Methods, devices, equipment and media for predicting the utilization rate of electric vehicle charging piles

Abstract

This invention belongs to the technical fields of intelligent transportation systems, urban computing, and geographic information systems. It provides a method, device, equipment, and medium for predicting the utilization rate of electric vehicle charging piles. The method includes: preprocessing the raw data of the target urban area; determining a geographic adjacency graph and a travel tidal flow graph based on the obtained static and dynamic features; then performing feature weighting fusion using a time feature embedding method; performing two-layer graph attention aggregation on the geographic adjacency graph and the travel tidal flow graph based on the obtained node fusion features to obtain spatial feature embedding; using a gated recurrent unit to learn the evolution of the charging load of the grid over time; using the prediction error of the effective grid as a loss function and training with a dataset to obtain a grid charging load prediction model for subsequent predictions. This invention can effectively improve the model's environmental adaptability and prediction accuracy in different time periods.

Description

Technical Field

[0001] This invention relates to the fields of intelligent transportation systems, urban computing and geographic information systems, and in particular to a method, device, equipment and medium for predicting the utilization rate of electric vehicle charging piles. Background Technology

[0002] With the rapid growth in the number of electric vehicles, the construction of charging infrastructure often lags behind the growth in demand, and there is a significant spatial mismatch problem. Accurately predicting the spatiotemporal distribution of charging load is fundamental to solving problems such as site selection optimization, grid load scheduling, and dynamic electricity pricing.

[0003] Currently, existing technologies mainly include:

[0004] (1) Time-series forecasting methods based on historical charging station data: These methods focus on patterns in the time dimension, employing models such as Autoregressive Moving Average (ARIMA), Support Vector Regression (SVR), or Long Short-Term Memory (LSTM) networks. They extrapolate trends based solely on the time series of historical charging station occupancy. These methods rely on operational data from existing stations, which can easily lead to a "data silo" effect, ignoring the spillover effect of spatially adjacent areas (such as surrounding commercial hotspots) on station load. This results in the model being unable to perceive demand fluctuations caused by changes in the surrounding environment.

[0005] (2) Prediction methods based on the neighborhood characteristics of charging stations: These methods introduce the spatial characteristics of charging stations, divide the study area into grids, and construct neural network models such as convolutional neural networks (CNNs) or graph neural networks (GCNs) based on the spatial relationships of the study unit grids to aggregate information from adjacent grids in geographic space. Most of these methods only construct spatial relationships based on "geographical proximity," but do not consider the significant "work-residence separation" characteristic of urban traffic, and ignore the fact that charging load often shifts between distant residential and work areas along with traffic flow. This prediction method based on geographic proximity cannot capture the tidal flow of travel with such strong functional correlation.

[0006] The existing technology has the following drawbacks:

[0007] Existing prediction methods based on the neighborhood characteristics of charging stations cannot capture the tidal dependence of traffic flow on the number of charging piles occupied by charging stations. They are limited to capturing the mutual influence of adjacent grids in geographic space. However, urban travel has a significant "work-residence separation" characteristic, and charging demand often shifts between residential and work areas along with traffic flow. The static geographic adjacency maps constructed by existing technologies cannot connect areas that are "geographically non-proximity but functionally closely related," resulting in a significant blind spot in the model when predicting cross-regional transfers (i.e., tidal effects) during morning and evening peak hours.

[0008] The existence of data silos and spatial uncertainty of charging stations means that time-series forecasting methods based on historical charging station data fragment the integrity of urban space and fail to perceive the spillover effect of charging demand from other areas to this area. In addition, when processing gridded data, existing technologies often use coarse "point inclusion" mapping rules, ignoring the geographical fact that the service range of charging stations crosses the grid boundaries. This leads to distortion of the statistical data on the actual number of charging piles occupied in the grid, further limiting the accuracy of fine-grained forecasts. Summary of the Invention

[0009] In order to at least solve one of the technical problems existing in the prior art, the present invention provides a method, device, equipment and medium for predicting the utilization rate of electric vehicle charging piles.

[0010] One aspect of the present invention provides a method for predicting the utilization rate of electric vehicle charging stations, comprising:

[0011] The raw data of the target city area is obtained and preprocessed to obtain preprocessed data. The preprocessing includes grid allocation based on service radius, static feature processing and feature normalization. The preprocessed data includes the grid, supply features, static features and dynamic features of the target city area.

[0012] Based on static and dynamic features, the grid is constructed into nodes using a time feature embedding method. Relationship edges between nodes are constructed using geospatial proximity and traffic tidal flow directionality. Based on the nodes and relationship edges, a geographic adjacency graph and a travel tidal flow direction graph are determined.

[0013] The static and dynamic features are weighted and fused using a time feature embedding method. Based on the obtained node fusion features, a two-layer graph attention aggregation is performed on the geographic adjacency graph and the travel tidal flow graph to obtain the spatial feature embedding.

[0014] Based on spatial feature embedding, gated recurrent units are used to learn the evolution of the grid's charging load over time, thereby obtaining the predicted value of the grid's charging load.

[0015] Based on the predicted grid charging load, the prediction error of the effective grid is used as the loss function, and the model is trained using the dataset to obtain the grid charging load prediction model.

[0016] Real-time data of the target city area is obtained, and the real-time data is predicted by the grid charging load prediction model to obtain the predicted power load of electric vehicle charging piles in the target area.

[0017] According to the electric vehicle charging pile utilization prediction method, the raw data of the target urban area is obtained, and the raw data is preprocessed to obtain preprocessed data, including:

[0018] Acquire raw data, which includes historical charging station record data, basic charging station information, urban points of interest data, population grid data, and traffic origin-destination (OD) trajectory data.

[0019] After cleaning the data recorded at the charging stations, the data is aggregated by time to obtain the load sequence of each charging station, where the load sequence is a dynamic feature.

[0020] The target urban area is divided into multiple grids. The allocation weight of the charging station is determined based on the intersection area between the buffer zone of the charging station and the grid. The unit grid load value is determined based on the allocation weight and the load sequence. The supply characteristics are determined based on the unit grid load value and the charging pile data of the charging station.

[0021] The average population density is determined based on the population raster data in each grid, and the static characteristics of the grid are determined based on the urban interest point data, average population density, and supply characteristics in the grid.

[0022] The traffic OD trajectory data is filtered and abnormal data is removed. Then, the number of vehicle flow directions between different grids in each time period is counted to obtain dynamic characteristics.

[0023] In addition, the load sequence and static feature vector are normalized.

[0024] According to the electric vehicle charging pile utilization prediction method, the target urban area is divided into multiple grids. The allocation weight of the charging stations is determined based on the intersection area between the charging station's buffer zone and the grid. The unit grid load value is determined based on the allocation weight and load sequence. Furthermore, the supply characteristics are determined based on the unit grid load value and the charging pile data of the charging stations, including:

[0025] The target city area is divided into multiple square grids, and the buffer zone of the charging stations is determined based on the service radius of the charging stations in the target city area.

[0026] Calculate the intersection area of ​​the buffer zone and the grid. Determine the allocation weights for:

[0027] ;

[0028] in, Indicates the first The grid and the first The intersecting area of ​​the charging stations Indicates the first The charging station for the first The weights assigned to each grid. The total number of grid cells. For grid number identification;

[0029] Based on the assigned weights The load sequence is assigned to the grid to obtain the grid load value. for:

[0030] ;

[0031] in, express Time of the first The grid load value of each grid. For the first Each charging station time The load sequence;

[0032] In addition, by assigning weights, the charging pile data of the charging station and the grid load value are allocated to the grid to obtain the supply characteristics, wherein the charging pile data includes the total number of piles, the number of fast charging piles and the number of slow charging piles.

