Energy storage system grid dispatching method based on dynamic spatiotemporal graph convolutional neural network

Abstract

The present invention provides a grid dispatching method for an energy storage system based on a dynamic spatiotemporal graph convolutional neural network. This method relates to the field of energy storage management and includes the following steps: S1: acquiring time series data of grid nodes; S2: sequentially performing graph convolution operations and time convolution operations using a dynamic spatiotemporal graph convolutional neural network to extract the spatiotemporal features of the time series data; S3: calculating attention-weighted features of the spatiotemporal features; S4: processing the uncertainty of renewable energy output based on fuzzy rules using the spatiotemporal features and attention-weighted features; S5: implementing distributed model training based on a federated learning framework and dynamically updating model parameters through an attention mechanism; S6: integrating the geographic location and connection relationships of nodes to output enhanced features; S7: outputting grid load forecast results based on the enhanced features through a fully connected layer and activation function, and dynamically regulating the charging and discharging strategy of the energy storage system. This method achieves a dynamic balance between power supply and demand in scenarios with high penetration of renewable energy, reducing operating costs and environmental impact.

Description

Technical Field

[0001] The present invention relates to the field of energy storage management technology, and in particular to a power grid deployment method for an energy storage system based on a dynamic spatiotemporal graph convolutional neural network. Background Art

[0002] An energy system dominated by renewable energy is a hallmark of future development, but it also places enormous pressure on power regulation. Insufficient peak-shaving capacity will lead to curtailment of wind and solar power. In a power system dominated by renewable energy, relying solely on controllable units will make it difficult to meet system regulation needs. The dispatchability of a given flexibility resource is primarily constrained by its ramp rate, which in turn determines the responsiveness of the system's flexibility needs.

[0003] In scenarios with high renewable energy penetration, traditional power dispatch systems have exposed the following shortcomings: Renewable energy output exhibits typical spatial and temporal imbalances, with minute-by-minute fluctuations in photovoltaic output exceeding 50% of installed capacity. This leads to a severe shortage of traditional peak-shaving capacity and a high wind and solar curtailment rate of 15%-20%, resulting in resource waste. Multi-agent coordination is difficult: data on flexible resources such as distributed charging stations and energy storage power stations is siloed, making it difficult for centralized dispatch models to balance privacy protection with global optimization. Centralized optimization methods that rely on global data aggregation pose privacy risks, and traditional graph neural networks, using fixed adjacency matrices, are unable to capture the spatiotemporal dynamic correlations of renewable energy output. Furthermore, when dealing with renewable energy fluctuations, traditional PID control or rule engines increase equipment start-up and shutdown times by 3-5 times, significantly shortening the lifespan of energy storage systems.

[0004] To solve the above problems, the present invention proposes an intelligent allocation method that integrates dynamic spatiotemporal graph modeling, federated learning privacy protection, and fuzzy uncertainty processing, constructs a three-layer collaborative architecture of data-driven, knowledge-guided, and physically constrained, and achieves a dynamic balance of power supply and demand under scenarios with high penetration of new energy. Summary of the Invention

[0005] To address the above problems, the present invention proposes a dynamic spatial-temporal graph convolutional neural networks algorithm FC-FL-AM-DST-GCN based on fuzzy control, federated learning and attention mechanism. Combined with the example of a photovoltaic renewable energy charging energy storage system, the FC-FL-AM-DST-GCN algorithm can monitor and analyze the supply of various energy sources and grid load in real time. It is especially suitable for the dynamic balance of power supply and demand in scenarios with a high proportion of renewable energy access, intelligently allocate charging resources, ensure the use of green electricity for charging during peak periods, and reduce operating costs and environmental impact.

[0006] To achieve the above objectives, the following technical solutions are adopted:

[0007] In a first aspect, the present invention provides a method for grid deployment of an energy storage system based on a dynamic spatiotemporal graph convolutional neural network, comprising:

[0008] Step S1: The data input layer obtains the time series data of the power grid nodes;

[0009] Step S2: The dynamic spatial-temporal graph convolutional layer uses a dynamic spatial-temporal graph convolutional neural network (DST-GCN) to perform graph convolution layer operations and temporal convolution operations in sequence to extract the spatiotemporal features of the time series data;

[0010] Step S3: Calculating the attention-weighted features of the spatiotemporal features through the attention mechanism layer;

[0011] Step S4: inputting the spatiotemporal features and the attention weighted features into a fuzzy control module, and processing the uncertainty of the new energy output based on fuzzy rules;

[0012] Step S5: Implement distributed model training based on the federated learning framework and dynamically update model parameters through the attention mechanism;

[0013] Step S6: Integrate the node geographic location and connection relationship through the physical information fusion layer and output enhanced features;

[0014] Step S7: Based on the enhanced features, the grid load forecast result is output through the fully connected layer and activation function, and the charging and discharging strategy of the energy storage system is dynamically adjusted.

