Load prediction method

Abstract

The invention provides a load prediction method. The method comprises the following steps: acquiring historical charging load data of the electric vehicle, and encrypting the historical charging load data to obtain encrypted data; performing feature extraction on the encrypted data to obtain a plurality of feature sequences, performing feature clustering on the plurality of feature sequences, and generating a dynamic graph structure according to a feature clustering result; and sending the dynamic graph structure to a client, so that the client fuses a feature sequence corresponding to the dynamic graph structure and a local feature extracted by the client to obtain a fused feature, performs load prediction according to the fused feature to obtain a load prediction result, and allocates power resources according to the load prediction result, the load prediction result shows the power consumption condition of the electric vehicle, and the cloud is in communication connection with the client. Through the load prediction method and device, the problem that a load prediction method in the prior art lacks privacy protection for users is solved.

Description

Technical Field

[0001] The present application relates to the technical field of load prediction of charging stations, and in particular to a load prediction method. Background Art

[0002] As global climate change and energy crisis issues have continued to attract widespread attention in recent years, the impact of traditional internal combustion engine vehicles on the environment has become increasingly prominent. These vehicles usually rely on fossil fuels such as oil, and they emit a large amount of harmful gases such as carbon dioxide and nitrogen oxides during operation, seriously polluting air quality and exacerbating environmental problems. Therefore, reducing carbon emissions during vehicle driving has become an urgent issue that needs to be addressed to mitigate the impact of climate change.

[0003] As an emerging energy technology, electric vehicles have quickly won the favor of the market with their low consumption of fossil energy, low pollution to the environment and positive social benefits. The scale of its industry has expanded rapidly and the number of electric vehicles has exploded. However, the incorporation of a large number of charging piles into the power grid has posed a huge challenge to the stability of power grid operation and the operating efficiency of electric vehicle charging stations. Specifically, it is reflected in: 1. The concentration of charging behavior may lead to an increase in the peak-to-valley difference in power grid load, which puts higher requirements on the dispatch of the power grid. The power grid needs to respond to load fluctuations more flexibly to maintain the stability of power supply. 2. The load volatility and spatiotemporal instability of electric vehicle charging and swapping stations increase the difficulty of power supply. Power grid equipment, such as transformers and transmission lines, will be under greater pressure when the load fluctuates greatly. If the charging piles are unevenly distributed or the charging time is too concentrated, it may cause voltage collapse in the local power grid, equipment damage, and even cause serious safety accidents such as systemic large-scale power outages. In order to ensure the overall operating efficiency, safety and stability of the power system, it is particularly important to accurately predict the load of charging and swapping stations. Therefore, it is necessary to predict the load of charging and swapping stations to ensure the overall operating efficiency, safety and stability of the power system, while improving the service efficiency and reliability of electric vehicle charging stations.

[0004] Traditional load forecasting models usually rely on a small number of electric vehicle charging records provided by a single charging station, and are mainly applicable to situations with simpler charging patterns. However, with the incorporation of a large number of charging piles into the power grid, the spatiotemporal instability of charging behavior and the high load volatility make it difficult for these traditional models to deeply explore users' charging patterns and accurately predict loads. Therefore, it is necessary to use the data of multiple charging stations for joint prediction, and use graph neural network technology to deeply explore the spatial dependency of load data to improve model accuracy. However, electric vehicle charging piles involve the core sensitive data of operators, which makes joint predictions face the risk of privacy leakage, because this process requires sharing load data between different power stations. In addition, as the number of power stations increases, cloud servers will be under tremendous communication and computing pressure. At present, research on load prediction of charging and swapping stations lacks consideration of the impact on user privacy protection and computing pressure.

[0005] Therefore, a load forecasting method that protects user privacy is needed. Summary of the invention

[0006] The main purpose of the present application is to provide a load forecasting method to at least solve the problem that the load forecasting method in the prior art lacks privacy protection for users.

[0007] In order to achieve the above-mentioned purpose, according to one aspect of the present application, a load forecasting method is provided, which is applied to the cloud, including: obtaining historical charging load data of electric vehicles, encrypting the historical charging load data to obtain encrypted data; extracting features from the encrypted data to obtain multiple feature sequences, performing feature clustering on the multiple feature sequences, and generating a dynamic graph structure according to the feature clustering results; sending the dynamic graph structure to a client, so that the client fuses the feature sequence corresponding to the dynamic graph structure with the local features extracted by the client to obtain fused features, and performs load forecasting according to the fused features to obtain load forecasting results, and allocating power resources according to the load forecasting results, wherein the load forecasting results indicate the power consumption of the electric vehicle, and the cloud and the client are communicatively connected.

[0008] Optionally, encrypting the historical charging load data to obtain encrypted data includes: dividing the historical charging load data into a plurality of time series of predetermined lengths, wherein the predetermined length is less than the length of the historical charging load data; and encrypting the historical charging load data by formula Encrypting a plurality of time series of the predetermined length to obtain the encrypted data, wherein: represents the encrypted data, X represents the historical charging load data, X=[[x0,...,x l-1 ],...,[x t-l+1 ,...,xt ]], l represents the predetermined length, t represents the length of the historical charging load data, W1 represents the weight matrix of the first fully connected layer, W2 represents the weight matrix of the second fully connected layer, b1 represents the first bias vector, and b2 represents the second bias vector.

[0009] Optionally, feature extraction is performed on the encrypted data to obtain multiple feature sequences, including: using a one-dimensional convolutional layer through the formula Extract features from the encrypted data to obtain multiple feature sequences, where (X i *w)[n] represents the convolution result of the i-th historical charging load data sequence and the convolution kernel w at position n, k represents the size of the convolution kernel, w[k] represents the weight of the convolution kernel at position k, X I [nk] represents the convolution result of the I-th historical charging load data sequence at position nk.

[0010] Optionally, feature clustering is performed on multiple feature sequences, including: normalizing the multiple feature sequences to obtain normalized feature sequences; randomly selecting multiple normalized feature sequences as centroids, calculating the distance from samples in each normalized feature sequence to the multiple centroids, and assigning the samples to the centroid with the smallest distance to obtain multiple clusters; calculating the mean of all samples corresponding to each cluster, and replacing the centroid with the mean to obtain a new centroid, and when the absolute value of the difference between the new centroid and the centroid is less than a preset threshold, stopping the iteration to obtain a feature clustering result.

[0011] Optionally, generating a dynamic graph structure according to the feature clustering result includes: generating a dynamic graph structure according to the feature clustering result by formula G i =(V i ,E i ) generates a dynamic graph structure, where G i represents the i-th dynamic graph structure, V i represents the i-th vertex of the dynamic graph structure, E i Represents the i-th edge of the dynamic graph structure.

[0012] According to another aspect of the present application, a load forecasting method is provided, which is applied to a client, including: obtaining historical charging load data of an electric vehicle, performing feature extraction on the historical charging load data, and obtaining local features; obtaining a dynamic graph structure sent by a cloud, and fusing a feature sequence corresponding to the dynamic graph structure with the local features to obtain a fused feature, wherein the dynamic graph structure is the cloud obtaining historical charging load data of the electric vehicle, encrypting the historical charging load data to obtain encrypted data, performing feature extraction on the encrypted data to obtain multiple feature sequences, and performing feature clustering on the multiple feature sequences, which are generated according to the feature sequences corresponding to the same feature; performing load forecasting according to the fused features to obtain a load forecasting result, and allocating power resources according to the load forecasting result, wherein the load forecasting result indicates the power consumption of the electric vehicle, and the cloud and the client are communicatively connected.

[0013] Optionally, fusing the feature sequence corresponding to the dynamic graph structure and the local feature to obtain a fused feature, including: normalizing the feature sequence corresponding to the dynamic graph structure and the local feature respectively to obtain a normalized feature sequence and a normalized local feature; and performing a formula on the normalized feature sequence and the normalized local feature. Fusion is performed to obtain the fusion feature, where G represents the fusion feature, represents the matrix concatenation operation, P * represents the normalized local feature, S * represents the normalized feature sequence.

[0014] Optionally, performing feature extraction on the historical charging load data to obtain local features includes: performing feature extraction on the historical charging load data through a one-dimensional convolution layer to obtain the local features.

[0015] Optionally, performing load forecasting according to the fusion feature to obtain a load forecast result includes: dividing the fusion feature into different scaling scales, and using the formula Graph convolution is performed on the fused features at each scale to capture spatial features, where H si is the hidden representation with the scaling factor si, R si is the final representation with the scaling factor si, represents the adjacency matrix, represents the degree matrix, W1 si represents the first weight matrix, represents the second weight matrix, represents the final weight matrix, σ() represents the ReLU activation function, p represents the Dropout ratio, BN() represents the batch normalization layer, Gsi represents the spatial feature with the scaling scale si, Indicates that G out Represents the sum of the spatial features of all the scaling scales.

