Charging pile energy management method, device, equipment and storage medium

Abstract

The present application relates to the technical field of charging piles, and discloses a charging pile energy management method, device, equipment and storage medium. The method includes: performing swarm-style data collection on the charging pile network to obtain a distributed data set; adaptively compressing the distributed data set based on a hybrid compression algorithm of wavelet transform and autoencoder to obtain a compressed data set; reconstructing the compressed data set through a graph neural network to obtain an optimized data set; performing multimodal fusion prediction analysis on the optimized data set to obtain a charging demand prediction result; dynamically dividing the charging pile network into microgrids and coordinating the scheduling according to the charging demand prediction result to obtain a target energy allocation plan. The present application improves the accuracy of charging pile energy allocation.

Description

Technical Field

[0001] The present application relates to the technical field of charging piles, and in particular to a charging pile energy management method, device, equipment and storage medium. Background Art

[0002] The rapid popularization of electric vehicles has led to a significant increase in the demand for charging piles, and the construction and optimization of charging infrastructure has become particularly important. The charging pile network requires efficient energy management to ensure the rational distribution of electricity and the satisfaction of charging needs. However, traditional energy management methods face many challenges, including low data collection efficiency, single data processing methods, insufficient accuracy in charging demand prediction, and insufficient refinement of microgrid division and scheduling optimization.

[0003] In existing charging pile energy management systems, data collection usually relies on a single sensor or centralized collection method, resulting in limited data coverage and poor real-time performance. At the same time, traditional data compression and reconstruction algorithms cannot effectively handle distributed and heterogeneous data, resulting in high data transmission and storage costs and low data reconstruction accuracy. In addition, most existing prediction methods are based on simple time series analysis, which fails to fully utilize the potential of multimodal data and is difficult to provide high-precision charging demand prediction. Summary of the invention

[0004] The present application provides a charging pile energy management method, device, equipment and storage medium for improving the accuracy of charging pile energy distribution.

[0005] In a first aspect, the present application provides a charging pile energy management method, the charging pile energy management method comprising:

[0006] Perform swarm data collection on the charging pile network to obtain a distributed data set;

[0007] Based on a hybrid compression algorithm of wavelet transform and autoencoder, the distributed data set is adaptively compressed to obtain a compressed data set;

[0008] Reconstructing the compressed data set through a graph neural network to obtain an optimized data set;

[0009] Performing multimodal fusion prediction analysis on the optimized data set to obtain a charging demand prediction result;

[0010] The charging pile network is dynamically divided into microgrids and coordinated according to the charging demand prediction result to obtain a target energy allocation plan.

[0011] In a second aspect, the present application provides a charging pile energy management device, the charging pile energy management device comprising:

[0012] The acquisition module is used to collect data from the charging pile network in a swarm-like manner to obtain a distributed data set;

[0013] A compression module, used for adaptively compressing the distributed data set based on a hybrid compression algorithm of wavelet transform and autoencoder to obtain a compressed data set;

[0014] A reconstruction module, used to reconstruct the compressed data set through a graph neural network to obtain an optimized data set;

[0015] An analysis module, used to perform multimodal fusion prediction analysis on the optimized data set to obtain a charging demand prediction result;

[0016] The scheduling module is used to dynamically divide and coordinate the charging pile network into microgrids according to the charging demand prediction result to obtain a target energy allocation plan.

[0017] The third aspect of the present application provides a charging pile energy management device, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the charging pile energy management device executes the above-mentioned charging pile energy management method.

[0018] A fourth aspect of the present application provides a computer-readable storage medium, in which instructions are stored. When the computer-readable storage medium is run on a computer, the computer executes the above-mentioned charging pile energy management method.

[0019] In the technical solution provided by this application, through the collaborative work of worker bee nodes and scout bee nodes, extensive and real-time charging pile data collection is achieved. The data collection has a wide coverage and can more accurately reflect the usage and environmental changes of charging piles. Based on the hybrid compression algorithm of wavelet transform and autoencoder, the distributed data set is adaptively compressed, which significantly reduces the data transmission and storage costs. The efficiency and effect of data compression are improved through sparse processing and importance scoring, and key feature information is retained. The compressed data set is reconstructed with high precision using graph neural network, which can restore the structure and characteristics of the original data. Multi-level message passing and graph attention mechanism effectively deal with noise and missing in the data and improve the quality of data reconstruction. The accuracy of charging demand prediction is improved by integrating structured data, text data and image data and using attention mechanism for feature fusion. Combining long short-term memory network and time series analysis, it can capture long-term dependent features and achieve more refined demand prediction. Through spatiotemporal clustering and power flow analysis, the microgrid structure is dynamically adjusted to optimize the load distribution of the charging pile network. The cross-microgrid energy routing protocol and optimal energy transmission path are constructed to improve the efficiency and reliability of energy distribution. Multi-agent collaborative optimization and hierarchical reinforcement learning modeling ensure the optimality of global energy scheduling, achieve efficient utilization of resources, and thus improve the accuracy of energy allocation of charging piles. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0021] Figure 1 This is a schematic diagram of an embodiment of the charging pile energy management method in the embodiment of the present application;

[0022] Figure 2 This is a schematic diagram of an embodiment of a charging pile energy management device in an embodiment of the present application. DETAILED DESCRIPTION

[0023] The embodiments of the present application provide a charging pile energy management method, device, equipment and storage medium. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of this application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0024] For ease of understanding, the specific process of the embodiment of the present application is described below. Figure 1 , an embodiment of the charging pile energy management method in the embodiment of the present application includes:

[0025] Step S101, performing swarm data collection on the charging pile network to obtain a distributed data set;

[0026] It is understandable that the execution subject of the present application may be a charging pile energy management device, or a terminal or a server, which is not specifically limited here. The present application embodiment is described by taking a server as the execution subject as an example.

[0027] Specifically, worker bee nodes are set on the charging piles in the charging pile network to obtain a worker bee node network. These nodes are connected to each other through the near field communication protocol to form an efficient worker bee node cluster. Mobile sensors are installed on electric vehicles to obtain scout bee nodes. These scout bee nodes collect data during the movement of electric vehicles and collect original environmental data sets. Data classification and preprocessing are performed on the original environmental data sets to obtain structured environmental data sets. The structured environmental data sets are prioritized according to the preset data importance evaluation indicators to obtain priority data sets, ensuring that important data can be given priority during data processing and transmission, thereby improving the overall efficiency and response speed of the system. The priority data sets are processed by the bee dance algorithm to obtain the data propagation strategy, and the data is distributed to the worker bee node cluster according to the data propagation strategy to obtain the initial data set. The bee dance algorithm is an optimization algorithm that simulates the foraging behavior of bees. By analyzing the priority data set, the optimal data propagation strategy is formulated, so as to efficiently distribute data in the worker bee node cluster and ensure that each node can quickly obtain and process the required information. On this basis, the performance indicators of the worker bee node cluster are evaluated to obtain the optimal performance node. The evaluation process includes monitoring key indicators such as communication efficiency, data processing capacity and energy consumption of each node to determine the best performance node. The best performance node is set as a queen node, and the initial data set is summarized and processed to obtain a distributed data set. The queen node serves as the central node of the entire network, summarizing the data from each worker node and performing further processing and optimization.

