Electric vehicle load gridding prediction method, device, equipment, medium and product

Abstract

The invention discloses an electric vehicle load gridding prediction method, device and equipment, a medium and a product, and the method comprises the steps: obtaining electric vehicle cluster charging station data of a target region, and constructing a charging station network diagram of the target region according to the charging station data; performing spatial feature extraction on the charging station network diagram, and splicing the spatial features and external environment vectors to obtain comprehensive feature representation of each charging station node; according to the comprehensive feature representation, modeling a load evolution process of each charging station node by using a graph diffusion denoising model, and generating a load prediction result of each charging station node; and mapping the load prediction result of each charging station node to the charging station network diagram, and generating load distribution of the space dimension and the time dimension of the target area. Load mapping and visualization from a station level to a distribution network level can be realized, and spatial correlation and load distribution randomness between electric vehicle charging stations can be effectively analyzed.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of vehicle charging, in particular to an electric vehicle load gridization prediction method, device, equipment, medium and product. BACKGROUND

[0002] With the rapid growth of the number of electric vehicles in cities, the concentrated charging behavior of electric vehicles has brought significant load impact on urban distribution networks, especially in peak periods, which easily causes local overload, voltage fluctuation and increased scheduling difficulty. In order to improve the operation safety and control accuracy of the distribution network, it is urgent to accurately predict the load evolution trend of the electric vehicle cluster in the city.

[0003] However, existing electric vehicle load prediction methods mostly model single stations, ignoring the spatial interaction between different charging stations, and are difficult to reflect the overall load characteristics of the city distribution network level. In addition, some methods only focus on deterministic prediction and fail to effectively represent the uncertainty factors existing in load changes, resulting in insufficient generalization ability of the model under complex conditions such as price fluctuations, weather changes, holidays, etc.

[0004] On the other hand, traditional time series prediction methods, such as long short-term memory network (LSTM) and recurrent neural network (RNN), often face problems such as high input dimension, weak spatial correlation, and difficult visualization output when dealing with multi-site joint prediction, making it difficult to meet the comprehensive prediction needs of city distribution networks for multi-source data fusion, spatial distribution modeling, and high-resolution load map output. SUMMARY

[0005] The present application provides an electric vehicle load gridization prediction method, device, equipment, medium and product, which is based on the fusion of spatial topology, supports multi-dimensional external condition input, and has uncertainty modeling and spatio-temporal visualization capabilities, providing technical support for regional power grid dispatching, load balancing analysis and emergency response.

[0006] To achieve the above purpose, the electric vehicle load gridization prediction method provided by the embodiments of the present application comprises:

[0007] Obtain electric vehicle cluster charging station data of a target area, and construct a charging station network graph of the target area according to the charging station data;

[0008] Extract spatial features from the charging station network graph, splice the spatial features and external environment vectors to obtain comprehensive feature representations of each charging station node;

[0009] According to the comprehensive feature representations, a graph diffusion denoising model is used to model the load evolution process of each charging station node, and load prediction results of each charging station node are generated;

[0010] mapping the load prediction result of each charging station node to the charging station network graph to generate the load distribution of the target area in the spatial and time dimensions.

[0011] As an improvement of the above scheme, the obtaining of the electric vehicle cluster charging station data of the target area, and the construction of the charging station network graph of the target area according to the charging station data, comprises:

[0012] obtaining electric vehicle cluster charging station data of a target area, and defining nodes and edges according to the charging station data to obtain a charging station weighted graph;

[0013] dividing the target area into a plurality of grids, mapping each charging station into the corresponding grid according to the charging station weighted graph to obtain a charging station network graph of the target area.

[0014] As an improvement of the above scheme, the spatial feature extraction of the charging station network graph, and the splicing of the spatial feature and the external environment vector to obtain the comprehensive feature representation of each charging station node, comprises:

[0015] obtaining the load curve of each charging station node and the connection relationship between the charging station nodes according to the charging station network graph;

[0016] training the load curve by using a graph neural network to obtain the node feature of each charging station node; and constructing an adjacency matrix between the charging station nodes according to the connection relationship between the charging station nodes;

[0017] combining the node feature and the adjacency matrix by using a graph convolution network to obtain the spatial feature of each charging station node;

[0018] splicing the spatial feature and the external environment vector to obtain the comprehensive feature representation of each charging station node.

[0019] As an improvement of the above scheme, the comprehensive feature representation is used to model the load evolution process of each charging station node by using a graph diffusion denoising model to generate the load prediction result of each charging station node, comprising:

[0020] using the comprehensive feature representation as the original input, gradually adding Gaussian noise according to a preset diffusion step number to convert the comprehensive feature representation into an input state in the training stage;

[0021] training the denoising network by using the input state to obtain the trained denoising network;

[0022] The trained denoising network is used for noise elimination of standard normal distribution noise, and the load evolution process of each charging station node is modeled by combining different comprehensive feature representations to generate the load prediction result of each charging station node.