[0016] Optionally, performing load forecasting according to the fusion feature to obtain a load forecasting result further includes: using an integrated LSTM method to determine the time feature of the fusion feature; according to the formula The spatial feature and the temporal feature are integrated to obtain the load forecast result, where P represents the load forecast result, W1 out represents the weight of the spatial feature, represents the weight of the time feature, h t represents the time feature, represents the historical charging load data sequence captured from the past n time steps, X represents the historical charging load data, W lo Represents the load weight.

[0017] By applying the technical solution of the present application, the historical charging load data of electric vehicles is obtained, the historical charging load data is encrypted to obtain encrypted data; the encrypted data is feature extracted to obtain multiple feature sequences, the multiple feature sequences are feature clustered, and a dynamic graph structure is generated according to the feature clustering results; the dynamic graph structure is sent to the client, so that the client fuses the feature sequence corresponding to the dynamic graph structure with the local features extracted by the client to obtain fused features, and performs load forecasting according to the fused features to obtain load forecasting results, and allocates power resources according to the load forecasting results. Compared with the prior art, in which the load forecasting method lacks user privacy protection, the present application can encrypt the load data, and then perform feature extraction and load forecasting, thereby improving the data security when performing load forecasting and ensuring the user's data security. Therefore, it can solve the problem of the lack of user privacy protection in the load forecasting method in the prior art, and achieve the effect of protecting user privacy while accurately predicting the load. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The drawings constituting part of the present application are used to provide a further understanding of the present application. The exemplary embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0019] Figure 1 A hardware structure block diagram of a mobile terminal for executing a load prediction method provided in an embodiment of the present application is shown;

[0020] Figure 2 A schematic diagram of a flow chart of a load forecasting method applied to a cloud provided in an embodiment of the present application is shown;

[0021] Figure 3 A schematic diagram of a flow chart of a load prediction method applied to a client provided in an embodiment of the present application is shown;

[0022] Figure 4 A schematic diagram of a flow chart of a specific load forecasting method provided in an embodiment of the present application is shown;

[0023] Figure 5 A schematic diagram of a two-layer framework for load forecasting based on cloud-edge collaboration and graph neural network provided by an embodiment of the present application is shown;

[0024] Figure 6 A schematic structural diagram of a MSTEM model provided in an embodiment of the present application is shown;

[0025] Figure 7 A structural block diagram of a load prediction device provided in an embodiment of the present application is shown;

[0026] Figure 8 A structural block diagram of another load forecasting device provided in an embodiment of the present application is shown.

[0027] The above drawings include the following reference numerals:

[0028] 102, processor; 104, memory; 106, transmission device; 108, input and output devices. DETAILED DESCRIPTION

[0029] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0030] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present application.

[0031] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequential order. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present application described here. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0032] For the convenience of description, some nouns or terms involved in the embodiments of the present application are explained below:

[0033] Cloud-Edge Collaboration: is a concept involving the combination of cloud computing and edge computing technologies. It provides a more efficient and flexible way to process data and run applications. The core idea of ​​cloud-edge collaboration is to transfer some computing tasks and data processing from the cloud (centralized data center) to the edge (close to the data source), thereby reducing latency, improving response speed, and reducing the burden on the central server.

[0034] Graph Neural Networks (GNN): A deep learning model for analyzing and processing graph structured data. Unlike traditional neural networks, GNNs can process irregular graph structured data, such as social networks, biological molecular structures, urban transportation networks, etc. GNNs have achieved remarkable success in many fields, including node classification, graph classification, link prediction, etc.

[0035] As introduced in the background technology, the existing research on load prediction of charging and swapping stations lacks consideration of the impact on user privacy protection and computing pressure. In order to solve the problem of lack of user privacy protection in load prediction methods, an embodiment of the present application provides a load prediction method.

[0036] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0037] The method embodiments provided in the embodiments of the present application can be executed in a mobile terminal, a computer terminal or a similar computing device. Taking running on a mobile terminal as an example, Figure 1 FIG. 1 is a hardware structure block diagram of a mobile terminal of a load forecasting method according to an embodiment of the present invention. Figure 1 As shown, the mobile terminal may include one or more ( Figure 1Only one is shown in the figure) a processor 102 (the processor 102 may include but is not limited to a processing device such as a microprocessor MCU or a programmable logic device FPGA) and a memory 104 for storing data, wherein the mobile terminal may also include a transmission device 106 and an input / output device 108 for communication functions. It can be understood by those skilled in the art that Figure 1 The structure shown is only for illustration and does not limit the structure of the mobile terminal. Figure 1 More or fewer components as shown, or with Figure 1 Different configurations shown.

[0038] The memory 104 can be used to store computer programs, for example, software programs and modules of application software, such as the computer program corresponding to the load forecasting method in the embodiment of the present invention. The processor 102 executes various functional applications and data processing by running the computer program stored in the memory 104, that is, the above method is implemented. The memory 104 may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some examples, the memory 104 may further include a memory remotely arranged relative to the processor 102, and these remote memories can be connected to the mobile terminal via a network. Examples of the above-mentioned network include but are not limited to the Internet, an intranet, a local area network, a mobile communication network, and a combination thereof. The transmission device 106 is used to receive or send data via a network. The specific example of the above-mentioned network may include a wireless network provided by a communication provider of the mobile terminal. In one example, the transmission device 106 includes a network adapter (Network Interface Controller, referred to as NIC), which can be connected to other network devices through a base station so as to communicate with the Internet. In one example, the transmission device 106 may be a radio frequency (RF) module, which is used to communicate with the Internet wirelessly.

[0039] In this embodiment, a load forecasting method running on a mobile terminal, a computer terminal or a similar computing device is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0040] Figure 2 is a flow chart of a load forecasting method according to an embodiment of the present application. Figure 2 As shown, applied to the cloud, the method includes the following steps:

[0041] Step S201, obtaining historical charging load data of the electric vehicle, and encrypting the historical charging load data to obtain encrypted data;

[0042] Specifically, a framework for cloud-edge collaborative load forecasting is established, including a cloud-based graph structure data generation module based on clustering (cloud center server, referred to as cloud), and a client prediction module based on graph neural network (client edge side, referred to as client). Taking the charging station as an example, the client uploads the historical load data of the local charging station, and the cloud establishes a privacy protection module, and uses the embedding layer to encrypt the data of each charging station to protect privacy. Including the cloud and the client, among which the cloud mainly establishes a cloud-based graph structure data generation module based on clustering, which includes a privacy protection module, a feature extraction module and a dynamic graph structure representation generation module, extracts the features of the load data and constructs a dynamic graph structure representation; then the corresponding dynamic graph structure representation is sent to the corresponding charging station client, and the client extracts the features of the local load data and aggregates it with the distribution features. Therefore, the cloud first obtains the historical charging load data of electric vehicles. For electric vehicle charging stations, the set of all charging records can be expressed as X={X1,X2,...,X N}, where N is the total number of EVCS (Electric Vehicle Charging Station, referred to as EVCS). For data from different clients, if the total length of their charging records is different, the number of fragments is kept consistent by truncating the sequence. After obtaining the load input matrix, the cloud encrypts and reconstructs it through the embedding layer. For data in the form of time series, the embedding layer will consist of fully connected layers.

[0043] Step S202, extracting features from the encrypted data to obtain multiple feature sequences, performing feature clustering on the multiple feature sequences, and generating a dynamic graph structure according to the feature clustering results;

[0044] Specifically, the prediction of the encrypted data corresponding to the above historical charging load data depends on capturing and identifying patterns in the data. After obtaining the embedded representation, the model uses a one-dimensional convolution layer to fully extract the features in the charging information. In the one-dimensional convolution layer, a sliding convolution operation along the time dimension is performed on the embedded representation sequence using a one-dimensional convolution kernel to extract features. When the convolution kernel slides on the input sequence, it calculates the dot product with the input data fragment at the current position and generates the output at the corresponding position, finally generating a feature sequence, and then generating a dynamic graph structure in the cloud based on the feature sequence.

[0045] Step S203, sending the above-mentioned dynamic graph structure to the client, so that the above-mentioned client fuses the above-mentioned feature sequence corresponding to the above-mentioned dynamic graph structure and the local feature extracted by the above-mentioned client to obtain the fused feature, and performs load forecasting based on the above-mentioned fused feature to obtain the load forecasting result, and allocates power resources according to the above-mentioned load forecasting result, wherein the above-mentioned load forecasting result indicates the power consumption of the above-mentioned electric vehicle, and the above-mentioned cloud and the above-mentioned client are communicatively connected.

[0046] Specifically, the feature aggregation module established by the client is then used to extract the local features of the local load data and fuse them with the dynamic graph structure sent by the cloud. The load forecast is further performed based on the fused features. The fusion of features obtained through multiple feature extractions can more accurately predict the power load, obtain the load forecast result, and allocate power resources based on the above load forecast results. Without sacrificing user privacy, it is possible to make full use of distributed data to improve prediction accuracy, reduce the computing and storage requirements for the central server, and reduce communication delays, thereby improving the response speed and efficiency of the overall system. In addition, this method can also effectively process non-graph data. It can deeply explore the user's charging mode by learning the dynamic graph structure representation of load data, provide additional spatiotemporal information and load characteristics for isolated EVCS, and provide charging station clients with personalized local models for more accurate predictions.