[0015] Furthermore, the step S1 further includes:

[0016] A three-dimensional tensor (N, T, F) is constructed based on the time series data of the grid nodes, where N is the number of nodes, T is the number of time steps, and F is the number of features. The nodes include photovoltaic power stations, charging stations, and grid access points, and the features include at least electricity production, consumption, and load.

[0017] Furthermore, the operations performed by the dynamic spatiotemporal graph convolution layer in step S2 include:

[0018] Dynamic graph construction: A dynamic graph structure is constructed based on the physical connection relationship and spatiotemporal dependency relationship between nodes. The physical connection relationship is determined by the grid topology structure to determine the physical connectivity between nodes; the spatiotemporal dependency relationship is determined by the Pearson correlation coefficient.

[0019] Graph convolution operation: for the dynamic graph structure and the three-dimensional tensor (N, T, F), the spatial features of each time step are extracted using graph convolution operation;

[0020] Temporal convolution operation: Perform a one-dimensional temporal convolution operation on the spatial features along the time dimension and output a spatiotemporal feature representation.

[0021] Furthermore, the dynamic graph construction includes:

[0022] Determine the physical connection relationship of nodes based on the power grid topology;

[0023] The spatiotemporal dependencies between nodes were quantified using the Pearson correlation coefficient;

[0024] Generates a time-varying sparse adjacency matrix where non-zero elements represent valid connections.

[0025] Furthermore, the step S3: calculating the attention-weighted features of the spatiotemporal features through the attention mechanism layer includes:

[0026] For the spatiotemporal feature representation, the attention weight of each node feature in the time dimension is calculated based on the self-attention mechanism, and the time attention weighted feature is output;

[0027] For the temporal attention weighted feature, the attention weights of each node and other nodes in the graph structure are calculated based on the graph attention mechanism, and the graph attention feature is output.

[0028] Furthermore, the step S4: inputting the spatiotemporal features and the graph attention features into a fuzzy control module, processing the uncertainty of the new energy output based on fuzzy rules, and generating a dynamic allocation strategy includes:

[0029] Fuzzify the graph attention features and define the fuzzy sets and membership functions of the input / output variables;

[0030] Define fuzzy rules based on expert knowledge and system characteristics, and apply the fuzzy rule base to perform Mamdani reasoning;

[0031] Through the center of gravity method or maximum membership method, defuzzification is used to generate control decisions, including charging scheduling instructions.

[0032] Furthermore, the step S5: implementing distributed model training based on the federated learning framework and dynamically updating model parameters through the attention mechanism includes:

[0033] Each distributed node uses local data to train a local model. Each node calculates its own loss function and updates the local model parameters according to the gradient of the loss function.

[0034] The federated learning server aggregates the local model parameters of each node to generate global model parameters;

[0035] Send the global model parameters to each node to synchronously update the local model, and output the updated spatiotemporal features of federated learning;

[0036] Among them, in the federated learning module: local model training adopts differential privacy technology and adds Gaussian noise to the gradient update process; model aggregation adopts a weighted average strategy, and weights are dynamically allocated according to the amount of node data.

[0037] Furthermore, the step S6: integrating the node geographic location and connection relationship through the physical information fusion layer to output enhanced features includes:

[0038] Encoding the physical information of the node into a physical constraint matrix, wherein the physical information includes: grid impedance parameters and line capacity limitations;

[0039] fusing the spatiotemporal features updated by the federated learning with the physical constraint matrix through a feature crossover operation to obtain a fused feature;

[0040] With respect to the fusion features, the relationships and dependencies between nodes are enhanced through physical information, and enhanced features are output.

[0041] Furthermore, the step S7: outputting the grid load forecast result based on the enhanced features through the fully connected layer and the activation function, and dynamically regulating the energy storage system charging and discharging strategy, includes:

[0042] The enhanced feature representation output by the physical information fusion layer is input into the fully connected layer, and a three-channel prediction output is generated through linear transformation, including charging and discharging power instructions, energy storage unit selection instructions, and grid interaction confidence.

[0043] The physical constraint activation function includes the charge and discharge instruction activation function, the energy storage selection instruction activation and the confidence activation; and the charge and discharge instructions are corrected in real time based on the safety constraints.