[0023] As an improvement of the above scheme, the training of the denoising network by using the input state comprises:

[0024] The UNet with an encoder-decoder structure is used as the main architecture, a graph residual module is introduced in each convolution block, and a multi-head time attention mechanism is introduced in the time dimension to obtain the denoising network.

[0025] The loss function of the denoising network is defined, and the denoising network is trained by using the input state through the loss function to obtain the trained denoising network.

[0026] As an improvement of the above scheme, the mapping of the load prediction result of each charging station node to the charging station network graph to generate the load distribution in the space and time dimensions of the target area comprises:

[0027] According to the charging station network graph, determine the grid to which each charging station belongs.

[0028] The load prediction results of all charging station nodes in each grid are accumulated to obtain the grid-level load.

[0029] Each grid-level load is combined to generate the load distribution in the space and time dimensions of the target area.

[0030] In order to achieve the above purpose, the embodiment of the application provides an electric vehicle load grid prediction device, which comprises:

[0031] A charging data acquisition module is configured to acquire electric vehicle cluster charging station data of a target area, and construct a charging station network graph of the target area according to the charging station data.

[0032] A feature representation acquisition module is configured to perform spatial feature extraction on the charging station network graph, splice the spatial features and external environment vectors to obtain comprehensive feature representations of each charging station node.

[0033] A prediction result generation module is configured to model the load evolution process of each charging station node by using a graph diffusion denoising model according to the comprehensive feature representations, and generate the load prediction result of each charging station node.

[0034] A load distribution generation module is configured to map the load prediction result of each charging station node to the charging station network graph to generate the load distribution in the space and time dimensions of the target area.

[0035] To achieve the above object, the embodiment of the present application provides an electric vehicle load gridization prediction device, comprising a processor, a memory and a computer program stored in the memory and configured to be executed by the processor, and the processor implements the electric vehicle load gridization prediction method when executing the computer program.

[0036] To achieve the above object, the embodiment of the present application also provides a computer readable storage medium, comprising a stored computer program, wherein the computer readable storage medium controls the device where the computer readable storage medium is located to execute the electric vehicle load gridization prediction method when the computer program runs.

[0037] To achieve the above object, the embodiment of the present application also provides a computer program product stored in a storage medium, and the program product is executed by at least one processor to implement the steps of the electric vehicle load gridization prediction method.

[0038] Compared with the prior art, the electric vehicle load gridization prediction method, device, equipment, medium and product disclosed by the embodiment of the present application can obtain electric vehicle cluster charging station data of a target area, construct a charging station network graph of the target area according to the charging station data, perform spatial feature extraction on the charging station network graph, splice the spatial features and external environment vectors to obtain comprehensive feature representation of each charging station node, model the load evolution process of each charging station node by using a graph diffusion denoising model according to the comprehensive feature representation, generate a load prediction result of each charging station node, and map the load prediction result of each charging station node to the charging station network graph to generate a load distribution in the spatial and time dimensions of the target area. According to the charging station data and external environment data, the graph diffusion denoising model can be introduced to simulate a random disturbance process in the load evolution, restore a load curve conforming to external conditions from pure noise, generate a load prediction result with diversity and reality, aggregate the prediction result from the charging station level to the regional grid by constructing a station-grid mapping relationship, finally output a load distribution graph in time and space, realize load mapping and visualization from the station level to the distribution network level, effectively analyze the spatial correlation between electric vehicle charging stations and the randomness of load distribution, and be suitable for load management of a large-scale electric vehicle cluster, thereby providing technical support for regional power grid dispatching, load balance analysis and emergency response. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 is a flowchart of an electric vehicle load gridization prediction method provided by the embodiment of the present application;

[0040] Figure 2is a structural schematic diagram of an electric vehicle load gridization prediction device provided by an embodiment of the present application;

[0041] Figure 3 is a structural block diagram of an electric vehicle load gridization prediction device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0042] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative labor fall within the protection scope of the present application.

[0043] It should be noted that the terms "comprise" and "specific" and any variations of them in the present application are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device comprising a series of steps or units does not have to be limited to the clearly listed steps or units, but can include other steps or units that are not clearly listed or inherent to the process, method, product or device.

[0044] Please refer to Figure 1 , Figure 1 is a flowchart of an electric vehicle load gridization prediction method provided by an embodiment of the present application, and the electric vehicle load gridization prediction method comprises the following steps:

[0045] S1, obtaining electric vehicle cluster charging station data of a target area, and constructing a charging station network graph of the target area according to the charging station data;

[0046] S2, performing spatial feature extraction on the charging station network graph, splicing the spatial features and an external environment vector to obtain a comprehensive feature representation of each charging station node;

[0047] S3, modeling a load evolution process of each charging station node by using a graph diffusion denoising model according to the comprehensive feature representation, to generate a load prediction result of each charging station node;

[0048] S4, mapping the load prediction result of each charging station node to the charging station network graph to generate a load distribution in the spatial and time dimensions of the target area.