[0047] Through this embodiment, the historical charging load data of the electric vehicle is obtained, the historical charging load data is encrypted to obtain encrypted data; the encrypted data is feature extracted to obtain multiple feature sequences, the multiple feature sequences are feature clustered, and a dynamic graph structure is generated according to the feature clustering results; the dynamic graph structure is sent to the client, so that the client fuses the feature sequence corresponding to the dynamic graph structure with the local features extracted by the client to obtain fused features, and performs load forecasting according to the fused features to obtain load forecasting results, and allocates power resources according to the load forecasting results. Compared with the prior art, in which the load forecasting method lacks user privacy protection, the present application can encrypt the load data, and then perform feature extraction and load forecasting, thereby improving the security of the data when performing load forecasting and ensuring the data security of the user. Therefore, the problem of the lack of user privacy protection in the load forecasting method in the prior art can be solved, and the effect of protecting user privacy while accurately predicting the load can be achieved.

[0048] In the specific implementation process, the above step S201 encrypts the above historical charging load data to obtain encrypted data, which can be achieved by the following steps: dividing the above historical charging load data into a plurality of time series of predetermined lengths, wherein the above predetermined length is less than the length of the above historical charging load data; Encrypting a plurality of the above-mentioned time series of predetermined lengths to obtain the above-mentioned encrypted data, wherein: represents the above encrypted data, X represents the above historical charging load data, X=[[x0,...,x l-1 ],...,[x t-l+1 ,...,x t ]], l represents the predetermined length, t represents the length of the historical charging load data, W1 represents the weight matrix of the first fully connected layer, W2 represents the weight matrix of the second fully connected layer, b1 represents the first bias vector, and b2 represents the second bias vector. The method performs encryption through the above steps, which can protect the user's data security.

[0049] Specifically, for data in the form of time series, the embedding layer will consist of a fully connected layer. The embedding process is shown in the above formula. The load data of the charging station essentially constitutes a time series, represented by X i represents the historical load data of the ith power station, then X i =[x0,x1,...,x t ]∈R t , where x j represents the load of the i-th power station at the j-th time step. Each client first sends the local load data to the trusted cloud server, which aggregates it into a load matrix X = [X1, X2, ..., X N ], and convert the charging load data into a time series segment X with a predetermined length l i =[[x0,...,x l-1 ],...,[x t-l+1 ,...,x t ]] to match the input of the prediction model. For data from different clients, if the total length of their charging records is different, the number of fragments is kept consistent by truncating the sequence. After obtaining the load input matrix, the cloud encrypts and reconstructs it through the embedding layer. For data in the form of time series, the embedding layer will consist of a fully connected layer. The embedding process is as follows: shown.

[0050] In some optional implementations, the above step S202 extracts features from the above encrypted data to obtain multiple feature sequences, which can be achieved by the following steps: using a one-dimensional convolutional layer through the formula Feature extraction is performed on the encrypted data to obtain multiple feature sequences, among which (X i *w)[n] represents the convolution result of the i-th historical charging load data sequence and the convolution kernel w at position n, k represents the size of the convolution kernel, w[k] represents the weight of the convolution kernel at position k, X I[nk] represents the convolution result of the Ith historical charging load data sequence at position nk. The method performs feature extraction through the above steps, so that the features of the user's historical electricity consumption can be extracted, so that the future load can be better predicted.

[0051] Specifically, the feature extraction module on the cloud is constructed as follows: The prediction of time series data depends on capturing and identifying patterns in the data. After obtaining the embedded representation, the model uses a one-dimensional convolution layer to fully extract the features in the charging information. In the one-dimensional convolution layer, a sliding convolution operation along the time dimension is performed on the embedded representation sequence using a one-dimensional convolution kernel to extract features. When the convolution kernel slides on the input sequence, it calculates the dot product with the input data fragment at the current position and generates the output at the corresponding position, finally generating a feature sequence. The expression of the entire calculation process is In order to ensure that the length of the input and output sequences is consistent, a padding strategy is used to control the size of the output feature map, and the padding range is half of the convolution kernel size.

[0052] In some optional embodiments, the above step S202 performs feature clustering on the multiple feature sequences, which can be achieved by the following steps: normalizing the multiple feature sequences to obtain normalized feature sequences; randomly selecting multiple normalized feature sequences as centroids, calculating the distances from the samples in each normalized feature sequence to the multiple centroids, assigning the samples to the centroids with the smallest distances, and obtaining multiple clusters; calculating the mean of all samples corresponding to each cluster, and replacing the centroids with the mean to obtain new centroids, and stopping iteration when the absolute value of the difference between the new centroid and the centroid is less than a preset threshold, to obtain feature clustering results. The method clusters similar features into one category through the above steps, so that the features can be classified for easy identification and distinction of features.

[0053] In the specific implementation process, the K-Means clustering algorithm is used to cluster the features, and the features of the same category are grouped into corresponding feature graphs. Given a feature sequence set X = {X1, X2, ..., X n}, where X i =[x n ,x n+1 ,...,x n+τ-1 ] represents the feature segment of length τ of the ith power station from time n. In order to reduce the impact of data scale differences on clustering effects, the input is first normalized: in, is the normalized feature sequence, μ and σ are the mean and standard deviation of all sample features respectively. After normalization, k feature sequences v are randomly selected i(i=1,2,...,k) is used as the initial centroid, and k is the final number of clusters, which is determined by the elbow method. Then, for each sample, the distance from it to each centroid is calculated and assigned to the nearest centroid to form K clusters. Using the Euclidean distance metric, the expression is: d(X i -v i )=||X i -v i ||2, for each category, calculate the mean of all samples assigned to it, and use the mean to update the centroid. Compare the new and old centroids. When the change in the centroid is less than the preset threshold or the change is very small after multiple consecutive iterations, the algorithm is considered to have converged. At this time, the iteration is stopped and the final clustering result is obtained. The feature graph is constructed based on the final clustering result. The feature graph structure corresponding to the i-th category is represented as G i (V i ,E i ), where V i Represents the vertex set of the graph, which consists of all feature sequences in the category, E i The edge set of the graph is composed of the edges connecting every two vertices. Finally, the cloud distributes the corresponding feature map data of each client according to its category, and further predicts it after aggregating it with the local client data. The loss function of the entire cloud model consists of two parts: reconstruction error and clustering loss: L = L c +L r , where L c is the reconstruction error. In this study, the mean square error (MSE) loss is used as the network reconstruction loss to measure the expressive power of the embedded representation on the original data. The formula is: Where Y i is the ith input sequence, is the sequence reconstructed by the network, and n is the total number of sequence samples. r is the clustering loss, using the Davis-Boulding Index (DBI) as the clustering loss, and the calculation formula is as follows: M ij =||c i -c j ||2, where k is the number of categories, R ij is the similarity between the i-th category and the j-th category, S i is the dispersion of each time series in the i-th category, n i is the total number of samples in the ith class, c i is the cluster center of the ith class, M ij is the distance between the i-th and j-th cluster centers.

[0054] In some other optional implementations, the above step S202 generates a dynamic graph structure according to the feature clustering result, which can be achieved by the following steps: According to the feature clustering result, the dynamic graph structure is generated by formula G i =(V i ,E i ) generates a dynamic graph structure, where G i represents the i-th dynamic graph structure, V i represents the i-th vertex of the above dynamic graph structure, E i Represents the i-th edge of the above dynamic graph structure. This method generates a dynamic graph structure through the above steps, which can reduce the prediction deviation caused by abnormal or incomplete local data by aggregating feature data, help the model better understand the relationship between features, and enhance the prediction accuracy and robustness of the model.

[0055] Specifically, a feature graph is constructed based on the final clustering result. The feature graph structure corresponding to the i-th category is represented as G i =(V i ,E i ), the vertices of the graph consist of all feature sequences within the category, and the edge set of the graph consists of the edges connecting every two vertices. Finally, the cloud distributes the corresponding feature graph data of each client according to its category, and further predicts it after aggregating it with the local client data.

[0056] Figure 3 is a flow chart of a load forecasting method according to an embodiment of the present application, which is applied to a client, such as Figure 3 As shown, the method comprises the following steps:

[0057] Step S204: acquiring historical charging load data of the electric vehicle, performing feature extraction on the historical charging load data, and obtaining local features;

[0058] Specifically, after the cloud extracts and clusters the features, the client first uses a one-dimensional convolutional layer to process the client's original load data X and extract local features P.