[0033] According to the electric vehicle charging pile utilization prediction method, based on static and dynamic characteristics, a time feature embedding method is used to construct a grid as nodes, and geographic spatial proximity and traffic tidal flow direction are used to construct the relationship edges of the nodes. Based on the nodes and relationship edges, a geographic adjacency graph and a travel tidal flow direction graph are determined, including:

[0034] Temporal context extraction is performed on static and dynamic features. The temporal context is mapped into dense vectors through a predefined embedding layer and then concatenated to obtain... The time context vector at a given moment for:

[0035] ;

[0036] in, For splicing operations, express A dense vector of time-hour information. for A dense vector of time and day information;

[0037] Static features, dynamic features, and time context vectors are used as grid features. Based on the grid and grid features, a relationship graph of relational edges and charging load is constructed.

[0038] The connection weights of two meshes are determined based on whether they share a common boundary or a common vertex in spatial geometry. According to connection weight By obtaining the geographical adjacency edges, a geographical adjacency graph can be obtained, in which... , and This represents two different grids;

[0039] Based on the number of vehicles flowing in different directions during typical time periods, directed edges are established between two grids. The probability distribution of flow direction within the target city area is calculated based on the number of vehicles flowing in different directions, thus obtaining the tidal adjacency matrix. :

[0040] ;

[0041] in, Represents a grid With grid In the The weights of the tidal adjacency matrix for each time period. Indicates the first Grid for each time period To grid The number of vehicles flowing in the direction of traffic. Indicates the first Time period from grid Total number of departing vehicles The grid number is used as an identifier; the tidal connection edges are determined according to the weights of the tidal adjacency matrix, thus obtaining the travel tidal flow map.

[0042] According to the electric vehicle charging pile utilization prediction method, static and dynamic features are weighted and fused using a time feature embedding method. Based on the obtained node fusion features, a two-layer graph attention aggregation is performed on the geographical adjacency graph and the travel tidal flow graph to obtain spatial feature embedding, including:

[0043] A learnable linear transformation matrix is ​​used to map static and dynamic features to a unified hidden layer dimension:

[0044] ;

[0045] ;

[0046] in, This is the dimensional mapping result of dynamic features. As a dynamic feature, The linear transformation matrix is ​​a dynamic feature. For dynamic feature dimensions, For the hidden layer dimension, For dimensions; This is the dimensionality mapping result for static features. It is a static feature. The linear transformation matrix is ​​a static feature. For static feature dimensions;

[0047] Using the aforementioned time context vector as prior knowledge, a multilayer perceptron is used to dynamically calculate the feature weights of the dimension-aligned static and dynamic features, resulting in the following weights for the static and dynamic features:

[0048] ;

[0049] in, For a moment The weights of static features at time. For a moment The weights of the dynamic features at time, where and , It is the Sigmoid activation function. and For linear layers, It is the ReLU activation function;

[0050] The node fusion features are obtained by element-wise weighted summation based on the weights of static and dynamic features:

[0051] ;

[0052] in, Represents a node At any moment The fusion characteristics of time, Represents a node static characteristics, node At any moment Dynamic characteristics of time;

[0053] Based on the node fusion characteristics, a graph attention mechanism is used to obtain the attention coefficient between two nodes with related edges in the geographic adjacency graph and the travel tidal flow graph. Based on the attention coefficient, the neighbor features of the nodes are aggregated to obtain the geographic adjacency graph features and the travel tidal flow graph features.

[0054] Geographic adjacency graph features and travel tidal flow graph features are fused through a linear layer to obtain spatial feature embeddings. for:

[0055] ;

[0056] in, Features of a geographic adjacency graph. Indicates a linear layer. This represents the characteristics of the travel tidal flow map.

[0057] According to the electric vehicle charging pile utilization prediction method, based on spatial feature embedding, a gated recurrent unit is used to learn the time evolution of the grid's charging load to obtain the grid charging load prediction value, including:

[0058] Feature sequences embedded with spatial features Reset gate via the gated loop unit and Update Gate To forget and to retain:

[0059] ;

[0060] ;

[0061] ;

[0062] ;

[0063] in, The current hidden state. This is the hidden state from the previous moment; In the candidate hidden state, The input weights are the candidate hidden states. The hidden weights of the candidate hidden states. The bias for the candidate hidden state; To reset the input weights of the gate, To reset the hidden weight of the door, To reset the door offset; To update the input weights of the gate, To update the hidden weight of the door, To update the door offset; For the Sigmoid function, It is the hyperbolic tangent function. This indicates element-wise multiplication;

[0064] The output of the last time step of the gated loop unit is obtained and processed by the decoder of the gated loop unit to obtain the charging load prediction value:

[0065] ;

[0066] in, To control the output of the last time step of the loop unit.

[0067] According to the electric vehicle charging pile utilization prediction method, the prediction error of the effective grid is used as the loss function based on the grid charging load prediction value, and a dataset is used for training to obtain the grid charging load prediction model, including:

[0068] The mask mean square error is used as the loss function. Calculate the prediction error of the effective grid:

[0069] ;

[0070] in, This represents the actual load value. To predict load values, As a mask indicator variable, when the mesh When the memory is at a valid charging station or historical data Otherwise, it is 0; For L2 regularization terms, For hyperparameters;

[0071] Based on loss function The model is trained using the Adam optimizer and a dataset, and the training is repeated until the model accuracy meets the preset requirements, thus obtaining the grid charging load prediction model.

[0072] Another aspect of the present invention provides an electric vehicle charging pile utilization prediction device, comprising:

[0073] The first module is used to acquire the raw data of the target city area, preprocess the raw data to obtain preprocessed data, and the preprocessing includes grid allocation based on service radius, static feature processing and feature normalization. The preprocessed data includes the grid, supply features, static features and dynamic features of the target city area.

[0074] The second module is used to construct the grid into nodes based on static and dynamic features using a time feature embedding method, construct the relationship edges of the nodes using geospatial proximity and traffic tidal flow, and determine the geographic adjacency graph and travel tidal flow graph based on the nodes and relationship edges.

[0075] The third module is used to perform feature weighting and fusion of static and dynamic features using time feature embedding. Based on the obtained node fusion features, a two-layer graph attention aggregation is performed on the geographic adjacency graph and the travel tidal flow graph to obtain spatial feature embedding.

[0076] The fourth module is used to learn the evolution of the grid's charging load over time by embedding spatial features and using gated recurrent units to obtain the predicted value of the grid's charging load.

[0077] The fifth module is used to obtain a trained grid charging load prediction model by using the prediction error of the effective grid as the loss function and training it with the dataset based on the grid charging load prediction value.

[0078] The sixth module is used to acquire real-time data of the target urban area, and to predict the utilization rate of electric vehicle charging piles in the target area by using a trained grid charging load prediction model.

[0079] Another aspect of the present invention provides an electronic device, including a processor and a memory;

[0080] The memory is used to store programs;

[0081] The processor executes the program to implement the method as described above.

[0082] This invention also discloses a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device can read the computer instructions from the computer-readable storage medium and execute the computer instructions, causing the computer device to perform the methods described above.

[0083] The beneficial effects of this invention are as follows: The area-weighted allocation method based on service radius enables refined allocation of load and supply characteristics, fully guaranteeing the physical consistency and high signal-to-noise ratio of grid-level ground truth data; By coupling a two-layer graph structure with geographic adjacency and dynamic travel tidal interactions, and utilizing a two-layer graph attention network to simultaneously aggregate local diffusion features and long-distance flow direction features, reliable fitting of interactive spatial dependencies is fully guaranteed; The dynamic and static feature gating fusion mechanism based on time embedding utilizes a gating network to dynamically generate weights according to the current time context, achieving decoupling and weighted fusion of static and dynamic features, effectively improving the model's environmental adaptability and prediction accuracy in different time periods (such as late night and peak hours). Attached Figure Description

[0084] Figure 1 This is a schematic diagram of the electric vehicle charging pile utilization prediction process according to an embodiment of the present invention;

[0085] Figure 2 This is a schematic diagram of the data preprocessing process according to an embodiment of the present invention;

[0086] Figure 3 This is a charging load data grid allocation diagram based on service radius according to an embodiment of the present invention;

[0087] Figure 4 This is a directed travel tidal flow edge graph according to an embodiment of the present invention;