[0049] Illustratively, the geographic coordinates of all charging stations in the target area, the number of fast and slow charging piles, the maximum service capacity, and the historical charging and discharging power curve of the charging station data are collected; the time-of-use electricity price, temperature, rainfall and other external environmental data of the corresponding period are synchronously obtained; a charging station weighted topology graph is constructed according to a preset distance threshold, and a charging station to grid mapping relationship is established, and a charging station network graph of the target area is obtained. The load curve of each charging station node of the charging station network graph and the connection relationship between the charging station nodes are taken as the input of a graph convolution network (GCN) model, and the spatial features between the sites based on the physical location and the service are extracted by using the graph neural network; the spatial features and the external environmental vector such as the time-of-use electricity price and the meteorological condition are spliced and fused to form a comprehensive feature representation of each charging station under the current external condition. Taking the representation as an input condition, a graph diffusion denoising model is started, the load evolution process is modeled, the uncertainty of the future load is simulated by forward noise, the real trend is restored by multi-stage denoising to generate the load curve of the to-be-predicted period, and a representative load prediction result of each charging station under the same external condition is generated. The load prediction result of each charging station node is projected to the corresponding grid location according to the mapping relationship, and is superimposed and counted in the regional range, and finally a time and space continuous load thermal map of the regional distribution network level is output, and the charging load distribution characteristics of different regions are displayed. The embodiment of the present application can introduce the graph diffusion denoising model according to the charging station data and the external environmental data, simulate the random disturbance process in the load evolution, restore the load curve conforming to the external condition from the pure noise, generate the load prediction result with diversity and reality, aggregate the prediction result from the charging station level to the regional grid through the construction of the site-grid mapping relationship, finally output the time and space continuous load distribution map, realize the load mapping and visualization from the site level to the distribution network level, effectively analyze the spatial correlation between the electric vehicle charging stations and the randomness of the load distribution, and be suitable for the load management of large-scale electric vehicle clusters, and provide technical support for regional power grid dispatching, load balancing analysis and emergency response.

[0050] Specifically, the step S1 comprises:

[0051] S11, obtaining electric vehicle cluster charging station data of a target area, defining nodes and edges according to the charging station data, and obtaining a charging station weighted graph;

[0052] S12, dividing the target area into a plurality of grids, mapping each charging station into the corresponding grid according to the charging station weighted graph, and obtaining a charging station network graph of the target area.

[0053] For example, charging station information is aggregated within the regional power distribution network. First, a comprehensive survey of all public and dedicated charging stations is conducted, recording the latitude and longitude coordinates of each station, the number of fast and slow charging piles, the rated power of a single pile, the maximum number of vehicles that can be served simultaneously, and historical charging and discharging power curves. A weighted graph is established, with each charging station corresponding to a node in the graph. When the driving distance between two charging stations is less than a preset threshold, it is considered that there is a vehicle diversion relationship in actual operation, and an undirected edge is added between the nodes. Edge weights are set, with the physical meaning of edge weight being the substitutable service intensity. In this embodiment of the invention, the reciprocal of distance is used as the weight, with stronger interaction between neighboring stations and weights approaching zero for stations that are farther away. The regional latitude and longitude are divided into several grids, and the ground area covered by each grid can be flexibly adjusted according to the load density. Each charging station is mapped to the corresponding grid, obtaining the charging station network map of the target area. This facilitates subsequent projection of station-to-grid prediction results. By combining the power of fast charging and slow charging stations with their real-time utilization rates, the total charging power of a single station can be obtained at any given time. If vehicle-to-grid discharge is also supported, the discharge power is then deducted to obtain the net load curve. This curve can serve as both a historical sample and a benchmark for model evaluation.

[0054] In practice, the first step is to construct a weighted graph of the charging station network to describe the relationships between the charging stations. Assume there are n charging stations in the regional distribution network, denoted as C1, C2, ..., Cn. n The network graph is defined as G = (V, E, W) a V represents the set of nodes; E represents the set of edges in the network graph; W represents the set of edges in the network graph. a This is the weighting function.

[0055] For the node set V = {C1, C2, ..., C} n Let V be the set of all charging stations. For any node C... k It includes the following attributes: geographic coordinates are (x k ,y k The charge / discharge load at time t is expressed as L. k (t); The total number of charging piles in charging station k is represented by N. k N k,fast and N k,slow N represents the number of fast charging piles and slow charging piles within electric vehicle charging station k, respectively. k =N k,fast +N k,slow The service capacity of charging station k is S. k , used to represent node C k The maximum number of vehicles that a charging station can serve simultaneously.

[0056] Node C of the electric vehicle charging station k eigenvectors ω k The definition is as follows:

[0057] ω k = [x k , y k , L k (t), N k,fast , N k,slow , S k ], (1)

[0058] The edge set E represents the connection relationship between the electric vehicle charging stations, and is defined as follows:

[0059]

[0060] In the formula, i and j represent the node index of the electric vehicle charging station, and the edge set E represents that if the driving distance between two stations is short enough, they can be considered as directly connected on the regional road network.