[0059] Step S205: obtaining a dynamic graph structure sent by the cloud, fusing the feature sequence corresponding to the dynamic graph structure with the local feature to obtain a fusion feature, wherein the dynamic graph structure is obtained by the cloud to obtain historical charging load data of the electric vehicle, encrypting the historical charging load data to obtain encrypted data, performing feature extraction on the encrypted data to obtain multiple feature sequences, and performing feature clustering on the multiple feature sequences, which are generated according to the feature sequences corresponding to the same feature;

[0060] Specifically, after normalization, local features and cloud features are integrated to obtain fused features.

[0061] Step S206: Perform load forecasting based on the above fusion features to obtain load forecasting results, and allocate power resources based on the above load forecasting results, wherein the above load forecasting results indicate the power consumption of the above electric vehicle, and the above cloud and the above client are communicated and connected.

[0062] Specifically, after aggregating the features, the client introduces the MSTEM model to predict the load of the local charging station and obtains the load prediction result.

[0063] In order to perform feature fusion, the above step S205 of the present application can be implemented by the following steps: Optionally, the feature sequence corresponding to the above dynamic graph structure and the above local feature are fused to obtain a fused feature, including: normalizing the feature sequence corresponding to the above dynamic graph structure and the above local feature respectively to obtain a normalized feature sequence and a normalized local feature; and performing the formula on the above normalized feature sequence and the above normalized local feature. Fusion is performed to obtain the above fusion features, where G represents the above fusion features. represents the matrix concatenation operation, P * represents the above normalized local features, S * Represents the above normalized feature sequence. Through the above steps, this method reduces the prediction deviation caused by abnormal or incomplete local data, helps the model better understand the relationship between features, and enhances the prediction accuracy and robustness of the model.

[0064] Specifically, since the scales of the extracted local features and cloud features are different, they cannot be directly used as input and need to be normalized first: In the formula, P represents the local features extracted by the client, S is the graph structure feature data sent from the cloud, and P * and S * are normalized feature data, μ is the mean of local and cloud feature samples, and σ is the standard deviation. After normalization, local and cloud features are integrated to obtain the final graph structure input. The aggregation expression is as follows:

[0065] In some optional implementations, the above step S204 extracts features from the above historical charging load data to obtain local features, which can be achieved by the following steps: extracting features from the above historical charging load data through a one-dimensional convolution layer to obtain the above local features. The method extracts local features through the above steps, so that feature fusion can be performed.

[0066] In the specific implementation process, a one-dimensional convolutional layer is used to process the client's original load data X and extract the local features P.

[0067] In some optional implementations, the above step S206 performs load forecasting according to the above fusion features to obtain a load forecasting result, which can be achieved by the following steps: dividing the above fusion features into different scaling scales, and using the formula Graph convolution is performed on the above fused features at each of the above scales to capture spatial features, where H si is the hidden representation of the above scaled to si, R si is the final representation of the above scaled to si, represents the adjacency matrix, represents the degree matrix, W1 si represents the first weight matrix, represents the second weight matrix, represents the final weight matrix, σ() represents the ReLU activation function, p represents the Dropout ratio, BN() represents the batch normalization layer, G si represents the spatial features with the above scaling scale si, Indicates that G out Represents the sum of all the spatial features of the above scaling scales. This method extracts spatial features through the above steps, so that features can be extracted from different dimensions to obtain more comprehensive and accurate features.

[0068] In the specific implementation process, after aggregating features, the client introduces the MSTEM model to predict the load of the local charging station. In the MSTEM model, the load data X is first divided into different scales, and a multi-scale graph learning module is used to perform graph convolution on the input of each scale to capture spatial features. Then, a time-enhanced neural network module with integrated LSTM units is used to enhance the model's ability to capture long-term temporal dependencies. This process can be expressed as:

[0069] h t =o t *tanh(c t ), where f t 、i t and t are the outputs of the forget gate, input gate, and output gate respectively. and c t are candidate and updated cell states, h t is the updated hidden state, W l and b l Represents weights and bias terms.

[0070] In some optional implementations, the step S206 performs load forecasting according to the fusion features to obtain load forecasting results, and further includes the following steps: using an integrated LSTM method to determine the time characteristics of the fusion features; according to the formula and The above spatial features are combined with the above temporal features to obtain the above load forecasting result, where P represents the above load forecasting result, W1 out represents the weight of the above spatial features, represents the weight of the above time features, h t Representing the above time characteristics, represents the historical charging load data sequence captured from the past η time steps, X represents the above historical charging load data, W lo The method performs load forecasting through the above steps to predict the future power load and allocate power resources according to the power load.

[0071] Specifically, the outputs of the above modules are finally integrated to generate the prediction results of the local charging station load. The above process expression is as follows: in is the historical load sequence captured from the past η time steps, W lo and W out are all weight matrices. P is the final load forecast result. In order to fit the actual situation, the model will set the low forecast value in P to zero according to the predefined threshold.

[0072] In order to enable those skilled in the art to more clearly understand the technical solution of the present application, the implementation process of the load forecasting method of the present application will be described in detail below in conjunction with specific embodiments.

[0073] This embodiment relates to a specific load forecasting method, such as Figure 4 As shown, the following steps are included:

[0074] Step S1: Start;

[0075] Step S2: The client uploads data (historical charging load data of the local charging station), including the charging start and end time and total energy consumption of each charging event; the cloud aggregates all client data into a load matrix and converts it into a fixed-length time series segment;

[0076] Step S3: A cloud-based graph structure data generation module based on clustering is established in the cloud, including a privacy protection module, a feature extraction module and a dynamic graph structure representation generation module, to extract the features of the load data and construct a dynamic graph structure representation;

[0077] Step S4: (the cloud sends the corresponding dynamic graph structure representation to the corresponding charging station client, and the client extracts the features of the local load data and aggregates it with the distribution features); the client establishes a client prediction module based on the graph neural network, obtains the aggregated graph structure data, and trains the local graph neural network-based model for load prediction, including a feature aggregation module and a local MSTEM prediction model training (module);

[0078] Step S5: Determine whether the model has reached the convergence condition. If yes, output the final prediction result. If no, continue to execute step S4.

[0079] Figure 5 A schematic diagram of a two-layer framework for load forecasting based on cloud-edge collaboration and graph neural networks is shown. The historical charging load data of electric vehicles is uploaded, embedded through the embedding layer, convolved through the one-dimensional convolution layer to obtain the feature vector, and further kmeans clustering is performed to obtain category A features, category B features, and category C features, which are then distributed. At the same time, the collected original data is feature extracted through a one-dimensional convolution layer to obtain local features, and the local features are aggregated with the features distributed by the cloud server through the aggregation layer, input into the prediction model for prediction, and the final prediction result is obtained.

[0080] Figure 6 The schematic diagram of the structure of a MSTEM model is shown. X represents the acquisition of historical charging load data of electric vehicles, and the multi-scale graph is constructed to obtain X s1 , X st Until X sk , respectively through GCN, GCN layer for graph learning, get G s1 , G st Until G sk , processed by multi-layer perceptron MLP; at the same time, the historical charging load data of electric vehicles is obtained through the time-enhanced neural network LSTM to obtain h t , further processed by the multi-layer perceptron MLP, the obtained element set is passed through the output control mechanism of the output layer P to obtain the prediction result Y.

[0081] The embodiments of the present application also provide a load forecasting device. It should be noted that the load forecasting device of the embodiments of the present application can be used to execute the load forecasting method provided by the embodiments of the present application. The device is used to implement the above-mentioned embodiments and preferred implementation modes, and the descriptions that have been made will not be repeated. As used below, the term "module" can implement a combination of software and / or hardware for a predetermined function. Although the devices described in the following embodiments are preferably implemented in software, the implementation of hardware, or a combination of software and hardware, is also possible and conceivable.

[0082] The load forecasting device provided in the embodiments of the present application is introduced below.

[0083] Figure 7 Schematic diagram of a load forecasting device according to an embodiment of the present application. Figure 7 As shown, the device comprises:

[0084] The encryption unit 10 is used to obtain the historical charging load data of the electric vehicle and encrypt the historical charging load data to obtain encrypted data;

[0085] Specifically, a framework for cloud-edge collaborative load forecasting is established, including a cloud-based graph structure data generation module based on clustering (cloud center server, referred to as cloud), and a client prediction module based on graph neural network (client edge side, referred to as client). Taking the charging station as an example, the client uploads the historical load data of the local charging station, and the cloud establishes a privacy protection module, and uses the embedding layer to encrypt the data of each charging station to protect privacy. Including the cloud and the client, among which the cloud mainly establishes a cloud-based graph structure data generation module based on clustering, which includes a privacy protection module, a feature extraction module and a dynamic graph structure representation generation module, extracts the features of the load data and constructs a dynamic graph structure representation; then the corresponding dynamic graph structure representation is sent to the corresponding charging station client, and the client extracts the features of the local load data and aggregates it with the distribution features. Therefore, the cloud first obtains the historical charging load data of electric vehicles. For electric vehicle charging stations, the set of all charging records can be expressed as X={X1,X2,...,X N}, where N is the total number of EVCS (Electric Vehicle Charging Station, referred to as EVCS). For data from different clients, if the total length of their charging records is different, the number of fragments is kept consistent by truncating the sequence. After obtaining the load input matrix, the cloud encrypts and reconstructs it through the embedding layer. For data in the form of time series, the embedding layer will consist of fully connected layers.