[0044] In a second aspect, the present invention further provides an energy storage system, comprising:

[0045] Photovoltaic power generation module, including photovoltaic panels and inverters, used to collect solar energy and convert it into electrical energy and convert direct current into alternating current for system use;

[0046] Energy storage module, including lithium-ion battery pack and bidirectional charge and discharge controller, used to store excess photovoltaic power and support peak power output;

[0047] Charging stations equipped with a multi-protocol intelligent charging interface and a battery management module to support fast charging of multiple electric vehicle battery specifications and monitor the status of each lithium-ion battery pack;

[0048] The energy management module is used to monitor the production of photovoltaic power, grid load, energy storage status and charging demand of charging stations in real time, and execute the energy storage system grid dispatch method based on dynamic spatiotemporal graph convolutional neural network as described in the first aspect, optimize the allocation of charging resources, give priority to the use of photovoltaic power for charging, and intelligently dispatch grid power during peak periods.

[0049] Compared with the prior art, the present invention achieves the following beneficial effects:

[0050] 1. Maximizing green energy utilization: In scenarios with a high proportion of new energy sources, fuzzy control is used to process uncertainty information, effectively capturing spatiotemporal correlations and improving adaptability to dynamic environments. Photovoltaic power is prioritized for battery charging, reducing reliance on fossil fuels.

[0051] 2. Reducing peak power demand: Taking physical constraints into account, the model improves real-world adaptability. Through intelligent power forecasting and dispatch, power demand is minimized during peak periods, reducing power costs and load pressure.

[0052] 3. Improve system efficiency and collaboration: A federated learning framework enables distributed model training and collaborative optimization while protecting the data privacy of all participants (PV power plants, charging stations, and power grids). By integrating physical constraints and fuzzy control, the system significantly improves power forecast accuracy and the feasibility of dispatch instructions, thereby improving overall energy efficiency and reducing system operating costs.

[0053] It should be understood that the contents described in the summary of the invention are not intended to limit the key or important features of the embodiments of the present invention, nor are they intended to limit the scope of the present invention. Other features of the present invention will become readily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] The above and other features, advantages and aspects of the embodiments of the present invention will become more apparent with reference to the following detailed description in conjunction with the accompanying drawings. The accompanying drawings are provided for a better understanding of the present invention and do not constitute a limitation of the present invention. In the accompanying drawings, the same or similar reference numerals represent the same or similar elements, among which:

[0055] Figure 1 1. It is a flow chart of a method for grid deployment of an energy storage system based on a dynamic spatiotemporal graph convolutional neural network according to an embodiment of the present invention;

[0056] Figure 2 This is a flow chart of the convolutional layer of the network dynamic spatiotemporal graph according to an embodiment of the present invention;

[0057] Figure 3 is a schematic diagram of a module of an energy storage system according to an embodiment of the present invention;

[0058] Figure 4 1 is a schematic diagram of the grid allocation process executed by the energy management module of the energy storage system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0059] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0060] In this document, the term "and / or" simply describes a relationship between related objects, indicating that three possible relationships exist. For example, "A and / or B" can represent: A exists alone, A and B exist simultaneously, or B exists alone. Furthermore, the character " / " in this document generally indicates that the related objects are in an "or" relationship.

[0061] Example 1:

[0062] Figure 1 The figure shows a flow chart of the energy storage system grid deployment method based on the dynamic spatiotemporal graph convolutional neural network according to an embodiment of the present invention. Figure 1 As shown in the figure, the energy storage system grid dispatching method based on dynamic spatiotemporal graph convolutional neural network is applied to Figure 3 The energy storage system 200 shown includes the following steps:

[0063] Step S1: The data input layer obtains the time series data of the power grid nodes;

[0064] Step S2: The dynamic spatiotemporal graph convolution layer uses a dynamic spatiotemporal graph convolutional neural network (DST-GCN) to sequentially perform graph convolution operations and temporal convolution operations to extract the spatiotemporal features of the time series data;

[0065] Step S3: Calculating the attention-weighted features of the spatiotemporal features through the attention mechanism layer;

[0066] Step S4: inputting the spatiotemporal features and the attention weighted features into a fuzzy control module, processing the uncertainty of new energy output based on fuzzy rules, and generating a dynamic allocation strategy;

[0067] Step S5: Implement distributed model training based on the federated learning framework and dynamically update model parameters through the attention mechanism;

[0068] Step S6: Integrate the node geographic location and connection relationship through the physical information fusion layer and output enhanced features;

[0069] Step S7: Based on the enhanced features, the grid load forecast result is output through the fully connected layer and activation function, and the charging and discharging strategy of the energy storage system is dynamically adjusted.