[0061] Suppose the distance dist(C i ,C j ) between two electric vehicle charging stations is less than a certain threshold d th , then there is an edge (C i ,C j ) dist(C i ,C j ) is calculated as follows:

[0062]

[0063] In the formula, the weight W a (C i ,C j ) between two stations is the adjacency weight, which is equal to the inverse of the distance dist(C i ,C j ).

[0064] The adjacency matrix of the electric vehicle charging station network graph is denoted as A ij , and is represented as follows:

[0065]

[0066] The node feature matrix Ω stores the features of all charging stations, and is represented as follows:

[0067] Ω = [ω1, ω2, …, ω n ], (5)

[0068] In the formula, each column of Ω corresponds to the comprehensive attributes of a charging station in the whole day or in each batch.

[0069] To calculate the spatial distribution of the electric vehicle charging station load, the regional distribution network is divided into regular grids, and it is assumed that the geographical range of the region is a rectangular area [Xmin ,X max ]×[Y min ,Y max Divide it into m×n grids. Each grid is denoted as G. ij G ij The range is calculated as follows:

[0070] G ij =[X min +(i-1)Δx,X min +iΔx]×[Y min +(j-1)Δy,Y min +jΔy], (6)

[0071] In the formula, i and j are grid indices, and Δx and Δy are calculated as follows:

[0072]

[0073] For electric vehicle charging station C k The coordinates are (x k ,y k ), the index of its grid (i k ,j k The calculation method is as follows:

[0074] (8)

[0076] In the formula, This indicates rounding down, used to map charging station coordinates to their respective grid indices.

[0077] The load of each electric vehicle charging station can be determined by the charging and discharging loads, with the charging load L being the charging load. k,charge The calculation formula is as follows:

[0078] L k,charge (t)=U k (t)N k,fast P k,fast +U k (t)N k,slow P k,slow (9)

[0079] In the formula, U k (t) represents the utilization rate of charging piles in electric vehicle charging station k at time t; P k,fast and P k,slow These are the power ratings of the fast charging piles and slow charging piles at charging station k, respectively.

[0080] Total load L k The calculation method for (t) is as follows:

[0081] Lk (t) = L k,charge (t) = -L k,discharge (t), (10)

[0082] wherein, L k,charge (t) and L k,discharge (t) are the charging and discharging load values at time t respectively.

[0083] Each grid G ij Among them, the grid load L ij of the electric vehicle cluster is calculated as follows:

[0084] L ij (t) =∑ k w k L k (t), (11)

[0085] wherein, w k is the weight, reflecting the importance of the charging station, the weight of the fast charging station can be set to 1, the weight of the slow charging station is 0.8, and the weight of the hybrid station is 0.9.

[0086] Specifically, the step S2 comprises:

[0087] S21, obtaining the load curve of each charging station node and the connection relationship between the charging station nodes according to the charging station network diagram;

[0088] S22, training the load curve by using a graph neural network to obtain the node feature of each charging station node; constructing an adjacency matrix between the charging station nodes according to the connection relationship between the charging station nodes;

[0089] S23, combining the node feature and the adjacency matrix by using a graph convolution network to obtain the spatial feature of each charging station node;

[0090] S24, splicing the spatial feature and an external environment vector to obtain a comprehensive feature representation of each charging station node.

[0091] Exemplarily, the load curve of the past several days is packaged in batches-station-time three dimensions. For example, different scenes such as sunny weekdays, cloudy weekends, etc. are recorded as batches respectively to obtain diversified samples, and the packaged tensor form is convenient for parallel training and input into the graph neural network for processing. An adjacency matrix representing the connection relationship between charging stations is constructed, and the connection edges are determined according to geographical distance or traffic connectivity. The adjacency matrix is symmetrically normalized to make the information propagation weight between different nodes uniform and avoid bias. The node features of each charging station are arranged into an input vector, including load values, the number of charging piles and other static or dynamic information. A multi-layer network is set according to the graph convolution structure, the adjacency relationship and the node features are combined layer by layer, and the node representation is updated through a nonlinear function. The feature output by each layer is used to represent the integrated spatial feature of the station state and its adjacent station state. The vector output by the last layer is the spatial feature of the station. External data is encoded into a fixed-dimensional vector. For each station, the graph convolution output and the external factor vector are spliced to form a comprehensive feature representation for subsequent prediction model input.

[0092] In specific implementation, the total number of electric vehicle charging stations is obtained from the regional operation monitoring system and the third-party road network service provider, and weather data records are obtained. A historical load tensor X is constructed hist As follows:

[0093]

[0094] In the formula, the historical load tensor can be specifically recorded as X hist [b,i,t];X hist [b,i,t] is the load comprehensive feature of the i th station in the b th batch at the t th time period, representing the power of station i at time t in the b th batch, that is, X hist [b,i,t]=L i (t);B is the number of batches; Q is the total number of monitored charging stations; and T is the length of the historical window. is a real number set.