[0086] A generating unit 20 is used to extract features from the encrypted data to obtain multiple feature sequences, perform feature clustering on the multiple feature sequences, and generate a dynamic graph structure according to the feature clustering results;

[0087] Specifically, the prediction of the encrypted data corresponding to the above historical charging load data depends on capturing and identifying patterns in the data. After obtaining the embedded representation, the model uses a one-dimensional convolution layer to fully extract the features in the charging information. In the one-dimensional convolution layer, a sliding convolution operation along the time dimension is performed on the embedded representation sequence using a one-dimensional convolution kernel to extract features. When the convolution kernel slides on the input sequence, it calculates the dot product with the input data fragment at the current position and generates the output at the corresponding position, finally generating a feature sequence, and then generating a dynamic graph structure in the cloud based on the feature sequence.

[0088] The first prediction unit 30 is used to send the above-mentioned dynamic graph structure to the client, so that the above-mentioned client fuses the above-mentioned feature sequence corresponding to the above-mentioned dynamic graph structure and the local feature extracted by the above-mentioned client to obtain a fused feature, and performs load prediction based on the above-mentioned fused feature to obtain a load prediction result, and allocates power resources according to the above-mentioned load prediction result, wherein the above-mentioned load prediction result indicates the power consumption of the above-mentioned electric vehicle, and the above-mentioned cloud and the above-mentioned client are communicatively connected.

[0089] Specifically, the feature aggregation module established by the client is then used to extract the local features of the local load data and fuse them with the dynamic graph structure sent by the cloud. The load forecast is further performed based on the fused features. The fusion of features obtained through multiple feature extractions can more accurately predict the power load, obtain the load forecast result, and allocate power resources based on the above load forecast results. Without sacrificing user privacy, it is possible to make full use of distributed data to improve prediction accuracy, reduce the computing and storage requirements for the central server, and reduce communication delays, thereby improving the response speed and efficiency of the overall system. In addition, the device can also effectively process non-graph data. It can deeply explore the user's charging mode by learning the dynamic graph structure representation of load data, provide additional spatiotemporal information and load characteristics for isolated EVCS, and provide charging station clients with personalized local models for more accurate predictions.

[0090] Through this embodiment, the historical charging load data of electric vehicles is obtained, the historical charging load data is encrypted to obtain encrypted data; the encrypted data is feature extracted to obtain multiple feature sequences, the multiple feature sequences are feature clustered, and a dynamic graph structure is generated according to the feature clustering results; the dynamic graph structure is sent to the client, so that the client fuses the feature sequence corresponding to the dynamic graph structure with the local features extracted by the client to obtain fused features, and performs load forecasting according to the fused features to obtain load forecasting results, and allocates power resources according to the load forecasting results. Compared with the prior art, in which the load forecasting device lacks user privacy protection, the present application can encrypt the load data, and then perform feature extraction and load forecasting, thereby improving the security of data when performing load forecasting and ensuring the data security of users. Therefore, it is possible to solve the problem of the lack of user privacy protection in the load forecasting device in the prior art, and achieve the effect of protecting user privacy while accurately predicting the load.

[0091] In the specific implementation process, the encryption unit includes a division module and an encryption module. The division module is used to divide the historical charging load data into a plurality of time series of predetermined lengths, wherein the predetermined length is less than the length of the historical charging load data; the encryption module is used to calculate the historical charging load data by formula Encrypting a plurality of the above-mentioned time series of predetermined lengths to obtain the above-mentioned encrypted data, wherein: represents the above encrypted data, X represents the above historical charging load data, X=[[x0,...,x l-1 ],...,[x t-l+1 ,...,x t ]], l represents the predetermined length, t represents the length of the historical charging load data, W1 represents the weight matrix of the first fully connected layer, W2 represents the weight matrix of the second fully connected layer, b1 represents the first bias vector, and b2 represents the second bias vector. The device is encrypted through the above steps, which can protect the user's data security.

[0092] Specifically, for data in the form of time series, the embedding layer will consist of a fully connected layer. The embedding process is shown in the above formula. The load data of the charging station essentially constitutes a time series, represented by X i represents the historical load data of the ith power station, then X i =[x0,x1,...,x t ]∈R t , where x j represents the load of the i-th power station at the j-th time step. Each client first sends the local load data to the trusted cloud server, which aggregates it into a load matrix X = [X1, X2, ..., X N], and convert the charging load data into a time series segment X with a predetermined length l i =[[x0,...,x l-1 ],...,[x t-l+1 ,...,x t ]] to match the input of the prediction model. For data from different clients, if the total length of their charging records is different, the number of fragments is kept consistent by truncating the sequence. After obtaining the load input matrix, the cloud encrypts and reconstructs it through the embedding layer. For data in the form of time series, the embedding layer will consist of a fully connected layer. The embedding process is as follows: shown.

[0093] In some optional embodiments, the generation unit includes a feature extraction module for using a one-dimensional convolutional layer through the formula Feature extraction is performed on the encrypted data to obtain multiple feature sequences, among which (X i *w)[n] represents the convolution result of the i-th historical charging load data sequence and the convolution kernel w at position n, k represents the size of the convolution kernel, w[k] represents the weight of the convolution kernel at position k, X I [nk] represents the convolution result of the Ith historical charging load data sequence at position nk. The device performs feature extraction through the above steps, so that the features of the user's historical electricity consumption can be extracted, so as to better predict the future load.

[0094] Specifically, the feature extraction module on the cloud is constructed as follows: The prediction of time series data depends on capturing and identifying patterns in the data. After obtaining the embedded representation, the model uses a one-dimensional convolution layer to fully extract the features in the charging information. In the one-dimensional convolution layer, a sliding convolution operation along the time dimension is performed on the embedded representation sequence using a one-dimensional convolution kernel to extract features. When the convolution kernel slides on the input sequence, it calculates the dot product with the input data fragment at the current position and generates the output at the corresponding position, finally generating a feature sequence. The expression of the entire calculation process is In order to ensure that the length of the input and output sequences is consistent, a padding strategy is used to control the size of the output feature map, and the padding range is half of the convolution kernel size.

[0095] In some optional embodiments, the generation unit further includes a normalization processing module, an allocation module and a replacement module. The normalization processing module is used to normalize the plurality of feature sequences to obtain normalized feature sequences; the allocation module is used to randomly select a plurality of normalized feature sequences as centroids, calculate the distance from each sample in the normalized feature sequence to the plurality of centroids, and allocate the samples to the centroids with the smallest distance to obtain a plurality of clusters; the replacement module is used to calculate the mean of all samples corresponding to each cluster, and replace the centroids with the mean to obtain new centroids. When the absolute value of the difference between the new centroid and the centroid is less than a preset threshold, the iteration is stopped to obtain a feature clustering result. The device clusters similar features into one category through the above steps, so that the features can be classified for easy identification and distinction.

[0096] In the specific implementation process, the K-Means clustering algorithm is used to cluster the features, and the features of the same category are grouped into corresponding feature graphs. Given a feature sequence set X = {X1, X2, ..., X n}, where X i =[x n ,x n+1 ,...,x n+τ-1 ] represents the feature segment of length τ of the ith power station from time n. In order to reduce the impact of data scale differences on clustering effects, the input is first normalized: in, is the normalized feature sequence, μ and σ are the mean and standard deviation of all sample features respectively. After normalization, k feature sequences v are randomly selected i (i=1,2,...,k) is used as the initial centroid, and k is the final number of clusters, which is determined by the elbow method. Then, for each sample, the distance from it to each centroid is calculated and assigned to the nearest centroid to form K clusters. Using the Euclidean distance metric, the expression is: d(X i -v i )=||X i -v i ||2, for each category, calculate the mean of all samples assigned to it, and use the mean to update the centroid. Compare the new and old centroids. When the change in the centroid is less than the preset threshold or the change is very small after multiple consecutive iterations, the algorithm is considered to have converged. At this time, the iteration is stopped and the final clustering result is obtained. The feature graph is constructed based on the final clustering result. The feature graph structure corresponding to the i-th category is represented as G i (V i ,E i ), where V i Represents the vertex set of the graph, which consists of all feature sequences in the category, Ei The edge set of the graph is composed of the edges connecting every two vertices. Finally, the cloud distributes the corresponding feature map data of each client according to its category, and further predicts it after aggregating it with the local client data. The loss function of the entire cloud model consists of two parts: reconstruction error and clustering loss: L = L c +L r , where L c is the reconstruction error. In this study, the mean square error (MSE) loss is used as the network reconstruction loss to measure the expressive power of the embedded representation on the original data. The formula is: Where Y i is the ith input sequence, is the sequence reconstructed by the network, and n is the total number of sequence samples. r is the clustering loss, using the Davis-Boulding Index (DBI) as the clustering loss, and the calculation formula is as follows: M ij =||c i -c j ||2, where k is the number of categories, R ij is the similarity between the i-th category and the j-th category, S i is the dispersion of each time series in the i-th category, n i is the total number of samples in the ith class, c i is the cluster center of the ith class, M ij is the distance between the i-th and j-th cluster centers.