[0088] Figure 5 This is a diagram of a time-embedded dynamic-static feature fusion method according to an embodiment of the present invention;

[0089] Figure 6This is a detailed network structure diagram of the model in an embodiment of the present invention;

[0090] Figure 7 This is a visual comparison chart of the predicted charging load heat map and the actual heat map according to an embodiment of the present invention;

[0091] Figure 8 This is a comparison chart of charging load prediction results in a portion of the actual dataset according to an embodiment of the present invention;

[0092] Figure 9 This is a schematic diagram of an electric vehicle charging pile utilization prediction device according to an embodiment of the present invention. Detailed Implementation

[0093] The embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings. Throughout the description, the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions. In the following description, suffixes such as "module," "part," or "unit" used to denote elements are used only for the purpose of illustrative purposes and have no specific meaning in themselves. Therefore, "module," "part," or "unit" can be used interchangeably. Terms such as "first," "second," etc., are used only to distinguish technical features and should not be construed as indicating or implying relative importance, or implicitly indicating the number of indicated technical features, or implicitly indicating the sequential relationship of the indicated technical features. In the following description, the consecutive reference numerals for method steps are for ease of review and understanding. Adjusting the implementation order of steps, in conjunction with the overall technical solution of the present invention and the logical relationship between the various steps, will not affect the technical effect achieved by the technical solution of the present invention. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0094] refer to Figure 1 , Figure 1 This is a schematic flowchart of the electric vehicle charging pile utilization prediction method according to an embodiment of the present invention, which includes, but is not limited to, steps S100 to S600:

[0095] S100: Obtain the raw data of the target city area, preprocess the raw data to obtain preprocessed data, including grid allocation based on service radius, static feature processing and feature normalization, and the preprocessed data includes the grid, supply features, static features and dynamic features of the target city area.

[0096] It should be noted that the raw data obtained in the embodiments of the present invention is the relevant data of electric vehicle charging piles / charging stations generated in the target urban area at historical times. Subsequent embodiments construct a grid charging load prediction model based on these raw data, and then obtain real-time data of the target urban area to predict the load of electric vehicle charging piles / charging stations at future times.

[0097] In some embodiments, reference Figure 2 The data preprocessing flowchart shown includes, but is not limited to, steps S110 to S160:

[0098] S110, Obtain raw data, which includes historical charging station record data, basic charging station information, urban point of interest data, population raster data, and traffic OD trajectory data.

[0099] In some embodiments, the charging station recorded data may be original charging order data; the basic information of the charging station includes the latitude and longitude of the charging station, the number of charging piles, etc.; the city point of interest (POI) data and population density raster data are obtained from a third-party map information provider platform, wherein the city point of interest (POI) data includes residential, office, commercial, transportation facilities, etc.; the traffic OD trajectory data may be taxi / ride-hailing OD trajectory data, etc.

[0100] It is understood that, before performing grid-level charging load prediction, the embodiments of the present invention require refined data preprocessing to ensure the quality and spatial alignment of multi-source data.

[0101] S120: After cleaning the charging station recorded data, aggregate it by time to obtain the load sequence of each charging station, where the load sequence is a dynamic feature.

[0102] In some embodiments, the original charging order data is cleaned to remove abnormal charging station data with high data missing rate or zero power, and the data is aggregated by hourly granularity to obtain the hourly load sequence of each charging station.

[0103] S130 divides the target city area into multiple grids, determines the allocation weight of charging stations based on the intersection area of ​​the buffer zone of the charging station and the grid, determines the unit grid load value based on the allocation weight and load sequence, and determines the supply characteristics based on the unit grid load value and the charging pile data of the charging station.

[0104] In some embodiments, such as Figure 3 The diagram shown is a grid allocation map of charging load data based on service radius. Figure 3The circles in the diagram represent the service radius of the charging station, and different colored areas represent the intersections of different grids with the circles. Furthermore, in this embodiment of the invention, the target city area is divided into multiple square grids, and the buffer zone of the charging station is determined based on the service radius of the charging station in the target city area.

[0105] Calculate the intersection area of ​​the buffer zone and the grid. Determine the allocation weights for:

[0106] ;

[0107] in, Indicates the first The grid and the first The intersecting area of ​​the charging stations Indicates the first The charging station for the first The weights assigned to each grid. The total number of grid cells. For grid number identification;

[0108] Based on the assigned weights The load sequence is assigned to the grid to obtain the grid load value. for:

[0109] ;

[0110] in, express Time of the first The grid load value of each grid. For the first Each charging station time The load sequence;

[0111] In addition, by assigning weights, the charging pile data of the charging station and the grid load value are allocated to the grid to obtain the supply characteristics, wherein the charging pile data includes the total number of piles, the number of fast charging piles and the number of slow charging piles.

[0112] In some embodiments, data is aggregated at an hourly granularity to obtain the hourly load sequence for each charging station. To address the boundary errors caused by traditional point mapping, this embodiment of the invention divides the study area into regions based on the aforementioned area-weighted mapping method using service radius. A square grid of size 10 is used to traverse the grid and encode the values ​​sequentially from 1 to 10. For any charging station Set the service radius according to its size. (e.g., 500 meters), construct a buffer zone.

[0113] S140, determine the average population density based on the population raster data in each grid, and determine the static characteristics of the grid based on the urban interest point data, average population density, and supply characteristics in the grid.

[0114] In some embodiments, the number of POIs of each type within each spatial grid is counted; the average population density of each grid is extracted using raster sampling. The supply-side features (number of stakes), POI features, and population features described above are combined to construct a static feature vector for the grid. .

[0115] S150 filters and removes abnormal data from traffic OD trajectory data, and then counts the number of vehicle flow directions between different grids in each time period to obtain dynamic characteristics.

[0116] In some embodiments, taxi / ride-hailing vehicle OD trajectory data is filtered to remove abnormal trajectories and statistically analyzed from the grid for each time period. Flow to grid The number of vehicle trajectories was determined by processing traffic OD data, which enabled the subsequent dynamic tidal map to incorporate supplementary data based on OD trajectory data, ensuring a reliable fit to the OD interactive spatial dependency relationship.

[0117] S160 normalizes the load sequence and static eigenvectors.

[0118] In some embodiments, the normalization formula is:

[0119] ;

[0120] in, This represents the result after normalization; It is a type of raw data before normalization; and These are the maximum and minimum values ​​in the original data, respectively. The normalization process in this embodiment normalizes the various statistical values ​​in the grid charging load sequence and the static feature vector to... Intervals are used to eliminate the influence of dimensions.

[0121] S200 uses time feature embedding to construct the grid into nodes based on static and dynamic features, and uses geospatial proximity and traffic tidal flow to construct the relationship edges of the nodes. Based on the nodes and relationship edges, the geographic adjacency graph and the travel tidal flow graph are determined.

[0122] In order to capture the complex spatial feature relationships related to charging load, this invention adopts a two-layer graph attention network modeling method oriented towards grid charging load. In this method, the grid is modeled as the nodes of a graph, and the edges of the graph are defined as the proximity relationship between two grids in geographic space and the tidal flow direction graph of urban travel patterns.

[0123] In some embodiments, temporal context extraction is performed on static and dynamic features. The temporal context is mapped into dense vectors through a predefined embedding layer and then concatenated to obtain... The time context vector at a given moment for:

[0124] ;

[0125] in, For splicing operations, express A dense vector of time-hour information. for A dense vector of time and day information;

[0126] Static features, dynamic features, and time context vectors are used as grid features. Based on the grid and grid features, a relationship graph of relational edges and charging load is constructed.

[0127] The connection weights of two meshes are determined based on whether they share a common boundary or a common vertex in spatial geometry. According to connection weight By obtaining the geographical adjacency edges, a geographical adjacency graph can be obtained, in which... , and This represents two different grids;

[0128] Based on the number of vehicles flowing in different directions during typical time periods, directed edges are established between two grids. The probability distribution of flow direction within the target city area is calculated based on the number of vehicles flowing in different directions, thus obtaining the tidal adjacency matrix. :

[0129] ;

[0130] in, Represents a grid With grid In the The weights of the tidal adjacency matrix for each time period. Indicates the first Grid for each time period To grid The number of vehicles flowing in the direction of traffic. Indicates the first Time period from grid Total number of departing vehicles The grid number is used as an identifier; the tidal connection edges are determined according to the weights of the tidal adjacency matrix, thus obtaining the travel tidal flow map.