[0095] Each batch can correspond to different working days / weekends, sunny / rainy days, etc. X hist It is flattened into BxQ rows, that is, the complete time series load of each station is input as a sample.

[0096] Assume that τ ij represents the average driving time from station i to station j. The weight matrix W is initialized as The calculation is as follows:

[0097]

[0098] In the formula, λ is a distance attenuation parameter for controlling the exponential decay rate; tmax Maximum adjacency threshold.

[0099] Compute the normalized adjacency matrix A ij , the calculation formula is as follows:

[0100]

[0101] The GCN module is constructed, and a multi-layer graph convolution network is used to extract spatial features for peak charging queuing and shunting phenomena. The feature mapping is represented as follows:

[0102] H (0) = reshape(X hist ,BQ,d0), d0 = T, (15)

[0103] In the formula, H (0) is the initial feature mapping, the historical tensor is unfolded according to the charging station, and each row corresponds to the time sequence feature of a station; d0 is the initial feature dimension of GCN; reshape represents the tensor shape operation.

[0104] For GCN, the calculation method of multi-layer graph convolution is as follows:

[0105] H (l+1) = ReLU(AH (l) W (l) +b (l) )+H (l) , (16)

[0106] In the formula, l is the number of graph convolution layers; H (l) is the output feature of the lth layer; W (l) is the learnable parameter of the lth layer; ReLU is an activation function used to capture the nonlinear response when the load suddenly rises; b (l) is the bias vector of the lth layer; +H (l) The residual term retains the original feature to enhance the stability of the model.

[0107] The output spatial feature H (l) is represented as follows:

[0108]

[0109] In the formula, d l is the final feature dimension, and each row is a high-dimensional spatial feature of a station.

[0110] External condition encoding is performed, and the condition data tensor K is represented as follows:

[0111]

[0112] In the formula, F is the number of channels; p t is the electricity price at the tth period; Ttem,t is the ambient temperature; R t is the rainfall.

[0113] Specifically, the step S3 comprises:

[0114] S31, converting the comprehensive feature representation into an input state of a training phase by adding Gaussian noise step by step according to a preset diffusion step number, taking the comprehensive feature representation as an original input;

[0115] S32, training the denoising network by using the input state to obtain a trained denoising network;

[0116] S33, performing noise elimination on the standard normal distribution noise by using the trained denoising network, and modeling the load evolution process of each charging station node in combination with different comprehensive feature representations to generate a load prediction result of each charging station node.

[0117] More specifically, the step S32 comprises:

[0118] adopting a UNet with an encoder-decoder structure as a main architecture, introducing a graph residual module in each convolution block, and introducing a multi-head time attention mechanism in a time dimension to obtain the denoising network;

[0119] defining a loss function of the denoising network, training the denoising network by using the input state through the loss function to obtain the trained denoising network.

[0120] For example, the spatial feature sequence output by the graph convolutional network is used as the original input tensor. Gaussian noise is progressively added to the features at each time step according to a preset number of diffusion steps. Each diffusion operation is based on a pre-defined noise scheduling scheme, controlling the noise amplitude to gradually increase with each step. After all diffusion steps are completed, the original load features are transformed into a noise form approximating a standard normal distribution, serving as the final input state for the training phase. The UNet neural network with an encoder-decoder structure is used as the main architecture of the denoising model. A graph residual module is introduced into each convolutional block of the UNet to preserve the spatial structure information of the input nodes. A multi-head temporal attention mechanism is introduced in the time dimension to enhance the model's ability to capture load trends at different time points. External conditions at each time step (electricity price, weather, holiday markers, etc.) are encoded into vectors and injected into the network during the decoder stage to guide the generated results to conform to the actual external environment. Pure Gaussian noise generated by a standard normal distribution is used as the initial input. Following the trained denoising network structure, reverse diffusion operations are progressively performed to gradually eliminate noise and restore features. Each restoration step references graph structure information and current external conditions to ensure the output load curve conforms to the charging station network structure and its environment. A loss function is defined and training is optimized. The loss function includes: noise prediction error, used to measure the model's ability to recover from injected noise; reconstruction error, used to evaluate the difference between the final generated load curve and the actual load; and temporal smoothness constraint, used to control the volatility of the generated curve. The three types of losses are weighted and summed according to set weights to form the final training objective. During model training, the weights of these three types can be dynamically adjusted based on the actual data characteristics to obtain more reasonable load prediction results.

[0121] In practice, during the forward diffusion phase, the initial input tensor is... Consider it as raw data; Z t The domain of the noise tensor in the i-th step is the same as Z0; the total number of diffusion steps is defined as T. d In each diffusion step, t = 1, 2, 3, ..., T d Define the true forward diffusion distribution q(Z). t |Z t-1 ), which is represented as follows:

[0122]

[0123] In the formula, β t Noise scheduling parameters, typically starting from the minimum value β min Linear or cosine increasing to the maximum value β max To ensure that the end-point noise is close to the standard normal; I represents a normal distribution; I is the identity matrix whose dimension is the same as that of the tensor Z. t-1 The dimensions are consistent when flattened.