[0097] In some other optional implementations, the above-mentioned generating unit includes a generating module for generating a clustering function according to the formula G i =(V i ,E i ) generates a dynamic graph structure, where G i represents the i-th dynamic graph structure, V i represents the i-th vertex of the above dynamic graph structure, E i Represents the i-th edge of the above dynamic graph structure. The device generates a dynamic graph structure through the above steps, which can reduce the prediction deviation caused by abnormal or incomplete local data by aggregating feature data, help the model better understand the relationship between features, and enhance the prediction accuracy and robustness of the model.

[0098] Specifically, a feature graph is constructed based on the final clustering result. The feature graph structure corresponding to the i-th category is represented as G i =(V i ,E i), the vertices of the graph consist of all feature sequences within the category, and the edge set of the graph consists of the edges connecting every two vertices. Finally, the cloud distributes the corresponding feature graph data of each client according to its category, and further predicts it after aggregating it with the local client data.

[0099] Figure 8 is a schematic diagram of another load forecasting device according to an embodiment of the present application. Figure 8 As shown, the device comprises:

[0100] The extraction unit 40 acquires historical charging load data of the electric vehicle, performs feature extraction on the historical charging load data, and obtains local features;

[0101] Specifically, after the cloud extracts and clusters the features, the client first uses a one-dimensional convolutional layer to process the client's original load data X and extract local features P.

[0102] The fusion unit 50 is used to obtain the dynamic graph structure sent by the cloud, fuse the feature sequence corresponding to the dynamic graph structure with the local feature, and obtain the fusion feature, wherein the dynamic graph structure is obtained by the cloud, the historical charging load data of the electric vehicle is obtained, the historical charging load data is encrypted, the encrypted data is extracted, a plurality of feature sequences are obtained, and the plurality of feature sequences are clustered, and the feature sequences are generated according to the feature sequences corresponding to the same feature;

[0103] Specifically, after normalization, local features and cloud features are integrated to obtain fused features.

[0104] The second prediction unit 60 is used to perform load prediction based on the above-mentioned fusion characteristics, obtain load prediction results, and allocate power resources according to the above-mentioned load prediction results, wherein the above-mentioned load prediction results indicate the power consumption of the above-mentioned electric vehicle, and the above-mentioned cloud and the above-mentioned client are communicated and connected.

[0105] Specifically, after aggregating the features, the client introduces the MSTEM model to predict the load of the local charging station and obtains the load prediction result.

[0106] In order to perform feature fusion, the above-mentioned fusion unit of the present application includes a normalization module and a fusion module. The normalization module is used to normalize the feature sequence corresponding to the above-mentioned dynamic graph structure and the above-mentioned local features respectively to obtain a normalized feature sequence and a normalized local feature; the fusion module is used to normalize the above-mentioned normalized feature sequence and the above-mentioned normalized local feature by formula Fusion is performed to obtain the above fusion features, where G represents the above fusion features. represents the matrix concatenation operation, P * represents the above normalized local features, S* Represents the above normalized feature sequence. Through the above steps, the device reduces the prediction deviation caused by abnormal or incomplete local data, helps the model to better understand the relationship between features, and enhances the prediction accuracy and robustness of the model.

[0107] Specifically, since the scales of the extracted local features and cloud features are different, they cannot be directly used as input and need to be normalized first: In the formula, P represents the local features extracted by the client, S is the graph structure feature data sent from the cloud, and P * and S * are normalized feature data, μ is the mean of local and cloud feature samples, and σ is the standard deviation. After normalization, local and cloud features are integrated to obtain the final graph structure input. The aggregation expression is as follows:

[0108] In some optional embodiments, the extraction unit includes an extraction module for extracting features from the historical charging load data through a one-dimensional convolution layer to obtain the local features. The device extracts local features through the above steps, so that feature fusion can be performed.

[0109] In the specific implementation process, a one-dimensional convolutional layer is used to process the client's original load data X and extract the local features P.

[0110] In some optional implementations, the second prediction unit includes a division module for dividing the fused features into different scaling scales, using the formula Graph convolution is performed on the above fused features at each of the above scales to capture spatial features, where H si is the hidden representation of the above scaled to si, R si is the final representation of the above scaled to si, represents the adjacency matrix, represents the degree matrix, W1 si represents the first weight matrix, represents the second weight matrix, represents the final weight matrix, σ() represents the ReLU activation function, p represents the Dropout ratio, BN() represents the batch normalization layer, G si represents the spatial features with the above scaling scale si, Indicates that G out The sum of all the spatial features of the above scaling scales is represented. The device extracts spatial features through the above steps, so that features can be extracted from different dimensions to obtain more comprehensive and accurate features.

[0111] In the specific implementation process, after aggregating features, the client introduces the MSTEM model to predict the load of the local charging station. In the MSTEM model, the load data X is first divided into different scales, and a multi-scale graph learning module is used to perform graph convolution on the input of each scale to capture spatial features. Then, a time-enhanced neural network module with integrated LSTM units is used to enhance the model's ability to capture long-term temporal dependencies. This process can be expressed as:

[0112] h t =o t *tanh(c t ), where f t 、i t and t are the outputs of the forget gate, input gate, and output gate respectively. and c t are candidate and updated cell states, h t is the updated hidden state, W l and b l Represents weights and bias terms.

[0113] In some optional implementations, the second prediction unit further includes a determination module and a fusion module, wherein the determination module is used to determine the time characteristics of the fusion features using an integrated LSTM device; and the fusion module is used to determine the time characteristics of the fusion features according to the formula and The above spatial features are combined with the above temporal features to obtain the above load forecasting result, where P represents the above load forecasting result, W1 out Represents the weight of the above spatial features, W2 out represents the weight of the above time features, h t Representing the above time characteristics, represents the historical charging load data sequence captured from the past η time steps, X represents the above historical charging load data, W lo The device performs load forecasting through the above steps to predict the future power load and allocate power resources according to the power load.

[0114] Specifically, the outputs of the above modules are finally integrated to generate the prediction results of the local charging station load. The above process expression is as follows: in is the historical load sequence captured from the past η time steps, W lo and W out are all weight matrices. P is the final load forecast result. In order to fit the actual situation, the model will set the low forecast value in P to zero according to the predefined threshold.

[0115] The load prediction device includes a processor and a memory. The encryption unit, the generation unit, the first prediction unit, etc. are all stored in the memory as program units, and the processor executes the program units stored in the memory to implement corresponding functions. The modules are all located in the same processor; or, the modules are located in different processors in any combination.

[0116] The processor contains a kernel, which calls the corresponding program unit from the memory. One or more kernels can be set, and the load can be predicted by adjusting the kernel parameters while strengthening the privacy protection of users.

[0117] The memory may include non-permanent memory in a computer-readable medium, random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM, and the memory includes at least one memory chip.

[0118] An embodiment of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium includes a stored program, wherein when the program is executed, the device where the computer-readable storage medium is located is controlled to execute the load forecasting method.

[0119] Specifically, the load forecasting method includes:

[0120] Step S201, obtaining historical charging load data of the electric vehicle, and encrypting the historical charging load data to obtain encrypted data;

[0121] Specifically, a framework for cloud-edge collaborative load forecasting is established, including a cloud-based graph structure data generation module based on clustering (cloud center server, referred to as cloud), and a client prediction module based on graph neural network (client edge side, referred to as client). Taking the charging station as an example, the client uploads the historical load data of the local charging station, and the cloud establishes a privacy protection module, and uses the embedding layer to encrypt the data of each charging station to protect privacy. Including the cloud and the client, among which the cloud mainly establishes a cloud-based graph structure data generation module based on clustering, which includes a privacy protection module, a feature extraction module and a dynamic graph structure representation generation module, extracts the features of the load data and constructs a dynamic graph structure representation; then the corresponding dynamic graph structure representation is sent to the corresponding charging station client, and the client extracts the features of the local load data and aggregates it with the distribution features. Therefore, the cloud first obtains the historical charging load data of electric vehicles. For electric vehicle charging stations, the set of all charging records can be expressed as X={X1,X2,...,X N}, where N is the total number of EVCS (Electric Vehicle Charging Station, referred to as EVCS). For data from different clients, if the total length of their charging records is different, the number of fragments is kept consistent by truncating the sequence. After obtaining the load input matrix, the cloud encrypts and reconstructs it through the embedding layer. For data in the form of time series, the embedding layer will consist of fully connected layers.