[0131] In some embodiments, geographic proximity edge modeling uses the geographic spatial adjacency relationships between grids as a first type of edge relationship constraint. A geographic adjacency matrix is ​​then constructed. The judgment rules are as follows: If the grid... With grid If there are common boundaries or common vertices in spatial geometry (i.e., belonging to an eight-neighborhood relationship), then connection weights are set. Otherwise, set it to 0.

[0132] In some embodiments, such as Figure 4 The directed travel tidal flow edge graph shown in this embodiment of the invention demonstrates time-segmented directed travel tidal flow edge modeling, including constructing a tidal graph reflecting long-distance dependence using taxi / ride-hailing vehicle OD trajectory data. Considering the significant time-varying characteristics of traffic flow, a simple static graph cannot reflect the differences in pedestrian movement during morning and evening peak hours caused by the separation of work and residence. The specific construction process is as follows:

[0133] (1) Time division: Divide a day into 1000-12 ... Typical time periods (such as morning rush hour, off-peak hour, evening rush hour, and nighttime).

[0134] (2) Directed flow statistics: for the current time period Time period Track the origin-destination (OD) trajectories of all vehicles within that time period. Statistically analyze the data collected from the grid within that time period. To grid Number of vehicles transferred And establish directed edge connections.

[0135] (3) Calculation of transition probability (row normalization): In order to eliminate the influence of the difference in the dimensions of traffic flow in different areas, the probability of the flow direction of urban travel is calculated.

[0136] In some embodiments, the selection of a dynamic graph includes the generation thereof. Zhang's sparse adjacency matrix In subsequent prediction processes, the model will dynamically index the corresponding tidal map based on the timestamps of the input historical data to perform feature aggregation.

[0137] S300 uses time feature embedding to perform feature weighting and fusion of static and dynamic features. Based on the obtained node fusion features, it performs two-layer graph attention aggregation on the geographic adjacency graph and the travel tidal flow graph to obtain spatial feature embedding.

[0138] It should be noted that traditional feature concatenation methods cannot handle static inherent properties. With dynamic historical trends To address the issues of inconsistent dimensions and weights changing over time, this invention employs a fusion network that includes feature projection alignment and a temporal gating mechanism to perform the following processing:

[0139] refer to Figure 5 The diagram of the motion-static feature fusion method based on temporal embedding is shown below:

[0140] A learnable linear transformation matrix is ​​used to map static and dynamic features to a unified hidden layer dimension:

[0141] ;

[0142] ;

[0143] in, This is the dimensional mapping result of dynamic features. As a dynamic feature, The linear transformation matrix is ​​a dynamic feature. For dynamic feature dimensions, For the hidden layer dimension, For dimensions; This is the dimensionality mapping result for static features. It is a static feature. The linear transformation matrix is ​​a static feature. For static feature dimensions;

[0144] Using the temporal context vector as prior knowledge, a multilayer perceptron is used to dynamically calculate the feature weights of the dimension-aligned static and dynamic features, resulting in the following weights for the static and dynamic features:

[0145] ;

[0146] in, For a moment The weights of static features at time. For a moment The weights of the dynamic features at time, where and For example, at night High, peak period High; It is the Sigmoid activation function. and For linear layers, It is the ReLU activation function;

[0147] The node fusion features are obtained by element-wise weighted summation based on the weights of static and dynamic features:

[0148] ;

[0149] in, Represents a node At any moment The fusion characteristics of time, Represents a node static characteristics, node At any moment Dynamic characteristics of time;

[0150] Based on the node fusion characteristics, a graph attention mechanism is used to obtain the attention coefficient between two nodes with related edges in the geographic adjacency graph and the travel tidal flow graph. Based on the attention coefficient, the neighbor features of the nodes are aggregated to obtain the geographic adjacency graph features and the travel tidal flow graph features.

[0151] In some embodiments, in order to simultaneously capture the local geospatial features of the grid and the tidal features of travel interactions, a graph attention mechanism is applied in parallel on the geographic adjacency graph and the urban travel tidal flow map, respectively:

[0152] Taking a geographic adjacency graph as an example, calculate the nodes. with neighbors Attention coefficient between :

[0153] ;

[0154] ;

[0155] In the formula, This indicates a splicing operation. This is the attention vector.

[0156] Using normalized attention coefficients Aggregate neighbor features to obtain geographical adjacency graph features. :

[0157] ;

[0158] Geographic adjacency graph features and travel tidal flow graph features are fused through a linear layer to obtain spatial feature embeddings. for:

[0159] ;

[0160] in, Features of a geographic adjacency graph. Indicates a linear layer. This represents the characteristics of the travel tidal flow map.

[0161] S400, based on spatial feature embedding, uses gated cyclic units to learn the evolution of the grid's charging load over time, and obtains the predicted value of the grid's charging load.

[0162] In some embodiments, a feature sequence in which spatial features are embedded Reset gate via the gated loop unit and Update Gate To forget and to retain:

[0163] ;

[0164] ;

[0165] ;

[0166] ;

[0167] in, The current hidden state. This is the hidden state from the previous moment; In the candidate hidden state, The input weights are the candidate hidden states. The hidden weights of the candidate hidden states. The bias for the candidate hidden state; To reset the input weights of the gate, To reset the hidden weight of the door, To reset the door offset; To update the input weights of the gate, To update the hidden weight of the door, To update the door offset; For the Sigmoid function, It is the hyperbolic tangent function. This indicates element-wise multiplication;

[0168] The output of the last time step of the gated loop unit is obtained and processed by the decoder of the gated loop unit to obtain the charging load prediction value:

[0169] ;

[0170] in, To control the output of the last time step of the loop unit.

[0171] S500 uses the prediction error of the effective grid as the loss function based on the grid charging load prediction value and trains it using the dataset to obtain the grid charging load prediction model.

[0172] In some embodiments, the mask mean square error is used as the loss function. Calculate the prediction error of the effective grid:

[0173] ;

[0174] in, This represents the actual load value. To predict load values, As a mask indicator variable, when the mesh When the memory is at a valid charging station or historical data Otherwise, it is 0; For L2 regularization terms, For hyperparameters;

[0175] Based on loss function The model is trained using the Adam optimizer and a dataset, and the training is repeated until the model accuracy meets the preset requirements, thus obtaining the grid charging load prediction model.

[0176] In some embodiments, all samples are divided into a training set (first 80%) and a test set (last 20%) in chronological order; then, the time series of the training set is input into the model constructed in step S300, and the corresponding tidal map is dynamically selected according to the current time. The model is used for calculation; the Adam optimizer is used to adjust the model parameters according to the loss function until the error converges or the maximum number of iterations is reached; finally, the optimal parameters are saved, and the model accuracy (such as RMSE and MAE indicators) is verified using the test set. If the accuracy meets the requirements, the final grid charging load prediction model is obtained.

[0177] The S600 acquires real-time data for the target urban area, uses a grid charging load prediction model to predict the real-time data, and obtains the predicted electricity load of electric vehicle charging piles in the target area.

[0178] In some embodiments, the detailed network structure diagram of the model of the present invention is as follows: Figure 6 As shown, the construction process is as follows: real-time acquisition of the static features and dynamic load sequence of the entire city grid at the current moment; matching the tidal map of the current time period according to rules; inputting the data into the trained model, the model uses the time embedding module to calculate the current dynamic and static weights, and performs inference through two-layer ST-GAT and GRU; outputting the charging load prediction matrix of each grid in the city at future moments, providing decision support for power grid dispatch.