[0124] In the charging scenario of the electric vehicle cluster, the forward diffusion is equivalent to simulating the random fluctuations of the future load, such as the sudden surge of the charging peak or the random disturbance of the distributed discharge, to provide the learning basis for the subsequent denoising model.

[0125] In the denoising stage, the over-parameterized inverse distribution gradually purifies the noisy data, and the inverse distribution p θ of the model parameterization is defined as follows:

[0126]

[0127] In the formula, θ is the set of network parameters; μ θ is the mean function of network prediction; σ t is the learnable noise variance, which is used to control the randomness of the reverse sampling, and the network can use the UNet architecture combined with the graph convolution residual block to extract the spatial topological features; the time series attention layer is used to capture the time dependence of the load; e is the external condition encoding, which corresponds to the environmental information such as weather and electricity price; the UNet main structure is used for multi-scale feature extraction, combined with the graph convolution residual block to capture the charging station network topological information; and the denoised load distribution conforming to the current scene is restored in the denoising stage.

[0128] Reverse sampling initialization is represented as follows:

[0129]

[0130] In the formula, t = T d is sequentially sampled to t = 0, and the final denoised load prediction is obtained.

[0131] The loss function is designed as follows: the model training target is to minimize the noise prediction error L diff , the reconstruction error L rec and the smoothing regular L smooth , and the definitions of the loss functions are as follows:

[0132]

[0133]

[0134] In the formula, is the expected operator; ∈ is the true noise sample, which is used to simulate the forward process; ∈ θ is the noise tensor of network prediction, which is used to calculate the noise reconstruction error; is the final load prediction tensor obtained by reverse sampling; T' is the time step length defined in the forward diffusion, and u is the time step index; the smoothing loss avoids abrupt changes by comparing the predicted values at adjacent time points.

[0135] The overall loss function L is calculated as follows:

[0136] L = λ1L rec + λ2L smooth + λ3L diff , (25)

[0137] where λ1, λ2, λ3 are hyperparameters to balance the contribution of each part of the loss.

[0138] Specifically, the step S4 comprises:

[0139] S41, determining the grid to which each charging station belongs according to the charging station network diagram;

[0140] S42, accumulating the load prediction results of all charging station nodes in each grid to obtain a grid-level load;

[0141] S43, combining each grid-level load to generate a load distribution of the target area in space and time dimensions.

[0142] Illustratively, the area is divided into fixed-resolution two-dimensional grids, and is divided into a plurality of regular grid units according to latitude and longitude or area coordinates. Determine the grid to which each charging station belongs, which can be completed by comparing the geographical coordinates of the station with the grid boundary coordinates. Obtain the prediction result output by the model, i.e. the load sequence of each station at multiple future time steps, in the form of a two-dimensional structure of station x time. Traverse all stations, and according to the mapping table, accumulate the values of the prediction curve at the corresponding time of each station to the corresponding time point of the belonging grid. If multiple stations belong to the same grid, the total load of the grid at the time point is the accumulation of the load values of these stations. Generating a spatiotemporal load field combines the load values of each grid at each time step into a two-dimensional matrix of grid x time, which is used to construct a complete spatiotemporal load distribution. The final output form supports scenarios such as heat map display, time series dynamic diagram, geographic information system visualization, etc.

[0143] In specific implementation, for each batch b = 1, …, B; each node k = 1, …, n; and each time step t = 1, …, T', a load prediction tensor is generated is aggregated into a grid load, which is calculated as follows:

[0144]

[0145] where L b,m,t is the predicted load of batch b at grid m at time t; P m,k is the normalized mapping weight of grid m to node k; Z 0,b,k,t is the noise-free load prediction of batch b node k at time t obtained in step S3.

[0146] ​The electric vehicle load gridization prediction method disclosed by the embodiment of the application can obtain electric vehicle cluster charging station data of a target area, construct a charging station network graph of the target area according to the charging station data, perform spatial feature extraction on the charging station network graph, splice the spatial features and an external environment vector to obtain comprehensive feature representation of each charging station node, model a load evolution process of each charging station node by using a graph diffusion denoising model according to the comprehensive feature representation, generate a load prediction result of each charging station node, and map the load prediction result of each charging station node to the charging station network graph to generate a load distribution in the spatial and time dimensions of the target area. According to the charging station data and external environment data, the graph diffusion denoising model can be introduced to simulate a random disturbance process in the load evolution, restore a load curve that meets external conditions from pure noise, generate a load prediction result with diversity and reality, aggregate the prediction result from the charging station level to the regional grid by constructing a station-grid mapping relationship, and finally output a load distribution graph that is continuous in time and space, realize load mapping and visualization from the station level to the distribution network level, effectively analyze the spatial correlation between electric vehicle charging stations and the randomness of the load distribution, and be suitable for load management of a large-scale electric vehicle cluster, thereby providing technical support for regional power grid dispatching, load balance analysis and emergency response.