[0028] Step S102, based on a hybrid compression algorithm of wavelet transform and autoencoder, adaptively compress the distributed data set to obtain a compressed data set;

[0029] Specifically, the distributed data set is divided into multiple time series subsets of different scales by time series decomposition technology, and these subsets can better reflect the changing trend and characteristics of the original data. Wavelet transform is applied to the multi-scale time series subsets to obtain a wavelet coefficient matrix, and the wavelet transform can effectively capture the local characteristics of the data at different time scales. The wavelet coefficient matrix is ​​sparsely processed according to a preset threshold to obtain a sparse wavelet coefficient matrix. Sparse processing is to set a threshold and set the wavelet coefficients less than the threshold to zero, so that the data is more sparse, which is conducive to subsequent compression. The sparse wavelet coefficient matrix is ​​encoded and compressed by an autoencoder to obtain a first compressed feature vector. The autoencoder is a neural network model that realizes effective compression and reconstruction of data by learning the low-dimensional representation of data, generates a first compressed feature vector, and can significantly reduce the data dimension while retaining the original information as much as possible. Importance scoring is calculated for data items in the distributed data set to obtain a data importance scoring table. Scoring is performed according to the degree of influence of each data item in the overall data set to quantify the importance of the data. The first compressed feature vector is adaptively quantized according to the data importance scoring table to obtain a second compressed feature vector. Adaptive quantization adjusts the quantization step size so that data items with higher importance can retain more information during the quantization process, thereby improving the overall compression quality. The second compressed feature vector is entropy encoded to obtain a compressed bit stream. Entropy coding is a lossless compression algorithm that performs statistical analysis on the data and uses the probability distribution of the data for efficient encoding to further compress the data volume. At the same time, the parameters of the autoencoder and the data importance score table are packaged to obtain compressed metadata. The compressed metadata contains the structure and parameter information of the autoencoder and the data importance score table. The compressed bit stream and the compressed metadata are combined and packaged to obtain a compressed data set.

[0030] Step S103, reconstructing the compressed data set through a graph neural network to obtain an optimized data set;

[0031] Specifically, the compressed data set is processed by a decompression algorithm to restore the initial reconstructed data set. The initial reconstructed data set is modeled by graph structure to construct a target network graph that reflects the data nodes and the relationships between them. The nodes and edges of the network graph represent the data items in the reconstructed data set and their relationships. The target network graph is subjected to node feature extraction to obtain a node feature matrix. By analyzing the attributes of each node and its positional relationship in the graph, a vector representation reflecting the node characteristics is extracted to form a node feature matrix. According to the node feature matrix, a graph convolutional neural network model is constructed to obtain an initial graph convolutional model. The graph convolutional neural network is a neural network model that can effectively capture the local and global features of graph structured data. By performing convolution operations on the node feature matrix, the initial graph convolutional model can extract the potential complex patterns and relationships in the graph. The initial graph convolutional model is subjected to multi-level message passing iterations to obtain an updated node representation. Multi-level message passing is a method of propagating node information through multiple iterations, so that each node can obtain more levels of neighbor node information, thereby updating the node representation and increasing the expression ability of the model. According to the updated node representation, a graph attention layer is constructed to obtain a weighted node relationship matrix. The graph attention layer calculates the relative importance between nodes and generates a weighted node relationship matrix to enhance the flexibility and accuracy of the model in processing graph structure data. The missing data is interpolated and estimated through the weighted node relationship matrix to obtain a complete data set. Interpolation estimation is to use known data and node relationships to predict and fill in the missing parts of the data to make the data set more complete. Residual learning is performed on the complete data set to obtain a refined data representation. Residual learning is a method to improve the accuracy of the model by learning data residuals, making the data representation more detailed and accurate. Deconvolution is performed on the refined data representation to obtain the target reconstructed data set. The deconvolution operation reconstructs a high-quality target reconstructed data set by restoring the data representation to the original space. The target reconstructed data set is post-processed and normalized to obtain an optimized data set. The post-processing and normalization steps include eliminating data noise, adjusting the data range, etc. to ensure that the final optimized data set has better applicability and consistency.

[0032] Step S104: performing multimodal fusion prediction analysis on the optimized data set to obtain a charging demand prediction result;

[0033] Specifically, the optimized data set is analyzed by data type identification technology, and the data is divided into different modal subsets, including structured data, text data and image data. Feature engineering is performed on the structured data, and structured feature vectors that can reflect the intrinsic characteristics of the data are extracted through steps such as data cleaning, feature selection and feature generation. Natural language processing is performed on the text data in the multimodal data subset to obtain text feature vectors. Natural language processing technology extracts text feature vectors by performing operations such as word segmentation, word vector generation and topic modeling on text data. These vectors can effectively capture the semantic information of text data. At the same time, convolutional neural network processing is performed on the image data in the multimodal data subset, and high-level feature representations of the image are extracted through operations such as convolutional layers, pooling layers and fully connected layers to form image feature vectors. The structured feature vectors, text feature vectors and image feature vectors are fused by attention mechanism to obtain fused feature vectors. The attention mechanism calculates the relative importance between different feature vectors and assigns different weights to different features, thereby achieving effective fusion of multimodal data and forming a comprehensive fused feature vector. A multi-head self-attention model is constructed based on the fused feature vector to obtain context-aware features. The multi-head self-attention model captures the complex dependencies between feature vectors through parallel calculation of multiple attention heads, generating richer context-aware features. The context-aware features are analyzed in time series to obtain time series feature vectors. Time series analysis technology extracts time series feature vectors that can reflect the trend of data changes by modeling the time series dependencies of data. The time series feature vectors are input into the long short-term memory network to obtain long-term dependency features. The long short-term memory network is a recurrent neural network that can capture long-term dependencies. It extracts long-term dependency features of data through memory units and gating mechanisms. The long-term dependency features are processed by residual connection and layer normalization to obtain optimized feature representation. Residual connection alleviates the gradient vanishing problem in deep networks by adding jump connections between network layers, improving the training efficiency and accuracy of the model; layer normalization stabilizes the training process by normalizing the features. The charging demand prediction result is obtained by calculating the fully connected layer and activation function based on the optimized feature representation. The fully connected layer performs linear transformation on the features, and the activation function performs nonlinear transformation, and finally outputs the prediction result to complete the accurate prediction of charging demand.

[0034] Step S105: dynamically divide the charging pile network into microgrids and coordinate scheduling according to the charging demand prediction result to obtain a target energy allocation plan.