[0179] In some embodiments, the specific implementation process of the present invention is illustrated using electric vehicle charging order data (including station ID, charging start and end times, charging amount, etc., from a charging operator management platform) for a city within a time period (one month), basic geographic information data (including road network and population heat map raster data, from WorldPop and OpenStreetMap), POI data (including four core POI data types: residential areas, office buildings, shopping malls, and parking lots, from an open platform), and taxi OD trajectory data for a city.

[0180] (1) In this embodiment, a city-wide grid was selected as the study area, and hourly charging load data from 00:00 to 23:00 daily during the aforementioned time period was used. The study area and gridding diagram are shown above. Figure 3 .

[0181] (2) Data preprocessing: Before performing grid-level charging load prediction, this embodiment of the invention uses refined data preprocessing to ensure the quality of multi-source data and the accuracy of spatial mapping:

[0182] The service radius weighted allocation of charging load data is performed, dividing the study area into... A square grid of size 10 is used, and the grid is traversed using Morton coding, with each element encoded sequentially from 1 to 1. For each charging station Set the effective service radius according to its size. Construct a circular buffer. Calculate the relationship between this buffer and each grid cell. Geometric intersection area Calculate the assigned weights :

[0183] ;

[0184] Use this weight to optimize the site In time slice The original load Precisely assigned to the grid To obtain the true value of the grid-level load. :

[0185] ;

[0186] The total number of charging piles, fast charging piles, and slow charging piles of each station are allocated to the grid using this weight, which serves as the supply-side feature of the grid.

[0187] Static data feature preprocessing is performed to extract four types of facilities—residential, work, commercial, and transportation—from the POI dataset. The number of each type of POI in each spatial grid of the study area is counted. Population heat map raster data is extracted, and the average population density in each grid is calculated, thereby completing the extraction and spatial alignment of static features.

[0188] Urban travel interaction flow data preprocessing was performed, utilizing taxi origin-destination (OD) trajectory data to divide the day into four typical time periods: morning peak (7-10 am), off-peak (10-5 pm), evening peak (5-8 pm), and nighttime (8-7 pm). Data from the grid was then statistically analyzed within each time period. Flow to grid The number of vehicle trajectories is used as the basic data for constructing the tidal map.

[0189] Finally, data normalization was performed. To eliminate the influence of data dimensions, grid charging load, the number of various POI statistics, population density, and the number of supply-side charging piles were normalized to [specific values ​​to be filled in]. middle.

[0190] (3) Grid charging load relationship graph and feature construction. In this step, it is necessary to model the grid graph nodes, dynamic and static data feature fusion, and two-layer graph network structure; first, graph nodes and feature extraction for the grid; given a certain time slice The feature set of the entire city grid is represented as a matrix. For the first Each grid has its eigenvector. From static features (Including POI, population, number of stakes) and dynamic features (Historical load sequence) composition; in order to capture the time-to-time influence mechanism, this invention designs a time embedding and gating mechanism; the current hour and week Mapped to a dense vector through the Embedding layer A gating network is constructed using a multilayer perceptron (MLP) to generate weights. (Static weights) and (Dynamic weights):

[0191] ;

[0192] ;

[0193] In the formula, Use the Sigmoid activation function to ensure that the weights are within the range of 1 / 2 Hz. Between these steps, the importance of dynamic and static features is dynamically adjusted based on the temporal context (such as late night or peak hours); then, multi-relationship edge modeling is performed on a grid-oriented two-layer graph network; this invention constructs two graph structures: a geographic adjacency graph and an urban travel tidal graph; in the construction of the geographic adjacency graph, a geographic adjacency matrix is ​​constructed. If the grid With grid If there are common boundaries or vertices in geographic space (i.e., eight-neighborhood relationships), then let... Otherwise, it is 0.

[0194] In urban travel tidal view modeling, this embodiment constructs a dynamic tidal network based on the actual traffic operation characteristics of a city and utilizing four typical time periods (morning peak, off-peak, evening peak, and night) defined in the preprocessing stage. This network is based on the current time period. Time period Dynamically load the corresponding tidal adjacency matrix The matrix is ​​constructed as follows:

[0195] OD mapping and statistics, using taxi / ride-hailing trajectory data, statistical analysis over time periods. Inner from grid Flow to grid Number of vehicles transferred .

[0196] Transition probability calculation (row normalization): To accurately reflect the probability distribution of vehicle flow direction in different areas, the flow matrix is ​​row normalized.

[0197] ;

[0198] In the formula, the denominator represents the time period from the grid. The total number of all departing vehicles is the result of the grid calculation. The vehicles flowed into the grid in the next moment. The probability of.

[0199] Sparsification: In order to reduce computational complexity and remove noise, a sparsification strategy is set (only valid edges with transition probabilities greater than 0 are retained) to obtain the final sparse directed adjacency matrix.

[0200] (4) Construction of grid charging load prediction model. In this step, it is necessary to complete the temporal fusion and alignment of grid dynamic and static features, the construction of charging load prediction sample sequence, the construction of spatial dependency learning module based on two-layer graph attention network, the construction of temporal dependency learning module based on gated recurrent unit (GRU), and the construction of charging load decoding model.

[0201] First, temporal fusion and alignment of the static and dynamic features of the grid are performed. Electric vehicle charging load is influenced by both static inherent attributes (such as POI, population, and number of charging stations) and dynamic historical trends. To align these two types of heterogeneous data at the model input, a unified feature embedding space needs to be constructed. The number of grid cells is set to... For any time period Extract the static feature matrix at that moment. and dynamic load characteristic matrix To address the issue of shifting importance of dynamic and static features across different time periods (such as late night and peak hours), this invention does not employ simple splicing and filling. Instead, it utilizes temporal embedding to generate temporally gated weights. Specifically, this embodiment uses the following parameter configuration to achieve feature alignment and fusion:

[0202] Feature definition, extracting static features of the mesh. Its dimensions (Including the normalized total number of stakes, fast / slow fill counts, population density, and the number of POIs in the four categories); extract dynamic features. Its dimensions (i.e., historical load values);

[0203] Feature projection, setting a uniform hidden layer dimension By using two independent fully connected layers, the 8-dimensional static features and 1-dimensional dynamic features are projected onto a 64-dimensional feature space, respectively, thus solving the dimension mismatch problem.

[0204] Gating weight generation and time slice extraction The "hours" (0-23) and "weekdays" (0-6) are concatenated into a 20-dimensional time context vector through an embedding layer with an embedding dimension of 10; a gated network MLP is constructed, with the following structure: [Input layer 20-dimensional...] 32-dimensional hidden layer (ReLU) The output layer is 2D (Sigmoid), which outputs two scalar weights between 0 and 1. and .

[0205] Weighted fusion uses weights to sum the projected features, resulting in a feature of dimension 1. fusion feature matrix This serves as the input for the subsequent two-layer GAT network.

[0206] Then, a sample sequence for grid charging load prediction is constructed. The grid charging load exhibits significant temporal autocorrelation and periodicity; given the time slice to be predicted... The future charging load of any of its grid nodes is used as the output label; in order to capture time-series dependencies, a time slice is selected. Previous The feature matrix of each consecutive historical moment is used as the input feature sequence. Wherein, the length of the input sequence is... The settings are based on the periodic characteristics of the charging load (e.g., in this embodiment). (representing a historical window of the past 12 hours); the entire city For each time slice of a grid, "input feature sequence - output label" pairs are constructed in the manner described above; specifically, for a total time span of... The dataset can be constructed using the sliding window method. There are training samples; the dimension of the input feature sequence is . ,in For batch size, The feature dimension is [value]; the output label dimension is [value]. .

[0207] Subsequently, a spatial dependency learning module for travel flows is constructed based on a two-layer graph attention network (GAT). The nonlinear dependencies of grid load features from different spatial perspectives (geographical adjacency and travel tidal interaction flows) are fused using a two-layer GAT. Given a fused feature matrix in a time slice... Feature aggregation is performed on both the geographic adjacency graph and the travel tidal graph. For the geographic adjacency graph, a geographic adjacency matrix is ​​constructed using the eight-neighbor relationships of the grid. . No. The grid node and the first The process of calculating the attention coefficient of each neighbor node in the geographic view is as follows:

[0208] ;

[0209] ;

[0210] In the formula, For attention vectors, The learnable weight matrix under the geographic view. This indicates a splicing operation. Represents a grid The set of geographical neighbors.