[0122] Step S202, extracting features from the encrypted data to obtain multiple feature sequences, performing feature clustering on the multiple feature sequences, and generating a dynamic graph structure according to the feature clustering results;

[0123] Specifically, the prediction of the encrypted data corresponding to the above historical charging load data depends on capturing and identifying patterns in the data. After obtaining the embedded representation, the model uses a one-dimensional convolution layer to fully extract the features in the charging information. In the one-dimensional convolution layer, a sliding convolution operation along the time dimension is performed on the embedded representation sequence using a one-dimensional convolution kernel to extract features. When the convolution kernel slides on the input sequence, it calculates the dot product with the input data fragment at the current position and generates the output at the corresponding position, finally generating a feature sequence, and then generating a dynamic graph structure in the cloud based on the feature sequence.

[0124] Step S203, sending the above-mentioned dynamic graph structure to the client, so that the above-mentioned client fuses the above-mentioned feature sequence corresponding to the above-mentioned dynamic graph structure and the local feature extracted by the above-mentioned client to obtain the fused feature, and performs load forecasting based on the above-mentioned fused feature to obtain the load forecasting result, and allocates power resources according to the above-mentioned load forecasting result, wherein the above-mentioned load forecasting result indicates the power consumption of the above-mentioned electric vehicle, and the above-mentioned cloud and the above-mentioned client are communicatively connected.

[0125] Specifically, the feature aggregation module established by the client is then used to extract the local features of the local load data and fuse them with the dynamic graph structure sent by the cloud. The load forecast is further performed based on the fused features. The fusion of features obtained through multiple feature extractions can more accurately predict the power load, obtain the load forecast result, and allocate power resources based on the above load forecast results. Without sacrificing user privacy, it is possible to make full use of distributed data to improve prediction accuracy, reduce the computing and storage requirements for the central server, and reduce communication delays, thereby improving the response speed and efficiency of the overall system. In addition, this method can also effectively process non-graph data. It can deeply explore the user's charging mode by learning the dynamic graph structure representation of load data, provide additional spatiotemporal information and load characteristics for isolated EVCS, and provide charging station clients with personalized local models for more accurate predictions.

[0126] An embodiment of the present invention provides an electronic device, including a processor, a memory, and a program stored in the memory and executable on the processor, wherein when the processor executes the program, at least the following steps are implemented:

[0127] Step S201, obtaining historical charging load data of the electric vehicle, and encrypting the historical charging load data to obtain encrypted data;

[0128] Specifically, a framework for cloud-edge collaborative load forecasting is established, including a cloud-based graph structure data generation module based on clustering (cloud center server, referred to as cloud), and a client prediction module based on graph neural network (client edge side, referred to as client). Taking the charging station as an example, the client uploads the historical load data of the local charging station, and the cloud establishes a privacy protection module, and uses the embedding layer to encrypt the data of each charging station to protect privacy. Including the cloud and the client, among which the cloud mainly establishes a cloud-based graph structure data generation module based on clustering, which includes a privacy protection module, a feature extraction module and a dynamic graph structure representation generation module, extracts the features of the load data and constructs a dynamic graph structure representation; then the corresponding dynamic graph structure representation is sent to the corresponding charging station client, and the client extracts the features of the local load data and aggregates it with the distribution features. Therefore, the cloud first obtains the historical charging load data of electric vehicles. For electric vehicle charging stations, the set of all charging records can be expressed as X={X1,X2,...,X N}, where N is the total number of EVCS (Electric Vehicle Charging Station, referred to as EVCS). For data from different clients, if the total length of their charging records is different, the number of fragments is kept consistent by truncating the sequence. After obtaining the load input matrix, the cloud encrypts and reconstructs it through the embedding layer. For data in the form of time series, the embedding layer will consist of fully connected layers.

[0129] Step S202, extracting features from the encrypted data to obtain multiple feature sequences, performing feature clustering on the multiple feature sequences, and generating a dynamic graph structure according to the feature clustering results;

[0130] Specifically, the prediction of the encrypted data corresponding to the above historical charging load data depends on capturing and identifying patterns in the data. After obtaining the embedded representation, the model uses a one-dimensional convolution layer to fully extract the features in the charging information. In the one-dimensional convolution layer, a sliding convolution operation along the time dimension is performed on the embedded representation sequence using a one-dimensional convolution kernel to extract features. When the convolution kernel slides on the input sequence, it calculates the dot product with the input data fragment at the current position and generates the output at the corresponding position, finally generating a feature sequence, and then generating a dynamic graph structure in the cloud based on the feature sequence.

[0131] Step S203, sending the above-mentioned dynamic graph structure to the client, so that the above-mentioned client fuses the above-mentioned feature sequence corresponding to the above-mentioned dynamic graph structure and the local feature extracted by the above-mentioned client to obtain the fused feature, and performs load forecasting based on the above-mentioned fused feature to obtain the load forecasting result, and allocates power resources according to the above-mentioned load forecasting result, wherein the above-mentioned load forecasting result indicates the power consumption of the above-mentioned electric vehicle, and the above-mentioned cloud and the above-mentioned client are communicatively connected.

[0132] Specifically, the feature aggregation module established by the client is then used to extract the local features of the local load data and fuse them with the dynamic graph structure sent by the cloud. The load forecast is further performed based on the fused features. The fusion of features obtained through multiple feature extractions can more accurately predict the power load, obtain the load forecast result, and allocate power resources based on the above load forecast results. Without sacrificing user privacy, it is possible to make full use of distributed data to improve prediction accuracy, reduce the computing and storage requirements for the central server, and reduce communication delays, thereby improving the response speed and efficiency of the overall system. In addition, this method can also effectively process non-graph data. It can deeply explore the user's charging mode by learning the dynamic graph structure representation of load data, provide additional spatiotemporal information and load characteristics for isolated EVCS, and provide charging station clients with personalized local models for more accurate predictions.

[0133] The devices in this article can be servers, PCs, PADs, mobile phones, etc.

[0134] The present application also provides a computer program product, including a computer program. When the computer program is executed by a processor, the steps of the above method in each embodiment of the present application are implemented:

[0135] Step S201, obtaining historical charging load data of the electric vehicle, and encrypting the historical charging load data to obtain encrypted data;

[0136] Specifically, a framework for cloud-edge collaborative load forecasting is established, including a cloud-based graph structure data generation module based on clustering (cloud center server, referred to as cloud), and a client prediction module based on graph neural network (client edge side, referred to as client). Taking the charging station as an example, the client uploads the historical load data of the local charging station, and the cloud establishes a privacy protection module, and uses the embedding layer to encrypt the data of each charging station to protect privacy. Including the cloud and the client, among which the cloud mainly establishes a cloud-based graph structure data generation module based on clustering, which includes a privacy protection module, a feature extraction module and a dynamic graph structure representation generation module, extracts the features of the load data and constructs a dynamic graph structure representation; then the corresponding dynamic graph structure representation is sent to the corresponding charging station client, and the client extracts the features of the local load data and aggregates it with the distribution features. Therefore, the cloud first obtains the historical charging load data of electric vehicles. For electric vehicle charging stations, the set of all charging records can be expressed as X={X1,X2,...,X N}, where N is the total number of EVCS (Electric Vehicle Charging Station, referred to as EVCS). For data from different clients, if the total length of their charging records is different, the number of fragments is kept consistent by truncating the sequence. After obtaining the load input matrix, the cloud encrypts and reconstructs it through the embedding layer. For data in the form of time series, the embedding layer will consist of fully connected layers.

[0137] Step S202, extracting features from the encrypted data to obtain multiple feature sequences, performing feature clustering on the multiple feature sequences, and generating a dynamic graph structure according to the feature clustering results;

[0138] Specifically, the prediction of the encrypted data corresponding to the above historical charging load data depends on capturing and identifying patterns in the data. After obtaining the embedded representation, the model uses a one-dimensional convolution layer to fully extract the features in the charging information. In the one-dimensional convolution layer, a sliding convolution operation along the time dimension is performed on the embedded representation sequence using a one-dimensional convolution kernel to extract features. When the convolution kernel slides on the input sequence, it calculates the dot product with the input data fragment at the current position and generates the output at the corresponding position, finally generating a feature sequence, and then generating a dynamic graph structure in the cloud based on the feature sequence.

[0139] Step S203, sending the above-mentioned dynamic graph structure to the client, so that the above-mentioned client fuses the above-mentioned feature sequence corresponding to the above-mentioned dynamic graph structure and the local feature extracted by the above-mentioned client to obtain the fused feature, and performs load forecasting based on the above-mentioned fused feature to obtain the load forecasting result, and allocates power resources according to the above-mentioned load forecasting result, wherein the above-mentioned load forecasting result indicates the power consumption of the above-mentioned electric vehicle, and the above-mentioned cloud and the above-mentioned client are communicatively connected.