[0035] Specifically, the prediction results of charging demand are processed by spatiotemporal clustering analysis technology to identify the concentrated areas of charging demand in time and space and determine the hot spots of charging demand. According to the hot spots of charging demand and the geographical location information of charging piles, the charging pile network is preliminarily divided into microgrids to obtain the initial microgrid structure. Charging piles with adjacent geographical locations and intensive demand are divided into the same microgrid to form multiple initial microgrids. The initial microgrid structure is subjected to power flow analysis to obtain the load distribution of each microgrid. The power flow analysis calculates the power flow of each charging pile in the microgrid to understand the load characteristics of the microgrid, and adjusts and optimizes the boundaries of the initial microgrid structure according to the load distribution to obtain the target microgrid structure. The purpose of boundary adjustment and optimization is to balance the loads between microgrids, ensure the stability and efficient operation of each microgrid, and obtain a more reasonable target microgrid structure. The target microgrid structure is subjected to distributed consensus processing to obtain the energy exchange strategy of each microgrid. Distributed consensus processing determines the energy sharing and complementary strategies between microgrids through information exchange and negotiation between microgrids to ensure the maximum energy utilization efficiency of the entire charging pile network. According to the energy exchange strategy, an inter-microgrid energy routing protocol is constructed to obtain an energy transmission network. The inter-microgrid energy routing protocol defines the rules and paths for energy transmission between different microgrids to ensure that energy can be efficiently transmitted between different microgrids. The energy transmission network is processed by the minimum spanning tree algorithm to obtain the optimal energy transmission path. The minimum spanning tree algorithm is used to find the transmission path with the lowest cost in the energy transmission network to ensure the efficiency and economy of energy transmission. According to the optimal energy transmission path, hierarchical reinforcement learning modeling is performed on each microgrid to obtain the microgrid energy scheduling strategy. Hierarchical reinforcement learning modeling simulates the operation of each microgrid in different environments, learns and optimizes the energy scheduling strategy of each microgrid, and forms a set of scheduling schemes that can adapt to various situations. The microgrid energy scheduling strategy is optimized by multi-agent collaborative optimization to obtain a global collaborative scheduling scheme. Multi-agent collaborative optimization optimizes the energy scheduling strategy through cooperation and competition of multiple agents to ensure the coordinated operation of the entire charging pile network. According to the global collaborative scheduling scheme, energy allocation calculation is performed on each charging pile to obtain the target energy allocation scheme. The energy allocation calculation takes into account the real-time demand of each charging pile and the overall situation of the microgrid, ensuring that each charging pile can get sufficient energy supply when needed, thereby ultimately achieving efficient management and optimized operation of the charging pile network.

[0036] Through density cluster analysis technology, the hot spots of charging demand are processed, the central location of the high-demand area is identified, and the initial cluster center is determined. According to the initial cluster center and the geographical location information of the charging pile, the charging pile is divided into Voronoi diagrams to obtain the first partition result. Voronoi diagram partitioning is a spatial segmentation method that divides the geographical location of the charging pile according to the nearest cluster center to form multiple preliminary partitions. The boundary of the first partition result is optimized to obtain the second partition result. Boundary optimization ensures that the number and distribution of charging piles in each partition are more uniform by adjusting the partition boundaries, thereby improving the rationality and stability of the partition. And the optimal connection structure in each partition is determined based on the second partition result. The optimal connection structure determines the optimal network topology in each partition by analyzing the physical and electrical connections between the charging piles, ensuring that the charging piles in the partition can be efficiently interconnected. According to the optimal connection structure, the electrical characteristics of each partition are analyzed to obtain the impedance matrix inside the partition. The electrical characteristics analysis calculates the electrical parameters of each charging pile and its connection, such as resistance and reactance, to generate an impedance matrix that reflects the electrical characteristics inside the partition. Based on the impedance matrix, the voltage sensitivity analysis of each partition is performed to obtain the voltage sensitivity matrix. Voltage sensitivity analysis studies the voltage response of each node in the partition and generates a voltage sensitivity matrix to reflect the sensitivity of the node voltage to load changes. A subnetwork with balanced electrical characteristics is generated based on the voltage sensitivity matrix. The subnetwork generation process ensures the balanced electrical characteristics of each subnetwork by adjusting the node and connection structure to reduce voltage fluctuations and power losses. A state estimation model is created for each subnetwork, and the boundary nodes and interconnection relationships of each subnetwork are identified and constructed based on the state estimation model to obtain the initial microgrid structure. The state estimation model monitors and predicts the operating status of the subnetwork, identifies the key nodes at the partition boundary, and constructs the interconnection relationship between these nodes to form a stable initial microgrid structure.

[0037] In the embodiment of the present application, through the collaborative work of worker bee nodes and scout bee nodes, extensive and real-time charging pile data collection is achieved. The data collection has a wide coverage and can more accurately reflect the usage and environmental changes of charging piles. Based on the hybrid compression algorithm of wavelet transform and autoencoder, the distributed data set is adaptively compressed, which significantly reduces the data transmission and storage costs. The efficiency and effect of data compression are improved through sparse processing and importance scoring, and key feature information is retained. The compressed data set is reconstructed with high precision using graph neural network, which can restore the structure and characteristics of the original data. Multi-level message passing and graph attention mechanism effectively deal with noise and missing in the data, and improve the quality of data reconstruction. The accuracy of charging demand prediction is improved by integrating structured data, text data and image data and using attention mechanism for feature fusion. Combining long short-term memory network and time series analysis, it can capture long-term dependent features and achieve more refined demand prediction. Through spatiotemporal clustering and power flow analysis, the microgrid structure is dynamically adjusted to optimize the load distribution of the charging pile network. The cross-microgrid energy routing protocol and the optimal energy transmission path are constructed to improve the efficiency and reliability of energy distribution. Multi-agent collaborative optimization and hierarchical reinforcement learning modeling ensure the optimality of global energy scheduling, achieve efficient utilization of resources, and thus improve the accuracy of energy allocation of charging piles.

[0038] In a specific embodiment, the process of executing step S101 may specifically include the following steps:

[0039] (1) Setting worker bee nodes for charging piles in the charging pile network to obtain a worker bee node network, and configuring the worker bee node network with a near field communication protocol to obtain a worker bee node cluster;

[0040] (2) Install mobile sensors on electric vehicles to obtain scout bee nodes, and collect data through the scout bee nodes to obtain the original environmental data set;

[0041] (3) Classify and preprocess the original environmental data set to obtain a structured environmental data set, and prioritize the structured environmental data set according to the preset data importance evaluation index to obtain a priority data set;

[0042] (4) The priority data set is processed by the bee dance algorithm to obtain the data dissemination strategy, and the data is distributed to the worker bee node cluster according to the data dissemination strategy to obtain the initial data set;

[0043] (5) Evaluate the performance indicators of the worker bee node cluster to obtain the optimal performance node;

[0044] (6) Set the best performance node as the queen node, summarize and process the initial data set to obtain a distributed data set.