[0147] Referring to Figure 2 , Figure 2 is a structural schematic diagram of an electric vehicle load gridization prediction device 10 provided by the embodiment of the application. The electric vehicle load gridization prediction device 10 comprises:

[0148] A charging data acquisition module 11 is configured to obtain electric vehicle cluster charging station data of a target area, and construct a charging station network graph of the target area according to the charging station data.

[0149] A feature representation acquisition module 12 is configured to perform spatial feature extraction on the charging station network graph, splice the spatial features and an external environment vector, and obtain comprehensive feature representation of each charging station node.

[0150] A prediction result generation module 13 is configured to model a load evolution process of each charging station node by using a graph diffusion denoising model according to the comprehensive feature representation, and generate a load prediction result of each charging station node.

[0151] A load distribution generation module 14 is configured to map the load prediction result of each charging station node to the charging station network graph, and generate a load distribution in the spatial and time dimensions of the target area.

[0152] The electric vehicle load gridization prediction device 10 provided by the embodiment of the present application can realize all processes of the electric vehicle load gridization prediction method of the above embodiment, the functions of each module in the device and the technical effects achieved are respectively the same as the functions of the electric vehicle load gridization prediction method of the above embodiment and the technical effects achieved, and will not be repeated here.

[0153] Referring to Figure 3 , Figure 3 is a structural schematic diagram of an electric vehicle load gridization prediction device 20 provided by the embodiment of the present application. The electric vehicle load gridization prediction device 20 of the embodiment comprises a processor 21, a memory 22, and a computer program stored in the memory 22 and executable on the processor 21. The processor 21 implements the steps in the above electric vehicle load gridization prediction method embodiment when executing the computer program. Alternatively, the processor 21 implements the functions of each module in the above electric vehicle load gridization prediction device embodiment when executing the computer program.

[0154] For example, the computer program can be divided into one or more modules, which are stored in the memory 22 and executed by the processor 21 to complete the present application. The one or more modules can be a series of computer program instruction segments capable of completing a specific function, which are used to describe the execution process of the computer program in the electric vehicle load gridization prediction device 20.

[0155] The electric vehicle load gridization prediction device 20 can be a desktop computer, a notebook, a palm computer, and a cloud server, etc. The electric vehicle load gridization prediction device 20 can include, but is not limited to, the processor 21, the memory 22. Those skilled in the art can understand that the schematic diagram is only an example of the electric vehicle load gridization prediction device 20, and does not constitute a limitation on the electric vehicle load gridization prediction device 20, and can include more or fewer components than the diagram, or combine certain components, or different components, for example, the electric vehicle load gridization prediction device 20 can also include an input / output device, a network access device, a bus, etc.

[0156] The processor 21 can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic components, discrete hardware components, or the like. The general-purpose processor can be a microprocessor or can also be any conventional processor. The processor 21 is a control center of the electric vehicle load grid prediction device 20, and is connected to various parts of the electric vehicle load grid prediction device 20 through various interfaces and lines.

[0157] The memory 22 can be used to store computer programs and / or modules, and the processor 21 realizes various functions of the electric vehicle load grid prediction device 20 by running or executing the computer programs and / or modules stored in the memory 22, and calling data stored in the memory 22. The memory 22 can mainly include a program storage area and a data storage area. The program storage area can store an operating system, at least one application program required for a function (such as a sound playing function, an image playing function, etc.), and the like; and the data storage area can store data created according to use of the mobile phone (such as audio data, a phone book, etc.), and the like. In addition, the memory 22 can include a high-speed random access memory, and can also include a nonvolatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, at least one disk storage device, a flash memory device, or other volatile solid-state memory device.

[0158] The modules integrated in the electric vehicle load gridization prediction device 20 can be stored in a computer readable storage medium if they are implemented in the form of software function units and sold or used as independent products. Based on this understanding, all or part of the processes in the above-mentioned embodiment methods can also be completed by a computer program instructing related hardware, and the computer program can be stored in a computer readable storage medium. The computer program can be executed by the processor 21 to implement the steps of the above-mentioned various method embodiments. The computer program includes computer program code, which can be in the form of source code, object code, executable files or some intermediate forms, etc. The computer readable medium can include any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal and software distribution medium, etc. It should be noted that the content included in the computer readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction, for example, in some jurisdictions, according to legislation and patent practice, the computer readable medium does not include electrical carrier signals and telecommunication signals.

[0159] It should be noted that the above-described device embodiments are only schematic, and the units described as separate components can or can not be physically separated, and the components shown as units can or can not be physical units, that is, they can be located in one place or distributed on multiple network units. Part or all of the modules can be selected according to actual needs to achieve the purpose of the embodiment scheme. In addition, the connection relationship between the modules in the device embodiment provided by the present application indicates that there is a communication connection between them, which can be realized as one or more communication buses or signal lines. Those skilled in the art can understand and implement it without creative labor.

[0160] The embodiment of the present application also provides a computer readable storage medium, which includes a stored computer program, wherein the computer program controls the device where the computer readable storage medium is located to perform the electric vehicle load gridization prediction method of the above-mentioned embodiment when the computer program is running.