[0211] For the travel tidal map, the constructed time-segmented tidal adjacency matrix is ​​used. According to the current time slice The corresponding tidal matrix is ​​dynamically selected based on the time period (morning peak, off-peak, etc.); similarly, the attention coefficient under the tidal view is calculated. First, aggregate the features of distantly related nodes; finally, fuse the aggregated features of the two views. Spatial feature embedding vector of each grid node The calculation is as follows:

[0212] ;

[0213] ;

[0214] ;

[0215] In the formula, GAT represents the graph attention calculation process. That is, a time slice The city-wide grid feature matrix after dual-view space aggregation.

[0216] Therefore, for the time slice to be predicted Input it into the feature sequence The feature matrix of each time step is used to perform feature transfer and aggregation across all nodes according to the GRU formula for predicting load size, thus obtaining the past... The sequence of graph space feature embedding vector matrices for each time slice is represented as follows:

[0217] ;

[0218] In the formula, This represents a temporal feature tensor that includes information on geographic proximity and tidal spatial interaction.

[0219] Then, a charging load time dependency learning module is constructed based on a gated recurrent unit (GRU). The GRU learns the time dependency of the grid space feature embedding vector matrix. Given the sequence output by the GAT... The Middle Embedded vector matrix As input to the current time step, As the output of the hidden layer at the current time step As the output of the hidden layer at the previous time step, a single computation of GRU can be performed according to the following formula:

[0220] ;

[0221] ;

[0222] ;

[0223] ;

[0224] In the formula, This represents the Hadamard product (element-by-element multiplication). These represent the training parameter matrices of the gated recurrent unit, obtained during model training; This represents the bias variable. According to the formula... arrive The calculations are performed sequentially, and the final output is obtained. That is, it includes the past. The final hidden state of spatiotemporal information at a given moment. It is worth noting that when... When it is the input of the first time step, Initialize to a zero matrix.

[0225] Finally, a grid charging load decoding model is constructed to obtain the time slice to be predicted. Based on the grid charging load value, this invention constructs a decoder model consisting of a two-layer fully connected network (MLP) and a ReLU activation function, which can be derived from the final hidden state. Decode the whole city Predicted load size for each grid:

[0226] ;

[0227] In the formula, These are the learning parameters for the first and second layers of a fully connected network, respectively. It is the time slice to be predicted output by the model. The grid charging load matrix, and All represent bias terms. The ReLU activation function is specifically introduced here as a constraint on the output layer to ensure that the predicted charging load value is always non-negative. ), which is consistent with the physical meaning.

[0228] (5) Training and application of grid charging load prediction model.

[0229] This step includes three parts: loss function design, model training, and application.

[0230] First, the Masked Mean Square Error (MaskedMSE) loss function is designed. Considering the sparsity of the grid data, the error is calculated only for the grid containing valid data:

[0231] ;

[0232] In the formula, As a mask indicator variable, when the mesh When memory is in a valid charging facility Otherwise, it is 0; This is an L2 regularization term. Then, model training is performed. The first 80% of the time slices from all samples are used as training data, and the last 20% as test data. The training data is input into the constructed model, and the corresponding tidal map is dynamically loaded according to the time label of each batch. The Adam optimizer is used to adjust the model parameters based on the MaskedMSE loss until the error converges. Finally, the optimal parameters are saved, and the model accuracy (e.g., RMSE, MAE metrics) is validated using test data.

[0233] Finally, the grid charging load prediction model is applied. The static features and dynamic load sequences of the entire city's grid are obtained in real time using the above method and input into the trained model. The model first uses the time embedding gating module to calculate the static and dynamic weights, then performs parallel convolution through the geographic adjacency graph and tidal graph, and finally outputs the charging load prediction matrix of each grid in the city at future time through GRU inference.

[0234] (6) Experimental Results and Analysis

[0235] To verify the effectiveness of the method proposed in this embodiment of the invention, the moment with the highest total load in the entire city in the test set was selected ( Spatial distribution visualization and quantitative assessment were performed, and the results are as follows: Figure 7 The following is a visualization comparison of the predicted and actual charging load heat maps:

[0236] Spatial distribution consistency analysis shows that the predicted heatmap generated by the model in this embodiment of the invention has a very high degree of consistency with the actual value heatmap in terms of spatial distribution.

[0237] Fitting high-load areas, in core business districts and densely populated residential areas such as Nanshan District and Futian District of a certain city ( Figure 7 (Mid-deep red patches) The model accurately predicted the high-intensity charging demand without any obvious peak flattening. This indicates that the area-weighted mapping based on the service radius effectively preserves the spatial heterogeneity of the load, and the time-embedded gating module accurately captures the surge trend of dynamic traffic during peak hours.

[0238] Edge region fitting: In urban edges or areas with low load (light-colored patches), the predicted values ​​transition smoothly without abnormal noise, verifying the effectiveness of dual-view GAT in constraining spatial dependencies.

[0239] In terms of quantitative accuracy index analysis, to eliminate the interference of numerous invalid areas (such as water bodies and green spaces with zero load) in the urban grid on the evaluation indicators, this embodiment adopts a dual evaluation standard, namely, global indicators and effective area indicators (Load>1):

[0240] Global assessment: The model achieved an accuracy of 84.20% across all grid areas in the city, with a root mean square error (RMSE) as low as 1.43.

[0241] Effective region evaluation: For effective active grids with a charging load greater than 1.0 (i.e., regions with Mask=1), the model's accuracy was further improved to 84.87%, and the RMSE was 2.63.

[0242] Therefore, combining Figure 8 The charging load prediction results of the model in this embodiment of the invention shown are obtained on a portion of the real-world dataset. The model represents the prediction error for different grids (including their corresponding TOP values) over the past 48 hours. It can be determined that the accuracy of the model in the effective region is slightly higher than the global accuracy. The mask mean squared error loss function during training allows the model to focus on learning load change patterns with real physical meaning, rather than simply fitting a zero-value background. The numerical difference in RMSE (2.63 vs 1.43) is within a reasonable range because the load base in the active region is inherently large (exceeding 80 for some grids), yet the relative error remains at an extremely low level.

[0243] Figure 9 This is a schematic diagram of an electric vehicle charging pile utilization prediction device according to an embodiment of the present invention. The device includes a first module 1110, a second module 1120, a third module 1130, a fourth module 1140, a fifth module 1150, and a sixth module 1160.

[0244] The system comprises three modules: The first module acquires raw data for the target city area and preprocesses it to obtain preprocessed data. Preprocessing includes grid allocation based on service radius, static feature processing, and feature normalization. The preprocessed data includes the grid, supply characteristics, static features, and dynamic features of the target city area. The second module constructs nodes from the grid using temporal feature embedding based on the static and dynamic features. It then constructs relationship edges between nodes using geospatial proximity and traffic tidal flow directionality. Based on the nodes and relationship edges, it determines the geographic adjacency graph and the travel tidal flow graph. The third module performs weighted feature fusion of the static and dynamic features using temporal feature embedding. Based on the obtained node fusion features, a two-layer graph attention aggregation is performed on the geographic adjacency graph and the travel tidal flow graph to obtain spatial feature embeddings. The fourth module is used to learn the evolution of the grid's charging load over time using gated recurrent units based on the spatial feature embeddings to obtain the grid charging load prediction value. The fifth module is used to train the grid charging load prediction model using the prediction error of the effective grid as the loss function and the dataset. The sixth module is used to acquire real-time data of the target urban area and predict the real-time data using the trained grid charging load prediction model to obtain the predicted results of the electric vehicle charging pile utilization rate of the target area.