[0045] Specifically, worker bee nodes are installed on each charging pile, which are responsible for collecting and transmitting data to form a worker bee node network covering the entire charging pile network. The worker bee nodes are configured using the near field communication protocol so that they can transmit data at high speed within a short distance to form a tightly connected worker bee node cluster. Mobile sensors are installed on electric vehicles to obtain scout bee nodes, and data is collected through the scout bee nodes to obtain the original environmental data set. Mobile sensors are installed on each electric vehicle. These sensors serve as scout bee nodes to collect environmental data, including air quality, temperature, humidity, etc., during the movement of the vehicle. The scout bee nodes transmit the collected data to the nearest worker bee nodes through wireless communication and are aggregated into the original environmental data set. The original environmental data set is classified and preprocessed to obtain a structured environmental data set. The data classification and preprocessing steps include removing noise, filling missing data, and normalizing data to ensure data quality and consistency. The structured environmental data set is prioritized according to the preset data importance evaluation indicators to obtain a priority data set. The data importance evaluation indicators may include the time sensitivity, spatial correlation, and the degree of influence of the data on the charging demand prediction. The priority data set is processed by the Bee Dance algorithm to obtain the data propagation strategy, and the data is distributed to the worker bee node cluster according to the data propagation strategy to obtain the initial data set. The Bee Dance algorithm is an optimization algorithm based on the foraging behavior of bees. It optimizes the data transmission strategy by simulating the information transmission and resource allocation between bees. The Bee Dance algorithm determines the optimal data propagation strategy by calculating the priority of each data item and the cost of the transmission path. The core formula of the Bee Dance algorithm is as follows:

[0046] ;

[0047] in, Indicates the path The transmission cost, Indicates the path The set of data items on , Represents a data item The importance rating of Represents a data item Through the path The transmission distance. According to the data propagation strategy obtained by the bee dance algorithm, data is distributed to the worker bee node cluster to form an initial data set. The performance indicators of the worker bee node cluster are evaluated to obtain the optimal performance node. Performance indicators include data transmission rate, node energy consumption, data packet loss rate, etc. Through the comprehensive analysis of these indicators, the node with the best performance is determined. The queen bee node is set for the optimal performance node, and the initial data set is summarized and processed to obtain a distributed data set. The queen bee node serves as the central node of the entire network, summarizing the data from each worker bee node and further processing and analyzing it. In this way, the distributed data set contains detailed information on the entire charging pile network, which can provide support for subsequent energy management and optimization.

[0048] In a specific embodiment, the process of executing step S102 may specifically include the following steps:

[0049] (1) Decompose the distributed data set into time series to obtain multi-scale time series subsets, and perform wavelet transform on the multi-scale time series subsets to obtain the wavelet coefficient matrix;

[0050] (2) performing sparse processing on the wavelet coefficient matrix according to a preset threshold to obtain a sparse wavelet coefficient matrix, and encoding and compressing the sparse wavelet coefficient matrix through an autoencoder to obtain a first compressed feature vector;

[0051] (3) Calculating the importance scores of the data items in the distributed data set to obtain a data importance score table, and adaptively quantizing the first compressed feature vector according to the data importance score table to obtain a second compressed feature vector;

[0052] (4) performing entropy coding on the second compressed feature vector to obtain a compressed bit stream, and performing parameter packaging on the autoencoder and the data importance score table to obtain compressed metadata;

[0053] (5) The compressed bit stream and the compressed metadata are combined and encapsulated to obtain a compressed data set.

[0054] Specifically, the distributed data set is decomposed into time series to obtain multi-scale time series subsets. Time series decomposition is a technique for decomposing original time series data into multiple sub-sequences of different time scales, which can reflect the characteristics and trends of the data at different time scales. Wavelet transform is performed on the multi-scale time series subset to obtain a wavelet coefficient matrix. Wavelet transform is a mathematical tool that can decompose time series data into a series of wavelet coefficients with different frequency and time localization characteristics, thereby providing detailed information about the data at different time scales. The wavelet coefficient matrix is ​​sparsely processed according to a preset threshold to obtain a sparse wavelet coefficient matrix. Sparse processing sets a threshold and sets the wavelet coefficients less than the threshold to zero, thereby reducing the amount of data and improving the sparsity and compression of the data. The formula for sparse processing is as follows:

[0055] ;

[0056] in, represents the elements of the sparse wavelet coefficient matrix, represents the elements of the wavelet coefficient matrix, is a preset threshold. In this way, the amount of data that needs to be stored and transmitted is greatly reduced. The sparse wavelet coefficient matrix is ​​encoded and compressed by the autoencoder to obtain the first compressed feature vector. The autoencoder is an unsupervised neural network model that achieves effective compression and reconstruction of data by learning the low-dimensional representation of data. The autoencoder consists of two parts: an encoder and a decoder. The encoder converts the input data into a low-dimensional feature vector, and the decoder reconstructs the original data based on the low-dimensional feature vector. By training the autoencoder, the sparse wavelet coefficient matrix is ​​compressed into the first compressed feature vector. The importance score of the data items in the distributed data set is calculated to obtain a data importance score table. The data importance score calculation is to score each data item according to the degree of influence in the overall data set, thereby quantifying the importance of the data. According to the data importance score table, the first compressed feature vector is adaptively quantized to obtain a second compressed feature vector. Adaptive quantization is a technology that adjusts the quantization accuracy according to the importance of data, and the purpose is to compress the data volume under the premise of ensuring the accuracy of important data. The second compressed feature vector is entropy encoded to obtain a compressed bit stream. Entropy coding is a lossless compression algorithm that efficiently encodes data and compresses the data volume by utilizing the probability distribution of data. Entropy coding improves coding efficiency by assigning shorter codes to high-frequency data and longer codes to low-frequency data based on the frequency of occurrence of data items. The parameters of the autoencoder and the data importance score table are packaged to obtain compressed metadata. The compressed metadata includes the structure and parameter information of the autoencoder and the data importance score table, which are essential for the decompression and reconstruction of the data. The compressed bitstream is combined and packaged with the compressed metadata to obtain a compressed data set. In this way, the compressed data set not only contains the data information that is efficiently compressed, but also includes the necessary metadata to ensure that the original data can be correctly restored when decompressed.

[0057] In a specific embodiment, the process of executing step S103 may specifically include the following steps:

[0058] (1) Decompress the compressed data set to obtain an initial reconstructed data set, and perform graph structure modeling on the initial reconstructed data set to obtain the target network graph;

[0059] (2) Extract node features from the target network graph to obtain a node feature matrix, and build a graph convolutional neural network model based on the node feature matrix to obtain an initial graph convolutional model;

[0060] (3) Perform multi-level message passing iterations on the initial graph convolutional model to obtain updated node representations;

[0061] (4) Construct a graph attention layer based on the updated node representation to obtain a weighted node relationship matrix, and use the weighted node relationship matrix to interpolate and estimate the missing data to obtain a complete data set;

[0062] (5) Perform residual learning on the complete data set to obtain a refined data representation, and perform deconvolution on the refined data representation to obtain the target reconstructed data set;

[0063] (6) Post-process and normalize the target reconstructed dataset to obtain the optimized dataset.