[0140] Specifically, the feature aggregation module established by the client is then used to extract the local features of the local load data and fuse them with the dynamic graph structure sent by the cloud. The load forecast is further performed based on the fused features. The fusion of features obtained through multiple feature extractions can more accurately predict the power load, obtain the load forecast result, and allocate power resources based on the above load forecast results. Without sacrificing user privacy, it is possible to make full use of distributed data to improve prediction accuracy, reduce the computing and storage requirements for the central server, and reduce communication delays, thereby improving the response speed and efficiency of the overall system. In addition, this method can also effectively process non-graph data. It can deeply explore the user's charging mode by learning the dynamic graph structure representation of load data, provide additional spatiotemporal information and load characteristics for isolated EVCS, and provide charging station clients with personalized local models for more accurate predictions.

[0141] Obviously, those skilled in the art should understand that the above modules or steps of the present invention can be implemented by a general computing device, they can be concentrated on a single computing device, or distributed on a network composed of multiple computing devices, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, and in some cases, the steps shown or described can be executed in a different order than here, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. Thus, the present invention is not limited to any specific combination of hardware and software.

[0142] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0143] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0144] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0145] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0146] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0147] The memory may include non-permanent memory in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. The memory is an example of a computer-readable medium.

[0148] Computer readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.

[0149] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0150] From the above description, it can be seen that the above embodiments of the present application achieve the following technical effects:

[0151] 1) In the load forecasting method of the present application, the historical charging load data of the electric vehicle is obtained, the historical charging load data is encrypted to obtain encrypted data; the encrypted data is feature extracted to obtain multiple feature sequences, the multiple feature sequences are feature clustered, and a dynamic graph structure is generated according to the feature clustering results; the dynamic graph structure is sent to the client, so that the client fuses the feature sequence corresponding to the dynamic graph structure with the local features extracted by the client to obtain fused features, and performs load forecasting according to the fused features to obtain load forecasting results, and allocates power resources according to the load forecasting results. Compared with the prior art, in which the load forecasting method lacks user privacy protection, the present application can encrypt the load data, and then perform feature extraction and load forecasting, thereby improving the security of the data when performing load forecasting and ensuring the data security of the user. Therefore, the problem of the lack of user privacy protection in the load forecasting method in the prior art can be solved, and the effect of protecting user privacy while accurately predicting the load can be achieved.

[0152] 2) In the load forecasting device of the present application, the historical charging load data of the electric vehicle is obtained, the historical charging load data is encrypted to obtain encrypted data; the encrypted data is feature extracted to obtain multiple feature sequences, the multiple feature sequences are feature clustered, and a dynamic graph structure is generated according to the feature clustering results; the dynamic graph structure is sent to the client, so that the client fuses the feature sequence corresponding to the dynamic graph structure with the local features extracted by the client to obtain fused features, and performs load forecasting according to the fused features to obtain load forecasting results, and allocates power resources according to the load forecasting results. Compared with the prior art, in which the load forecasting method lacks user privacy protection, the present application can encrypt the load data, and then perform feature extraction and load forecasting, thereby improving the data security when performing load forecasting and ensuring the user's data security. Therefore, it is possible to solve the problem of the lack of user privacy protection in the load forecasting method in the prior art, and achieve the effect of protecting user privacy while accurately predicting the load.

[0153] The above description is only the preferred embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A load forecasting method, characterized in that: Applied in the cloud, including: Acquire historical charging load data of the electric vehicle, and encrypt the historical charging load data to obtain encrypted data; Extracting features from the encrypted data to obtain multiple feature sequences, performing feature clustering on the multiple feature sequences, and generating a dynamic graph structure according to the feature clustering results; The dynamic graph structure is sent to the client, so that the client fuses the feature sequence corresponding to the dynamic graph structure and the local feature extracted by the client to obtain a fused feature, and performs load forecasting based on the fused feature to obtain a load forecasting result, and allocates power resources based on the load forecasting result, wherein the load forecasting result indicates the power consumption of the electric vehicle, and the cloud and the client are communicated.

2. The load forecasting method according to claim 1, characterized in that: The historical charging load data is encrypted to obtain encrypted data, including: Dividing the historical charging load data into a plurality of time series of predetermined lengths, wherein the predetermined length is less than the length of the historical charging load data; By formula Encrypting a plurality of time series of the predetermined length to obtain the encrypted data, wherein: represents the encrypted data, X represents the historical charging load data, X=[[x0,...,x l-1 ],...,[x t-l+1 ,...,x t ]], l represents the predetermined length, t represents the length of the historical charging load data, W1 represents the weight matrix of the first fully connected layer, W2 represents the weight matrix of the second fully connected layer, b1 represents the first bias vector, and b2 represents the second bias vector.

3. The load forecasting method according to claim 1, characterized in that: Feature extraction is performed on the encrypted data to obtain multiple feature sequences, including: Use a one-dimensional convolutional layer through the formula Extract features from the encrypted data to obtain multiple feature sequences, where (X i *w)[n] represents the convolution result of the i-th historical charging load data sequence and the convolution kernel w at position n, k represents the size of the convolution kernel, w[k] represents the weight of the convolution kernel at position k, X I [nk] represents the convolution result of the I-th historical charging load data sequence at position nk.

4. The load forecasting method according to claim 1, characterized in that: Performing feature clustering on the plurality of feature sequences comprises: Normalizing the plurality of feature sequences to obtain a normalized feature sequence; Randomly select multiple normalized feature sequences as centroids, calculate the distances from samples in each normalized feature sequence to the multiple centroids, and assign the samples to the centroids with the smallest distances to obtain multiple clusters; The mean of all samples corresponding to each cluster is calculated, and the centroid is replaced by the mean to obtain a new centroid. When the absolute value of the difference between the new centroid and the centroid is less than a preset threshold, the iteration is stopped to obtain a feature clustering result.

5. The load forecasting method according to claim 1, characterized in that: Generate a dynamic graph structure based on the feature clustering results, including: According to the feature clustering results, the formula G i =(V i ,E i ) generates a dynamic graph structure, where G i represents the i-th dynamic graph structure, V i represents the i-th vertex of the dynamic graph structure, E i Represents the i-th edge of the dynamic graph structure.

6. A load forecasting method, characterized in that: Applied to the client, including: Acquire historical charging load data of the electric vehicle, perform feature extraction on the historical charging load data, and obtain local features; Obtaining a dynamic graph structure sent by the cloud, fusing a feature sequence corresponding to the dynamic graph structure with the local feature to obtain a fusion feature, wherein the dynamic graph structure is obtained by the cloud acquiring historical charging load data of the electric vehicle, encrypting the historical charging load data to obtain encrypted data, performing feature extraction on the encrypted data to obtain multiple feature sequences, and performing feature clustering on the multiple feature sequences, and is generated according to the feature sequence corresponding to the same feature; Load forecasting is performed according to the fusion features to obtain load forecasting results, and power resources are allocated according to the load forecasting results, wherein the load forecasting results indicate the power consumption of the electric vehicle, and the cloud and the client are communicated and connected.

7. The load forecasting method according to claim 6, characterized in that: The feature sequence corresponding to the dynamic graph structure and the local feature are fused to obtain a fused feature, including: Normalizing the feature sequence and the local feature corresponding to the dynamic graph structure respectively to obtain a normalized feature sequence and a normalized local feature; The normalized feature sequence and the normalized local feature are calculated by the formula Fusion is performed to obtain the fusion feature, where G represents the fusion feature, represents the matrix concatenation operation, P * represents the normalized local feature, S * represents the normalized feature sequence.

8. The load forecasting method according to claim 6, characterized in that: Feature extraction is performed on the historical charging load data to obtain local features, including: The historical charging load data is subjected to feature extraction through a one-dimensional convolutional layer to obtain the local feature.

9. The load forecasting method according to claim 6, characterized in that: Load forecasting is performed according to the fusion features to obtain load forecasting results, including: The fusion features are divided into different scaling scales, and the formula Graph convolution is performed on the fused features at each scale to capture spatial features, where H si is the hidden representation with the scaling factor si, R si is the final representation with the scaling factor si, represents the adjacency matrix, represents the degree matrix, represents the first weight matrix, represents the second weight matrix, represents the final weight matrix, σ() represents the ReLU activation function, p represents the Dropout ratio, BN() represents the batch normalization layer, G si represents the spatial feature with the scaling scale si, Indicates that G out Represents the sum of the spatial features of all the scaling scales.

10. The load forecasting method according to claim 9, characterized in that: Performing load forecasting according to the fusion feature to obtain a load forecasting result also includes: Determine the temporal characteristics of the fused features using an integrated LSTM method; According to the formula The spatial feature and the temporal feature are integrated to obtain the load forecast result, wherein P represents the load forecast result, represents the weight of the spatial feature, represents the weight of the time feature, h t represents the time feature, represents the historical charging load data sequence captured from the past n time steps, X represents the historical charging load data, W lo Represents the load weight.

Images