[0070] Figure 3 Schematic diagram of the system architecture of an energy storage system 200 according to an embodiment of the present invention. Figure 3 As shown, an energy storage system 200 according to an embodiment of the present invention is used to implement a method for grid deployment of an energy storage system based on a dynamic spatiotemporal graph convolutional neural network. The energy storage system 200 includes:

[0071] Photovoltaic power generation module 210, including photovoltaic panels and inverters, is used to collect solar energy and convert it into electrical energy and convert direct current into alternating current for use by the system;

[0072] Furthermore, photovoltaic panels are installed at the charging station or nearby areas to collect solar energy and convert it into electrical energy; inverters convert the direct current generated by the photovoltaic panels into alternating current for use by the system.

[0073] Energy storage module 220, including a lithium-ion battery pack and a bidirectional charge and discharge controller, is used to store excess photovoltaic power and support peak power output;

[0074] Furthermore, battery packs are used to store excess photovoltaic power; charge and discharge controllers ensure that green power can be used at night or on cloudy days.

[0075] Charging Station 230, equipped with a multi-protocol intelligent charging interface and battery management module;

[0076] Furthermore, the intelligent charging interface supports the rapid replacement and charging of various electric vehicle battery specifications; the battery management module monitors the status of each battery, including charging status, health status and temperature, to ensure safe and efficient charging.

[0077] Energy management module 240: The energy management module is used to monitor the production of photovoltaic power, grid load, energy storage status, and charging demand of charging stations in real time. It also has a built-in FC-FL-AM-DST-GCN algorithm scheduling core, executes the energy storage system grid dispatch method based on dynamic spatiotemporal graph convolutional neural network, optimizes the allocation of charging resources, prioritizes the use of photovoltaic power for charging, and intelligently dispatches grid power during peak periods.

[0078] Furthermore, step S1: the data input layer obtains time series data of the power grid nodes, including:

[0079] Input data: This includes time series data from different nodes (e.g., photovoltaic power plants, charging stations, and power grids). Each node's characteristics include electricity production, consumption, and load. Data format: A three-dimensional tensor (N, T, F), where N is the number of nodes, T is the number of time steps, and F is the number of features.

[0080] The data input layer constructs a three-dimensional tensor (N, T, F) based on the time series data of the grid nodes, where N is the number of nodes, T is the number of time steps, and F is the number of features. Nodes include photovoltaic power stations, charging stations, and grid access points, and features include at least electricity production, consumption, and load.

[0081] Further, such as Figure 2 The figure shows the flow chart of the convolution layer of the network dynamic spatiotemporal graph according to an embodiment of the present invention. Figure 2 The numbers 1, 2, 3, ..., 34, and 35 in the grid represent different nodes of the grid topology structure. The processing of the dynamic spatiotemporal graph convolution layer in step S2 includes:

[0082] S21: Dynamic graph construction: Construct a dynamic graph structure (N, N, T) based on the physical connection relationship and spatiotemporal dependency relationship between nodes;

[0083] This can be achieved by constructing an adjacency matrix, where each node is represented as 1 if it has connections to its neighboring nodes (based on physical location or relatedness), and 0 otherwise. A is (N, N), which represents the connection relationship between nodes.

[0084] Abstract photovoltaic power stations, charging stations, and grid nodes into graph nodes, adjacency matrix A(t) Dynamically encodes two types of relationships:

[0085] Physical connections (e.g., cable topology), where the physical connection relationship determines the physical connectivity between nodes through the grid topology;

[0086] Spatiotemporal correlation (such as regional lighting similarity), quantifying the spatiotemporal dependence between nodes through the Pearson correlation coefficient:

[0087]

[0088] Representation node i With node j The Pearson correlation coefficient is used to quantify the node i With node j The time series similarity of , W represents the sliding time window (corresponding to 3 hours of historical data, step length 15 minutes, covering the new energy fluctuation cycle), and the value W=12; Representation node i At the moment τ characteristic values ​​(such as photovoltaic output), Representation node i The feature mean within the time window, Representation node j The mean of the feature within the time window.

[0089] Adjacency Matrix A Elements :

[0090]

[0091] in, When it is 1, it means the node i With node j In time t There is a connection, when it is 0, it means the node i With node j In time t No connection exists; represents the elements of the physical connection adjacency matrix, where Representation node i With node j Establish a physical connection through a direct cable connection; when >0.7 when establishing a connection;

[0092] The sliding time window is 15 minutes, that is, it is calculated every 15 minutes , update the adjacency matrix A.

