to obtain high-order time-series features; Based on the node features of the high-order time-series features, use the Sim-GCA network to aggregate the multi-hop spatial information of the new adjacency matrix to generate high-order spatial features; Fuse the high-order time-series features and high-order spatial features to obtain spatio-temporal fusion features; Use the LSTM network to capture the long-term dependence relationship in the spatio-temporal fusion features to generate a hidden state sequence; Use the TPA network to perform dynamic attention weight allocation and feature aggregation on the hidden state sequence, and then generate the prediction result of the charging pile occupancy information through linear transformation.

4. The method for predicting charging pile occupancy information according to claim 3, wherein, The Res-GTC network extracts the time features in the time-series input data to obtain high-order time-series features, including: Perform one-dimensional causal convolution on the time-series input data to obtain a causal convolution result; Dynamically adjust the time features of the causal convolution result through a gating mechanism to obtain the key time features of each time step; Superimpose the key time features of each time step through residual connection to obtain high-order time-series features.

5. The method for predicting the occupancy information of a charging pile according to claim 3, wherein After obtaining the high-order spatial features, the method further includes: Perform three-dimensional attention weight allocation on the high-order spatial features through the energy function in the Sim-GCA network to obtain refined high-order spatial features; Then, fusing the high-order time-series features and high-order spatial features to obtain spatio-temporal fusion features, specifically: Fuse the high-order temporal features and the refined high-order spatial features to obtain spatio-temporal fusion features.

6. The method for predicting the occupancy information of a charging pile according to claim 3, wherein, Fusing the high-order temporal features and the high-order spatial features to obtain spatio-temporal fusion features, including: Align the dimensional structures of the high-order temporal features and the high-order spatial features through a reshaping operation; Concatenate the high-order temporal features and the high-order spatial features with aligned dimensional structures to obtain spatio-temporal fusion features.

7. The method for predicting charging pile occupancy information according to claim 1, wherein Based on the historical occupancy information in the historical operation data, screen out the key attributes in the attribute data, including: Use the Pearson correlation coefficient to perform a correlation analysis on the attribute data and the historical occupancy information in the historical operation data, and screen out the key attributes in the attribute data.

8. A charging pile occupancy information prediction system, characterized in that, Including: A data acquisition unit for acquiring the attribute data, historical operation data, and adjacency matrix of the charging stations in the target area; The attribute data includes: time attribute, geographical attribute, weather attribute, and station attribute; the adjacency matrix is used to represent the spatial association relationship between each charging station; A data processing unit for: Based on the historical occupancy information in the historical operation data, screen out the key attributes in the attribute data; the key attributes are one or more of the time attribute, geographical attribute, weather attribute, and station attribute; Update the adjacency matrix based on the key attributes to obtain a new adjacency matrix; A prediction unit for obtaining a prediction result of the charging pile occupancy information based on the historical operation data, key attributes, and new adjacency matrix using a pre-trained prediction model.

Images