[0161] In addition, the embodiment of the present application also provides a computer program product stored in a storage medium, and the program product is executed by at least one processor to implement the steps of the electric vehicle load gridization prediction method of the above-mentioned embodiment.

[0162] The above is the preferred embodiment of the present application, it should be pointed out that, for those skilled in the art, without departing from the principles of the present application, can also make a number of improvements and refinements, these improvements and refinements are also considered to be within the scope of the present application.

Claims

1. An electric vehicle load gridding prediction method, characterized in that, The method comprises the following steps: obtaining electric vehicle cluster charging station data of a target area, and constructing a charging station network graph of the target area according to the charging station data; extracting spatial features of the charging station network graph, and splicing the spatial features and an external environment vector to obtain comprehensive feature representation of each charging station node; modeling a load evolution process of each charging station node by using a graph diffusion denoising model according to the comprehensive feature representation, and generating a load prediction result of each charging station node; mapping the load prediction result of each charging station node to the charging station network graph to generate a load distribution in the space and time dimensions of the target area.

2. The electric vehicle load gridding prediction method of claim 1, wherein, The method of obtaining electric vehicle cluster charging station data of a target area and constructing a charging station network graph of the target area according to the charging station data comprises: obtaining electric vehicle cluster charging station data of a target area, and defining nodes and edges according to the charging station data to obtain a charging station weighted graph; dividing the target area into a plurality of grids, and mapping each charging station to a corresponding grid according to the charging station weighted graph to obtain the charging station network graph of the target area.

3. The electric vehicle load gridding prediction method of claim 1, wherein, The method of extracting spatial features of the charging station network graph and splicing the spatial features and an external environment vector to obtain comprehensive feature representation of each charging station node comprises: obtaining a load curve of each charging station node and a connection relationship between charging station nodes according to the charging station network graph; training the load curve by using a graph neural network to obtain node features of each charging station node, and constructing an adjacency matrix between the charging station nodes according to the connection relationship between the charging station nodes; combining the node features and the adjacency matrix by using a graph convolution network to obtain spatial features of each charging station node; splicing the spatial features and the external environment vector to obtain comprehensive feature representation of each charging station node.

4. The electric vehicle load gridding prediction method of claim 1, wherein, The method of modeling a load evolution process of each charging station node by using a graph diffusion denoising model according to the comprehensive feature representation and generating a load prediction result of each charging station node comprises: taking the comprehensive feature representation as an original input, adding Gaussian noise step by step according to a preset diffusion step number, and converting the comprehensive feature representation into an input state in a training stage; training a denoising network by using the input state to obtain a trained denoising network; modeling the load evolution process of each charging station node by using the trained denoising network to eliminate noise of a standard normal distribution, and generating a load prediction result of each charging station node.

5. The electric vehicle load gridization prediction method of claim 4, wherein, The method of training a denoising network by using the input state to obtain a trained denoising network comprises: using a UNet with an encoder-decoder structure as a main architecture, introducing a graph residual module in each convolution block, and introducing a multi-head time attention mechanism in a time dimension to obtain a denoising network; defining a loss function of the denoising network, training the denoising network by using the input state through the loss function, and obtaining the trained denoising network.

6. The electric vehicle load gridding prediction method of claim 1, wherein, The load prediction result of each charging station node is mapped to the charging station network graph to generate the load distribution of the target area in the spatial and time dimensions, including: According to the charging station network graph, determine the grid to which each charging station belongs; The load prediction results of all charging station nodes in each grid are accumulated to obtain the grid-level load; Each grid-level load is combined to generate the load distribution of the target area in the spatial and time dimensions.

7. An electric vehicle load gridization prediction device characterized by comprising: It includes: The charging data acquisition module is used for acquiring electric vehicle cluster charging station data of a target area, and constructing a charging station network graph of the target area according to the charging station data; The feature representation acquisition module is used for extracting spatial features of the charging station network graph, splicing the spatial features and external environment vectors to obtain comprehensive feature representations of each charging station node; The prediction result generation module is used for modeling the load evolution process of each charging station node according to the comprehensive feature representations and using a graph diffusion denoising model to generate the load prediction result of each charging station node; The load distribution generation module is used for mapping the load prediction result of each charging station node to the charging station network graph to generate the load distribution of the target area in the spatial and time dimensions.

8. An electric vehicle load gridization prediction device characterized by comprising: The computer readable storage medium includes a computer program stored therein, wherein when the computer program runs, the device where the computer readable storage medium is located executes the electric vehicle load gridding prediction method as claimed in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer readable storage medium includes a computer program stored therein, wherein when the computer program runs, the device where the computer readable storage medium is located executes the electric vehicle load gridding prediction method as claimed in any one of claims 1-6.

10. A computer program product, characterised in that, The computer program product is stored in a storage medium, and the program product is executed by at least one processor to implement the steps of the electric vehicle load gridding prediction method as claimed in any one of claims 1-6.

Images