[0064] Specifically, the original data is extracted from the compressed data set. The decompression process includes decoding and dequantization. Through these steps, the data information before compression can be restored to obtain the initial reconstructed data set. The compressed bit stream is read and the data is decoded using the stored compressed metadata, and then dequantized to restore the original data information. The initial reconstructed data set is modeled with a graph structure to obtain a target network graph. The process of graph structure modeling includes defining the relationship between data points and representing them as nodes and edges. For example, in a charging pile network, each charging pile can be represented as a node, and the connection between charging piles can be represented as an edge. Node feature extraction is performed on the target network graph to obtain a node feature matrix. By analyzing the attributes of each node and its positional relationship in the graph, a vector representation reflecting the node characteristics is extracted to form a node feature matrix. Each row of the node feature matrix represents a node, and each column represents a feature. According to the node feature matrix, a graph convolutional neural network model is constructed to obtain an initial graph convolutional model. The graph convolutional neural network is a neural network model that can effectively capture the local and global features of graph structure data. Through the convolution operation of the node feature matrix, the initial graph convolutional model can extract the potential complex patterns and relationships in the graph. Perform multi-level message passing iterations on the initial graph convolutional model to obtain updated node representations. Multi-level message passing is a method of propagating node information through multiple iterations, so that each node can obtain more levels of neighbor node information, thereby updating the node representation and increasing the expressiveness of the model. The formula for multi-level message passing is as follows:

[0065] ;

[0066] in, Representation Node In the The representation of the layer, N Representation Node The set of neighbor nodes of Representation Node and neighbor nodes The normalization coefficient between Indicates The weight matrix of the layer, is the activation function. According to the updated node representation, a graph attention layer is constructed to obtain a weighted node relationship matrix. The graph attention layer generates a weighted node relationship matrix by calculating the relative importance between nodes, thereby enhancing the flexibility and accuracy of the model in processing graph structured data. The missing data is interpolated and estimated through the weighted node relationship matrix to obtain a complete data set. Using known data and node relationships, the missing parts in the data are predicted and filled to make the data set more complete. Residual learning is performed on the complete data set to obtain a refined data representation. Residual learning is a method to improve the accuracy of the model by learning data residuals, making the data representation more detailed and accurate. The formula for residual learning is as follows:

[0067] ;

[0068] in, represents the refined data representation, Represents input data, represents the residual function, Represents the parameters of the residual function. The refined data representation is deconvolved to obtain the target reconstructed data set. The deconvolution operation reconstructs a high-quality target reconstructed data set by restoring the data representation to the original space. The target reconstructed data set is post-processed and normalized to obtain an optimized data set. The post-processing and normalization steps include eliminating data noise, adjusting the data range, etc. to ensure that the final optimized data set has better applicability and consistency.

[0069] In a specific embodiment, the process of executing step S104 may specifically include the following steps:

[0070] (1) Identify and classify the data types of the optimized data set to obtain a multimodal data subset, and perform feature engineering on the structured data in the multimodal data subset to obtain a structured feature vector;

[0071] (2) Perform natural language processing on the text data in the multimodal data subset to obtain a text feature vector, and perform convolutional neural network processing on the image data in the multimodal data subset to obtain an image feature vector;

[0072] (3) Perform an attention mechanism to fuse the structured feature vector, text feature vector, and image feature vector to obtain a fused feature vector;

[0073] (4) A multi-head self-attention model is constructed based on the fused feature vector to obtain context-aware features, and a time series analysis is performed on the context-aware features to obtain a time series feature vector;

[0074] (5) Input the time series feature vector into the long short-term memory network to obtain the long-term dependency feature, and perform residual connection and layer normalization on the long-term dependency feature to obtain the optimized feature representation;

[0075] (6) Based on the optimized feature representation, the charging demand prediction result is obtained through the fully connected layer and activation function calculation.

[0076] Specifically, the data in the optimized data set is analyzed to identify different types of data and form a multimodal data subset. These subsets may include structured data, text data, and image data. Through data type identification and classification, each type of data is processed and analyzed to ensure that each type of data can be processed in the most appropriate way. The structured data in the multimodal data subset is feature engineered to obtain a structured feature vector. Feature engineering is to extract a vector representation that can better represent the data features by transforming and combining the original data. Feature engineering includes steps such as data cleaning, missing value processing, data standardization, feature selection, and feature generation. For example, for the usage data of charging piles, the daily usage frequency, peak usage time, average charging time and other features are extracted to form a structured feature vector. Natural language processing is performed on the text data in the multimodal data subset to obtain a text feature vector. Natural language processing technology includes steps such as text cleaning, word segmentation, word vector generation, sentiment analysis, and topic modeling. For example, for user comment data, word vector generation technology (such as Word2Vec or BERT) is used to convert text into a vector representation to obtain a text feature vector. At the same time, the image data in the multimodal data subset is processed by a convolutional neural network to obtain an image feature vector. A convolutional neural network is a deep learning model specifically used to process image data. It extracts image features through convolutional layers, pooling layers, and fully connected layers. For example, for monitoring images of charging piles, a pre-trained convolutional neural network model (such as ResNet or VGG) is used to extract image features to obtain an image feature vector. The structured feature vector, text feature vector, and image feature vector are fused using an attention mechanism to obtain a fused feature vector. The attention mechanism calculates the relative importance of different feature vectors and assigns different weights to different features to achieve effective fusion of multimodal data. The formula for the fused feature vector is as follows:

[0077] ;

[0078] in, represents the fused feature vector, Represents the feature vector The weight of Represents the number of feature vectors. A multi-head self-attention model is constructed based on the fused feature vectors to obtain context-aware features. The multi-head self-attention model captures the complex dependencies between feature vectors through parallel calculation of multiple attention heads and generates richer context-aware features. The formula for context-aware features is as follows:

[0079] ;

[0080] in, represents context-aware features, Indicates The output of an attention head is represents the number of attention heads, Represents the output weight matrix. Perform time series analysis on the context-aware features to obtain a time series feature vector. Time series analysis captures the time dependency of data and extracts time series features that reflect the trend of data changes. For example, for charging demand data, the sliding window technology is used to extract the daily charging demand changes and obtain a time series feature vector. The time series feature vector is input into the long short-term memory network (LSTM) to obtain long-term dependency features. LSTM is a recurrent neural network that can capture long-term dependencies. It extracts long-term dependency features of data through memory units and gating mechanisms. The formula of LSTM is as follows:

[0081] ;

[0082] in, Indicates time The hidden state of Indicates time The input data, Indicates time The hidden state of the long-term dependent features is processed by residual connection and layer normalization to obtain optimized feature representation. The residual connection alleviates the gradient vanishing problem in the deep network by adding jump connections between network layers, and improves the training efficiency and accuracy of the model; layer normalization stabilizes the training process by normalizing the features. The formula of residual connection is as follows:

[0083] ;

[0084] in, represents the optimized feature representation, represents the input features, represents the residual function, Represents the parameters of the residual function. The charging demand forecast result is obtained by calculating through the fully connected layer and the activation function according to the optimized feature representation. The fully connected layer performs a linear transformation on the features, and the activation function performs a nonlinear transformation to finally output the forecast result. For example, for the daily charging demand forecast, the daily charging demand forecast value is calculated using the fully connected layer and the ReLU activation function.