[0245] For example, with the cooperation of the first, second, third, fourth, fifth, and sixth modules in the device, the embodiment device can implement any of the aforementioned electric vehicle charging pile utilization prediction methods, namely, acquiring the original data of the target urban area, preprocessing the original data to obtain preprocessed data, wherein the preprocessing includes grid allocation based on service radius, static feature processing, and feature normalization, wherein the preprocessed data includes the grid, supply characteristics, static characteristics, and dynamic characteristics of the target urban area; based on the static and dynamic characteristics, the grid is constructed into nodes using a time feature embedding method, and the relationship edges of the nodes are constructed using geospatial proximity and traffic tidal flow directionality; based on the nodes and relationship edges, the geographical proximity is determined. The system integrates the geographic adjacency graph and the travel tidal flow graph. It employs temporal feature embedding for weighted feature fusion of static and dynamic features. Based on the obtained node fusion features, it performs two-layer graph attention aggregation on the geographic adjacency graph and the travel tidal flow graph to obtain spatial feature embedding. Using the spatial feature embedding, it employs gated recurrent units to learn the temporal evolution of the grid's charging load, obtaining the grid charging load prediction value. Based on the grid charging load prediction value, it uses the prediction error of the effective grid as the loss function and trains the model using a dataset to obtain the grid charging load prediction model. Finally, it acquires real-time data for the target urban area and uses the grid charging load prediction model to predict the real-time data, obtaining the predicted electric vehicle charging load for the target area. The beneficial effects of this invention are as follows: The area-weighted allocation method based on service radius enables refined allocation of load and supply characteristics, fully guaranteeing the physical consistency and high signal-to-noise ratio of grid-level ground truth data; By coupling a two-layer graph structure with geographic adjacency and dynamic travel tidal interactions, and utilizing a two-layer graph attention network to simultaneously aggregate local diffusion features and long-distance flow direction features, reliable fitting of interactive spatial dependencies is fully guaranteed; The dynamic and static feature gating fusion mechanism based on time embedding utilizes a gating network to dynamically generate weights according to the current time context, achieving decoupling and weighted fusion of static and dynamic features, effectively improving the model's environmental adaptability and prediction accuracy in different time periods (such as late night and peak hours).

[0246] This invention also provides an electronic device, which includes a processor and a memory;

[0247] The memory stores the program;

[0248] The processor executes a program to perform the aforementioned electric vehicle charging pile utilization prediction method; the electronic device has the function of carrying and running the software system for predicting the utilization rate of electric vehicle charging piles provided in the embodiments of the present invention, such as a personal computer, minicomputer, mainframe, workstation, network or distributed computing environment, standalone or integrated computer platform, or communicating with charged particle tools or other imaging devices, etc.

[0249] This invention also provides a computer-readable storage medium storing a program that is executed by a processor to implement the electric vehicle charging pile utilization prediction method described above.

[0250] In some alternative embodiments, the functions / operations mentioned in the block diagrams may not occur in the order shown in the operation diagrams. For example, depending on the functions / operations involved, two consecutively shown blocks may actually be executed substantially simultaneously, or the blocks may sometimes be executed in reverse order. Furthermore, the embodiments presented and described in the flowcharts of this invention are provided by way of example to provide a more comprehensive understanding of the technology. The disclosed methods are not limited to the operations and logic flows presented in the embodiments of this invention. Alternative embodiments are contemplated, in which the order of various operations is changed and sub-operations described as part of a larger operation are executed independently.

[0251] This invention also discloses a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device can read the computer instructions from the computer-readable storage medium and execute the computer instructions, causing the computer device to perform the aforementioned electric vehicle charging pile utilization prediction method.

[0252] Furthermore, although the invention has been described in the context of functional modules, it should be understood that, unless otherwise stated, one or more of the described functions and / or features may be integrated into a single physical device and / or software module, or one or more functions and / or features may be implemented in a separate physical device or software module. It is also understood that a detailed discussion of the actual implementation of each module is unnecessary for understanding the invention. Rather, considering the properties, functions, and internal relationships of the various functional modules in the apparatus disclosed in the embodiments of the invention, the actual implementation of the module will be understood within the scope of conventional skill of an engineer. Therefore, those skilled in the art can implement the invention as set forth in the claims using ordinary techniques without excessive experimentation. It is also understood that the specific concepts disclosed are merely illustrative and are not intended to limit the scope of the invention, which is determined by the full scope of the appended claims and their equivalents.

[0253] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0254] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can include, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0255] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0256] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0257] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0258] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

[0259] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A method for predicting the utilization rate of electric vehicle charging piles, characterized in that, include: The raw data of the target city area is obtained and preprocessed to obtain preprocessed data. The raw data includes historical charging station record data, basic information of charging stations, urban points of interest data, population raster data, and traffic OD trajectory data. The preprocessing includes grid allocation based on service radius, static feature processing, and feature normalization. The preprocessed data includes grids, supply features, static features, and dynamic features of the target urban area. Based on static and dynamic features, the grid is constructed into nodes using a time feature embedding method. Relationship edges between nodes are constructed using geospatial proximity and traffic tidal flow directionality. Based on the nodes and relationship edges, a geographic adjacency graph and a travel tidal flow direction graph are determined. The static and dynamic features are weighted and fused using a time feature embedding method. Based on the obtained node fusion features, a two-layer graph attention aggregation is performed on the geographic adjacency graph and the travel tidal flow graph to obtain the spatial feature embedding. Based on spatial feature embedding, gated recurrent units are used to learn the evolution of the grid's charging load over time, thereby obtaining the predicted value of the grid's charging load. Based on the predicted grid charging load, the prediction error of the effective grid is used as the loss function, and the model is trained using the dataset to obtain the grid charging load prediction model. Real-time data of the target city area is obtained, and the real-time data is predicted by the grid charging load prediction model to obtain the predicted power load of electric vehicle charging piles in the target area. The static and dynamic features are weighted and fused using a time feature embedding method. Based on the obtained node fusion features, a two-layer graph attention aggregation is performed on the geographic adjacency graph and the travel tidal flow graph to obtain spatial feature embedding, including: A learnable linear transformation matrix is ​​used to map static and dynamic features to a unified hidden layer dimension: in, This is the dimensional mapping result of dynamic features. As a dynamic feature, The linear transformation matrix is ​​a dynamic feature. For dynamic feature dimensions, For the hidden layer dimension, For dimensions; This is the dimensionality mapping result for static features. It is a static feature. The linear transformation matrix is ​​a static feature. For static feature dimensions; Using the temporal context vector as prior knowledge, a multilayer perceptron is used to dynamically calculate the feature weights of the dimension-aligned static and dynamic features, resulting in the following weights for the static and dynamic features: in, For a moment The weights of static features at time. For a moment The weights of the dynamic features at time, where and , It is the Sigmoid activation function. and For linear layers, It is the ReLU activation function; The node fusion features are obtained by element-wise weighted summation based on the weights of static and dynamic features: in, Represents a node At any moment The fusion characteristics of time, Represents a node static characteristics, node At any moment Dynamic characteristics of time; Based on the node fusion characteristics, a graph attention mechanism is used to obtain the attention coefficient between two nodes with related edges in the geographic adjacency graph and the travel tidal flow graph. Based on the attention coefficient, the neighbor features of the nodes are aggregated to obtain the geographic adjacency graph features and the travel tidal flow graph features. Geographic adjacency graph features and travel tidal flow graph features are fused through a linear layer to obtain spatial feature embeddings. for: in, Features of a geographic adjacency graph. Indicates a linear layer. This represents the characteristics of the travel tidal flow map.

2. The method for predicting the utilization rate of electric vehicle charging piles according to claim 1, characterized in that, The process of acquiring raw data for the target city area and preprocessing the raw data to obtain preprocessed data includes: After cleaning the data recorded at the charging stations, the data is aggregated by time to obtain the load sequence of each charging station, where the load sequence is a dynamic feature. The target urban area is divided into multiple grids. The allocation weight of the charging station is determined based on the intersection area between the buffer zone of the charging station and the grid. The unit grid load value is determined based on the allocation weight and the load sequence. The supply characteristics are determined based on the unit grid load value and the charging pile data of the charging station. The average population density is determined based on the population raster data in each grid, and the static characteristics of the grid are determined based on the urban interest point data, average population density, and supply characteristics in the grid. The traffic OD trajectory data is filtered and abnormal data is removed. Then, the number of vehicle flow directions between different grids in each time period is counted to obtain dynamic characteristics. In addition, the load sequence and static feature vector are normalized.