[0085] In a specific embodiment, the process of executing step S105 may specifically include the following steps:

[0086] (1) Perform spatiotemporal clustering analysis on the charging demand forecast results to obtain the hot spots of charging demand;

[0087] (2) Based on the charging demand hotspots and the geographical location information of the charging piles, the charging pile network is preliminarily divided into microgrids to obtain the initial microgrid structure;

[0088] (3) Conduct power flow analysis on the initial microgrid structure to obtain the load distribution of each microgrid, and adjust and optimize the boundaries of the initial microgrid structure based on the load distribution to obtain the target microgrid structure;

[0089] (4) Perform distributed consensus processing on the target microgrid structure to obtain the energy exchange strategy of each microgrid, and build a cross-microgrid energy routing protocol based on the energy exchange strategy to obtain the energy transmission network;

[0090] (5) The energy transmission network is processed through the minimum spanning tree algorithm to obtain the optimal energy transmission path. Based on the optimal energy transmission path, hierarchical reinforcement learning modeling is performed on each microgrid to obtain the microgrid energy scheduling strategy;

[0091] (6) Perform multi-agent collaborative optimization on the microgrid energy scheduling strategy to obtain a global collaborative scheduling plan. Based on the global collaborative scheduling plan, perform energy allocation calculations on each charging pile to obtain the target energy allocation plan.

[0092] Specifically, the charging demand data is clustered according to the time and space dimensions through the spatiotemporal clustering algorithm to identify the hot spots of charging demand. The spatiotemporal clustering algorithm can effectively capture the spatiotemporal distribution characteristics of the data and find out the areas where the charging demand is highly concentrated at a specific time and geographical location. Common spatiotemporal clustering algorithms include DBSCAN, K-means, etc. These algorithms can effectively cluster the data according to the density and distance of the data. According to the geographical location information of the hot spots of charging demand and the charging piles, the charging pile network is preliminarily divided into microgrids to obtain the initial microgrid structure. The preliminary microgrid division is based on the geographical location of the hot spots of charging demand and the charging piles, and the charging piles are divided into different microgrids through the spatial segmentation method. The Voronoi diagram is a commonly used spatial segmentation method. Through this method, the charging piles are divided into their respective microgrids according to the principle of the closest distance to form an initial microgrid structure. The initial microgrid structure is subjected to power flow analysis to obtain the load distribution of each microgrid. Power flow analysis is a method for calculating the power flow and voltage distribution in the power system, which can help the system understand the load distribution and power demand in each microgrid. The formula for power flow analysis is as follows:

[0093] ;

[0094] in, Representation Node The active power, and Respectively represent nodes and nodes The voltage, and Respectively represent nodes and nodes The real and imaginary parts of the admittance between and Respectively represent nodes and nodes The phase angle of . According to the load distribution, the boundary of the initial microgrid structure is adjusted and optimized to obtain the target microgrid structure. The purpose of boundary adjustment and optimization is to balance the load between microgrids, ensure the stability and efficient operation of each microgrid, and obtain a more reasonable target microgrid structure. The target microgrid structure is processed by distributed consensus to obtain the energy exchange strategy of each microgrid. Distributed consensus processing determines the energy sharing and complementary strategies between microgrids through information exchange and negotiation between microgrids to ensure that the energy utilization efficiency of the entire charging pile network is maximized. According to the energy exchange strategy, a cross-microgrid energy routing protocol is constructed to obtain an energy transmission network. The cross-microgrid energy routing protocol defines the rules and paths for energy transmission between different microgrids to ensure that energy can be efficiently transmitted between different microgrids. The energy transmission network is processed by the minimum spanning tree algorithm to obtain the optimal energy transmission path. The minimum spanning tree algorithm is used to find the transmission path with the lowest cost in the energy transmission network to ensure the efficiency and economy of energy transmission. The formula of the minimum spanning tree algorithm is as follows:

[0095] ;

[0096] in, Representation Node and nodes The transmission cost between Indicates whether to select a node and nodes The edge between Represents all possible edge sets. According to the optimal energy transmission path, hierarchical reinforcement learning modeling is performed on each microgrid to obtain the microgrid energy scheduling strategy. Hierarchical reinforcement learning modeling simulates the operation of each microgrid in different environments, learns and optimizes the energy scheduling strategy of each microgrid, and forms a scheduling plan that can adapt to various situations. Multi-agent collaborative optimization is performed on the microgrid energy scheduling strategy to obtain a global collaborative scheduling plan. Multi-agent collaborative optimization optimizes the energy scheduling strategy through the cooperation and competition of multiple agents to ensure the coordinated operation of the entire charging pile network. The goal of multi-agent collaborative optimization is to maximize the benefits of the entire system while taking into account the individual needs and constraints of each microgrid. According to the global collaborative scheduling plan, energy allocation calculation is performed on each charging pile to obtain the target energy allocation plan. The energy allocation calculation takes into account the real-time needs of each charging pile and the overall situation of the microgrid, ensuring that each charging pile can get sufficient energy supply when needed, and ultimately achieving efficient management and optimized operation of the charging pile network.

[0097] In a specific embodiment, the execution step performs preliminary microgrid division on the charging pile network according to the charging demand hotspot area and the charging pile geographical location information, and the process of obtaining the initial microgrid structure may specifically include the following steps:

[0098] (1) Perform density cluster analysis on charging demand hotspots to obtain the initial cluster center;

[0099] (2) Based on the initial cluster center and the geographical location information of the charging piles, the charging piles are divided into Voronoi diagrams to obtain the first partition result;

[0100] (3) Optimizing the boundaries of the first partition results to obtain the second partition results, and determining the optimal connection structure within each partition based on the second partition results;

[0101] (4) According to the optimal connection structure, the electrical characteristics of each partition are analyzed to obtain the impedance matrix within the partition;

[0102] (5) Based on the impedance matrix, perform voltage sensitivity analysis on each partition to obtain a voltage sensitivity matrix, and generate a sub-network with balanced electrical characteristics based on the voltage sensitivity matrix;

[0103] (6) Create a state estimation model for each sub-network, and based on the state estimation model, identify the boundary nodes and build the interconnection relationship of each sub-network to obtain the initial microgrid structure.

[0104] Specifically, density clustering analysis is performed on the hotspot areas of charging demand to obtain the initial cluster center. These data may include the frequency of use of charging piles, the spatiotemporal distribution of charging demand, etc. By using density clustering algorithms such as DBSCAN, high-density charging demand areas are identified to obtain the initial cluster center. DBSCAN can find dense areas of charging demand by specifying a radius parameter and a minimum sample number parameter, and use the centers of these areas as the initial cluster centers. According to the initial cluster center and the geographical location information of the charging piles, the charging piles are divided into Voronoi diagrams to obtain the first partition result. The Voronoi diagram is a method of dividing a plane into several polygons, each polygon contains a cluster center point, and the distance from all points in the polygon to the center point is less than the distance to other center points. In this way, the charging piles are allocated to their respective partitions according to the principle of the closest distance to form a preliminary partition structure. The boundary optimization of the first partition result is performed to obtain the second partition result. Boundary optimization adjusts the partition boundaries to make the distribution of charging piles in each partition more uniform, thereby improving the rationality and stability of the partition. For example, using the weighted Voronoi diagram method, the partition boundaries are adjusted according to the usage frequency and load conditions of each charging pile to obtain the optimized partition structure. The optimal connection structure within each partition is determined based on the second partition result. By analyzing the geographical location and electrical connection relationship between the charging piles, the optimal connection method of the charging piles in each partition is determined. This connection method must not only consider the shortest path, but also the efficiency and stability of power transmission. According to the optimal connection structure, the electrical characteristics of each partition are analyzed to obtain the impedance matrix within the partition. The electrical characteristics analysis calculates the electrical parameters of each charging pile and its connection, such as resistance, reactance, etc., to generate an impedance matrix that reflects the electrical characteristics within the partition. The formula of the impedance matrix is ​​as follows:

[0105] ;

[0106] in, Representation Node and nodes The impedance between Representation Node and nodes The resistance between Representation Node and nodes The reactance between Represents an imaginary unit. Based on the impedance matrix, voltage sensitivity analysis is performed on each partition to obtain a voltage sensitivity matrix. Voltage sensitivity analysis generates a voltage sensitivity matrix by studying the voltage response of each node in the partition, reflecting the sensitivity of the node voltage to load changes. The formula of the voltage sensitivity matrix is ​​as follows:

[0107] ;

[0108] in, Representation Node The voltage on the node The sensitivity of active power changes, Representation Node The voltage, Representation Node Active power. Generate a subnetwork with balanced electrical characteristics based on the voltage sensitivity matrix. The subnetwork with balanced electrical characteristics is to balance the electrical characteristics of each subnetwork by adjusting the node and connection structure, thereby reducing voltage fluctuations and power losses. In this way, it is ensured that each subnetwork can operate stably under different load conditions. Create a state estimation model for each subnetwork, and identify the boundary nodes and build interconnection relationships for each subnetwork based on the state estimation model to obtain the initial microgrid structure. The state estimation model monitors and predicts the operating status of the subnetwork, identifies the key nodes at the partition boundary, and builds the interconnection relationship between these nodes, thereby forming a stable initial microgrid structure. The formula of the state estimation model is as follows:

[0109] ;

[0110] in, represents the estimated value of the state, Indicates the current state. represents the Kalman gain, Indicates the measured value, represents the measurement matrix.

[0111] The above describes the charging pile energy management method in the embodiment of the present application. The following describes the charging pile energy management device in the embodiment of the present application. Figure 2 In the embodiment of the present application, an embodiment of the charging pile energy management device includes:

[0112] The collection module 201 is used to collect data from the charging pile network in a swarm-like manner to obtain a distributed data set;

[0113] A compression module 202, used to adaptively compress the distributed data set based on a hybrid compression algorithm of wavelet transform and autoencoder to obtain a compressed data set;

[0114] A reconstruction module 203 is used to reconstruct the compressed data set through a graph neural network to obtain an optimized data set;

[0115] The analysis module 204 is used to perform multi-modal fusion prediction analysis on the optimized data set to obtain a charging demand prediction result;

[0116] The scheduling module 205 is used to dynamically divide the charging pile network into microgrids and coordinate scheduling according to the charging demand prediction results to obtain a target energy allocation plan.

[0117] Through the collaboration of the above components, the worker bee nodes and the scout bee nodes work together to achieve extensive and real-time charging pile data collection. The data collection covers a wide range and can more accurately reflect the usage and environmental changes of charging piles. Based on the hybrid compression algorithm of wavelet transform and autoencoder, the distributed data set is adaptively compressed, which significantly reduces the data transmission and storage costs. The efficiency and effect of data compression are improved through sparse processing and importance scoring, and key feature information is retained. The compressed data set is reconstructed with high precision using graph neural network, which can restore the structure and characteristics of the original data. Multi-level message passing and graph attention mechanism effectively deal with noise and missing in the data and improve the quality of data reconstruction. The accuracy of charging demand prediction is improved by integrating structured data, text data and image data and using attention mechanism for feature fusion. Combining long short-term memory network and time series analysis, it can capture long-term dependency features and achieve more refined demand prediction. Through spatiotemporal clustering and power flow analysis, the microgrid structure is dynamically adjusted to optimize the load distribution of the charging pile network. The cross-microgrid energy routing protocol and optimal energy transmission path are constructed to improve the efficiency and reliability of energy distribution. Multi-agent collaborative optimization and hierarchical reinforcement learning modeling ensure the optimality of global energy scheduling, achieve efficient utilization of resources, and thus improve the accuracy of energy allocation of charging piles.

[0118] The present application also provides a charging pile energy management device, which includes a memory and a processor. The memory stores computer-readable instructions. When the computer-readable instructions are executed by the processor, the processor executes the steps of the charging pile energy management method in the above-mentioned embodiments.

[0119] The present application also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. Instructions are stored in the computer-readable storage medium. When the instructions are executed on a computer, the computer executes the steps of the charging pile energy management method.

[0120] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0121] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program codes.

[0122] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A charging pile energy management method, characterized in that: The charging pile energy management method comprises: Perform swarm data collection on the charging pile network to obtain a distributed data set; Based on a hybrid compression algorithm of wavelet transform and autoencoder, the distributed data set is adaptively compressed to obtain a compressed data set; Reconstructing the compressed data set through a graph neural network to obtain an optimized data set; Perform multimodal fusion prediction analysis on the optimized data set to obtain a charging demand prediction result; specifically including: identifying and classifying the data type of the optimized data set to obtain a multimodal data subset, and performing feature engineering processing on the structured data in the multimodal data subset to obtain a structured feature vector; performing natural language processing on the text data in the multimodal data subset to obtain a text feature vector, and performing convolutional neural network processing on the image data in the multimodal data subset to obtain an image feature vector; performing attention mechanism fusion on the structured feature vector, the text feature vector and the image feature vector to obtain a fused feature vector; constructing a multi-head self-attention model based on the fused feature vector to obtain context-aware features, and performing time series analysis on the context-aware features to obtain a time series feature vector; inputting the time series feature vector into a long short-term memory network to obtain a long-term dependency feature, and performing residual connection and layer normalization processing on the long-term dependency feature to obtain an optimized feature representation; obtaining a charging demand prediction result through full connection layer and activation function calculation based on the optimized feature representation; According to the charging demand prediction results, the charging pile network is dynamically divided into microgrids and coordinated to obtain a target energy allocation plan; specifically including: performing spatiotemporal clustering analysis on the charging demand prediction results to obtain charging demand hot spots; performing density clustering analysis on the charging demand hot spots to obtain initial cluster centers; performing Voronoi diagram division on the charging piles according to the initial cluster centers and the geographical location information of the charging piles to obtain a first partition result; performing boundary optimization on the first partition result to obtain a second partition result, and determining the optimal connection structure within each partition according to the second partition result; performing electrical characteristic analysis on each partition according to the optimal connection structure to obtain an impedance matrix within the partition; performing voltage sensitivity analysis on each partition based on the impedance matrix to obtain a voltage sensitivity matrix, and generating a sub-network with balanced electrical characteristics according to the voltage sensitivity matrix; creating a state estimation model for each sub-network, and according to the The state estimation model is used to identify the boundary nodes of each sub-network and build the interconnection relationship to obtain the initial microgrid structure; the power flow analysis is performed on the initial microgrid structure to obtain the load distribution of each microgrid, and according to the load distribution, the boundary of the initial microgrid structure is adjusted and optimized to obtain the target microgrid structure; the target microgrid structure is processed by distributed consensus to obtain the energy exchange strategy of each microgrid, and according to the energy exchange strategy, a cross-microgrid energy routing protocol is constructed to obtain the energy transmission network; the energy transmission network is processed by the minimum spanning tree algorithm to obtain the optimal energy transmission path, and according to the optimal energy transmission path, hierarchical reinforcement learning modeling is performed on each microgrid to obtain the microgrid energy scheduling strategy; the microgrid energy scheduling strategy is optimized by multi-agent collaborative optimization to obtain a global collaborative scheduling plan, and according to the global collaborative scheduling plan, energy allocation calculation is performed on each charging pile to obtain the target energy allocation plan.

2. The charging pile energy management method according to claim 1, characterized in that: The swarm-style data collection on the charging pile network is performed to obtain a distributed data set, including: Setting worker bee nodes on charging piles in the charging pile network to obtain a worker bee node network, and configuring the worker bee node network with a near field communication protocol to obtain a worker bee node cluster; Installing mobile sensors on electric vehicles to obtain scout bee nodes, and collecting data through the scout bee nodes to obtain original environmental data sets; Performing data classification and preprocessing on the original environment data set to obtain a structured environment data set, and prioritizing the structured environment data set according to a preset data importance evaluation index to obtain a priority data set; The priority data set is processed by a bee dance algorithm to obtain a data propagation strategy, and data is distributed to the worker bee node cluster according to the data propagation strategy to obtain an initial data set; Evaluate the performance indicators of the worker bee node cluster to obtain the optimal performance node; The best performance node is set as a queen node, and the initial data set is aggregated and processed to obtain a distributed data set.

3. The charging pile energy management method according to claim 1, characterized in that: The hybrid compression algorithm based on wavelet transform and autoencoder adaptively compresses the distributed data set to obtain a compressed data set, including: Performing time series decomposition on the distributed data set to obtain a multi-scale time series subset, and performing wavelet transform on the multi-scale time series subset to obtain a wavelet coefficient matrix; Performing sparse processing on the wavelet coefficient matrix according to a preset threshold value to obtain a sparse wavelet coefficient matrix, and encoding and compressing the sparse wavelet coefficient matrix through an autoencoder to obtain a first compressed feature vector; Performing importance scoring calculation on the data items in the distributed data set to obtain a data importance scoring table, and adaptively quantizing the first compressed feature vector according to the data importance scoring table to obtain a second compressed feature vector; Performing entropy coding on the second compressed feature vector to obtain a compressed bit stream, and performing parameter packaging on the autoencoder and the data importance score table to obtain compressed metadata; The compressed bit stream and the compressed metadata are combined and packaged to obtain a compressed data set.

4. The charging pile energy management method according to claim 1, characterized in that: The step of reconstructing the compressed data set by using a graph neural network to obtain an optimized data set includes: Decompressing the compressed data set to obtain an initial reconstructed data set, and performing graph structure modeling on the initial reconstructed data set to obtain a target network graph; Extracting node features from the target network graph to obtain a node feature matrix, and constructing a graph convolutional neural network model based on the node feature matrix to obtain an initial graph convolutional model; Performing multi-level message passing iterations on the initial graph convolutional model to obtain updated node representations; Constructing a graph attention layer according to the updated node representation to obtain a weighted node relationship matrix, and interpolating and estimating the missing data through the weighted node relationship matrix to obtain a complete data set; Performing residual learning on the complete data set to obtain a refined data representation, and performing a deconvolution operation on the refined data representation to obtain a target reconstructed data set; The target reconstructed data set is post-processed and normalized to obtain an optimized data set.

5. A charging pile energy management device, characterized in that: The charging pile energy management device comprises: The acquisition module is used to collect data from the charging pile network in a swarm-like manner to obtain a distributed data set; A compression module, used for adaptively compressing the distributed data set based on a hybrid compression algorithm of wavelet transform and autoencoder to obtain a compressed data set; A reconstruction module, used to reconstruct the compressed data set through a graph neural network to obtain an optimized data set; An analysis module is used to perform multimodal fusion prediction analysis on the optimized data set to obtain a charging demand prediction result; specifically comprising: identifying and classifying the data type of the optimized data set to obtain a multimodal data subset, and performing feature engineering processing on the structured data in the multimodal data subset to obtain a structured feature vector; performing natural language processing on the text data in the multimodal data subset to obtain a text feature vector, and performing convolutional neural network processing on the image data in the multimodal data subset to obtain an image feature vector; performing attention mechanism fusion on the structured feature vector, the text feature vector and the image feature vector to obtain a fused feature vector; constructing a multi-head self-attention model based on the fused feature vector to obtain a context-aware feature, and performing time series analysis on the context-aware feature to obtain a time series feature vector; inputting the time series feature vector into a long short-term memory network to obtain a long-term dependency feature, and performing residual connection and layer normalization processing on the long-term dependency feature to obtain an optimized feature representation; and obtaining a charging demand prediction result through full connection layer and activation function calculation based on the optimized feature representation; The scheduling module is used to dynamically divide the charging pile network into microgrids and coordinate scheduling according to the charging demand prediction results to obtain a target energy allocation plan; specifically includes: performing spatiotemporal clustering analysis on the charging demand prediction results to obtain charging demand hot spots; performing density clustering analysis on the charging demand hot spots to obtain initial cluster centers; performing Voronoi diagram division on the charging piles according to the initial cluster centers and the geographical location information of the charging piles to obtain a first partition result; performing boundary optimization on the first partition result to obtain a second partition result, and determining the optimal connection structure within each partition according to the second partition result; performing electrical characteristic analysis on each partition according to the optimal connection structure to obtain an impedance matrix within the partition; performing voltage sensitivity analysis on each partition based on the impedance matrix to obtain a voltage sensitivity matrix, and generating a sub-network with balanced electrical characteristics according to the voltage sensitivity matrix; creating a state estimation model for each sub-network, and According to the state estimation model, boundary nodes of each sub-network are identified and interconnected relationships are constructed to obtain an initial microgrid structure; power flow analysis is performed on the initial microgrid structure to obtain the load distribution of each microgrid, and based on the load distribution, the boundary of the initial microgrid structure is adjusted and optimized to obtain a target microgrid structure; distributed consensus processing is performed on the target microgrid structure to obtain the energy exchange strategy of each microgrid, and based on the energy exchange strategy, a cross-microgrid energy routing protocol is constructed to obtain an energy transmission network; the energy transmission network is processed by a minimum spanning tree algorithm to obtain an optimal energy transmission path, and based on the optimal energy transmission path, hierarchical reinforcement learning modeling is performed on each microgrid to obtain a microgrid energy scheduling strategy; multi-agent collaborative optimization is performed on the microgrid energy scheduling strategy to obtain a global collaborative scheduling plan, and based on the global collaborative scheduling plan, energy allocation calculation is performed on each charging pile to obtain a target energy allocation plan.

6. A charging pile energy management device, characterized in that: The charging pile energy management device comprises: a memory and at least one processor, wherein instructions are stored in the memory; The at least one processor calls the instruction in the memory so that the charging pile energy management device executes the charging pile energy management method according to any one of claims 1 to 4.

7. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instruction is executed by the processor, the charging pile energy management method according to any one of claims 1 to 4 is implemented.

Images