3. The method for predicting the utilization rate of electric vehicle charging piles according to claim 2, characterized in that, The process involves dividing the target urban area into multiple grids, determining the allocation weight of charging stations based on the intersection area between the charging station's buffer zone and the grid, determining the unit grid load value based on the allocation weight and load sequence, and determining supply characteristics based on the unit grid load value and charging station charging pile data, including: The target city area is divided into multiple square grids, and the buffer zone of the charging stations is determined based on the service radius of the charging stations in the target city area. Calculate the intersection area of ​​the buffer zone and the grid. Determine the allocation weights for: in, Indicates the first The grid and the first The intersecting area of ​​the charging stations Indicates the first The charging station for the first The weights assigned to each grid. The total number of grid cells. For grid number identification; Based on the assigned weights The load sequence is assigned to the grid to obtain the grid load value. for: in, express Time of the first The grid load value of each grid. For the first Each charging station time The load sequence; In addition, by allocating weights, the charging pile data of the charging station and the grid load value are distributed to the grid to obtain the supply characteristics, wherein the charging pile data includes the total number of piles, the number of fast charging piles and the number of slow charging piles.

4. The method for predicting the utilization rate of electric vehicle charging piles according to claim 2, characterized in that, The process involves constructing nodes from a grid using a time-feature embedding method based on static and dynamic features, constructing relational edges between nodes using geospatial proximity and traffic tidal flow directionality, and determining a geographic adjacency graph and a travel tidal flow direction graph based on the nodes and relational edges. This includes: Temporal context extraction is performed on static and dynamic features. The temporal context is mapped into dense vectors through a predefined embedding layer and then concatenated to obtain... The time context vector at a given moment for: in, For splicing operations, express A dense vector of time-hour information. for A dense vector of time and day information; Static features, dynamic features, and time context vectors are used as grid features. A relationship graph of relational edges and charging load is constructed based on the grid and grid features. The connection weights of two meshes are determined based on whether they share a common boundary or a common vertex in spatial geometry. According to connection weight By obtaining the geographical adjacency edges, a geographical adjacency graph can be obtained, in which... , and This represents two different grids; Based on the number of vehicles flowing in different directions during typical time periods, directed edges are established between two grids. The probability distribution of flow direction within the target city area is calculated based on the number of vehicles flowing in different directions, thus obtaining the tidal adjacency matrix. : in, Represents a grid With grid In the The weights of the tidal adjacency matrix for each time period. Indicates the first Grid for each time period To grid The number of vehicles flowing in the direction of traffic. Indicates the first Time period from grid Total number of departing vehicles The grid number is used as an identifier; the tidal connection edges are determined according to the weights of the tidal adjacency matrix, thus obtaining the travel tidal flow map.

5. The method for predicting the utilization rate of electric vehicle charging piles according to claim 1, characterized in that, The step of embedding based on spatial features and using gated recurrent units to learn the evolution of the grid's charging load over time to obtain the predicted grid charging load includes: Feature sequences embedded with spatial features Reset gate via the gated loop unit and Update Gate To forget and to retain: in, The current hidden state. This is the hidden state from the previous moment; In the candidate hidden state, The input weights are the candidate hidden states. The hidden weights of the candidate hidden states. The bias for the candidate hidden state; To reset the input weights of the gate, To reset the hidden weight of the door, To reset the door offset; To update the input weights of the gate, To update the hidden weight of the door, To update the door offset; For the Sigmoid function, It is the hyperbolic tangent function. This indicates element-wise multiplication; The output of the last time step of the gated loop unit is obtained and processed by the decoder of the gated loop unit to obtain the charging load prediction value: in, To control the output of the last time step of the loop unit.

6. The method for predicting the utilization rate of electric vehicle charging piles according to claim 5, characterized in that, The process of obtaining a grid charging load prediction model based on the grid charging load prediction value, using the prediction error of the effective grid as the loss function, and training with a dataset, includes: The mask mean square error is used as the loss function. Calculate the prediction error of the effective grid: in, This represents the actual load value. To predict load values, As a mask indicator variable, when the mesh When the memory is at a valid charging station or historical data Otherwise, it is 0; For L2 regularization terms, For hyperparameters; Based on loss function The model is trained using the Adam optimizer and a dataset, and the training is repeated until the model accuracy meets the preset requirements, thus obtaining the grid charging load prediction model.

7. A device for predicting the utilization rate of electric vehicle charging piles, characterized in that, include: The first module is used to acquire raw data of the target city area, preprocess the raw data to obtain preprocessed data, and the raw data includes historical charging station record data, basic information of charging stations, urban points of interest data, population raster data and traffic OD trajectory data. The preprocessing includes grid allocation based on service radius, static feature processing, and feature normalization. The preprocessed data includes grids, supply features, static features, and dynamic features of the target urban area. The second module is used to construct the grid into nodes based on static and dynamic features using a time feature embedding method, construct the relationship edges of the nodes using geospatial proximity and traffic tidal flow, and determine the geographic adjacency graph and travel tidal flow graph based on the nodes and relationship edges. The third module is used to perform feature weighting and fusion of static and dynamic features using time feature embedding. Based on the obtained node fusion features, a two-layer graph attention aggregation is performed on the geographic adjacency graph and the travel tidal flow graph to obtain spatial feature embedding. The fourth module is used to learn the evolution of the grid's charging load over time by embedding spatial features and using gated recurrent units to obtain the predicted value of the grid's charging load. The fifth module is used to obtain a trained grid charging load prediction model by using the prediction error of the effective grid as the loss function and training it with the dataset based on the grid charging load prediction value. The sixth module is used to acquire real-time data of the target urban area, and to predict the utilization rate of electric vehicle charging piles in the target area by using a trained grid charging load prediction model. The static and dynamic features are weighted and fused using a time feature embedding method. Based on the obtained node fusion features, a two-layer graph attention aggregation is performed on the geographic adjacency graph and the travel tidal flow graph to obtain spatial feature embedding, including: A learnable linear transformation matrix is ​​used to map static and dynamic features to a unified hidden layer dimension: in, This is the dimensional mapping result of dynamic features. As a dynamic feature, The linear transformation matrix is ​​a dynamic feature. For dynamic feature dimensions, For the hidden layer dimension, For dimensions; This is the dimensionality mapping result for static features. It is a static feature. The linear transformation matrix is ​​a static feature. For static feature dimensions; Using the temporal context vector as prior knowledge, a multilayer perceptron is used to dynamically calculate the feature weights of the dimension-aligned static and dynamic features, resulting in the following weights for the static and dynamic features: in, For a moment The weights of static features at time. For a moment The weights of the dynamic features at time, where and , It is the Sigmoid activation function. and For linear layers, It is the ReLU activation function; The node fusion features are obtained by element-wise weighted summation based on the weights of static and dynamic features: in, Represents a node At any moment The fusion characteristics of time, Represents a node static characteristics, node At any moment Dynamic characteristics of time; Based on the node fusion characteristics, a graph attention mechanism is used to obtain the attention coefficient between two nodes with related edges in the geographic adjacency graph and the travel tidal flow graph. Based on the attention coefficient, the neighbor features of the nodes are aggregated to obtain the geographic adjacency graph features and the travel tidal flow graph features. Geographic adjacency graph features and travel tidal flow graph features are fused through a linear layer to obtain spatial feature embeddings. for: in, Features of a geographic adjacency graph. Indicates a linear layer. This represents the characteristics of the travel tidal flow map.

8. An electronic device, characterized in that, Including the processor and memory; The memory is used to store programs; The processor executes the program to implement the electric vehicle charging pile utilization prediction method as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The storage medium stores a program that is executed by a processor to implement the electric vehicle charging pile utilization prediction method as described in any one of claims 1-6.

Images