[0093] The connectivity relationships were quantified based on the physical connections based on the grid topology (hard connections) and the dynamic establishment of spatiotemporal dependencies (soft connections) through a correlation coefficient threshold (0.7), thereby generating a time-varying sparse adjacency matrix with non-zero elements indicating valid connections.

[0094] S22: Graph convolution operation: Graph convolution is used to extract spatial features at each time step for dynamic graph structures (N, N, T) and three-dimensional tensors (N, T, F).

[0095]

[0096] in: Indicates that after l After the graph convolution operation of the layer l +1 layer at time step t The output feature matrix of Indicates the l The feature representation of the layer, the initial value is l =0, it is the input feature matrix X; A is the adjacency matrix; is the weight matrix of graph convolution; is the bias term; is the activation function, and the LeakyReLU function is optional;

[0097] Graph convolution is performed on each time step, outputting the spatial features (N, T, F') for each time step. F' is the feature dimension after graph convolution. Graph convolution is used to capture spatial dependencies between nodes, such as the coordinated changes in output of adjacent PV stations.

[0098] S23: Temporal convolution operation: performing a one-dimensional temporal convolution operation on the spatial features along the time dimension, and outputting a spatiotemporal feature representation.

[0099] Perform one-dimensional convolution operation on the spatial features in the time dimension to extract the time features.

[0100] Temporal convolution operation formula:

[0101]

[0102] in: Indicates that after l After the temporal convolution operation of the layer, l +1 layer output, at time step t The feature representation of all nodes is a matrix (number of nodes N × feature dimension F''); represents the feature representation of time step tk after the graph convolution operation; It is the weight matrix of time convolution, and the time convolution kernel K is 3.

[0103] Perform temporal convolution on each node and output spatiotemporal feature representation , F'' is the feature dimension after time convolution, for example, F"=64. Time convolution is used to extract features of ramp scenarios, such as a 50% drop in PV output within 15 minutes.

[0104] Furthermore, the step S3: calculating the attention-weighted features of the spatiotemporal features through the attention mechanism layer.

[0105] This step S3 dynamically adjusts the importance of each node and time step through self-attention and graph attention mechanisms to enhance the robustness of feature representation. Specifically, it includes:

[0106] S31: For the spatiotemporal feature representation H, calculate the attention weight of each node feature in the time dimension based on the self-attention mechanism, and output the time attention weighted feature;

[0107] Self-attention mechanism: Calculate the temporal attention weight matrix for the spatiotemporal feature H :

[0108]

[0109] in, is the trainable weight vector; Represents a splicing operation, It is a double-layer perceptron structure for splicing features; 、 At the source time step 、 The feature vectors of all nodes; LeakyReLU is the activation function; preferably, the temporal attention weight calculation adopts a multi-head mechanism, and the number of multi-heads M ≥ 4;

[0110]

[0111] After dimension transformation, output time attention weighted features , (F′′′ is the attention output dimension), in the temporal attention mechanism, it is calculated at the target time step The temporal attention weighted features of all nodes. The temporal attention module performs weighted aggregation on the original spatiotemporal features H in the time dimension, and The output is N × F''', which is the number of nodes. T is the total number of time steps. For example, T = 12 means 3 hours of data with a step length of 15 minutes.

[0112] This step S31 is based on the self-attention mechanism, focusing on important time step features and amplifying the weights of key time periods, such as the moments when photovoltaic output suddenly drops.

[0113] S32: For the time attention weighted feature, calculate the attention weight of each node and other nodes in the graph structure based on the graph attention mechanism, and output the graph attention feature.

[0114] Graph Attention Mechanism:

[0115] Calculate the spatial attention weight between nodes for the time-weighted feature H′ :

[0116]

[0117] is the projection matrix; is a trainable vector; Representation node i The neighbor set of , and the node j With node i Euclidean distance d ij ≤5km; , Node i and j The temporal attention weighted feature vector of .

[0118] Output graph attention weighted features: , dimension preservation: Optionally, the ELU activation function is used for σ in spatial attention.

[0119] This step S32 focuses on other nodes that are highly correlated with the node, focusing on key nodes, such as overloaded grid nodes.

[0120] Furthermore, the step S4: inputs the spatiotemporal features and the graph attention features into a fuzzy control module, processes the uncertainty of new energy output based on fuzzy rules, and generates a dynamic allocation strategy.

[0121] In step S4, the fuzzy control module processes uncertainty information through fuzzy rules to improve the robustness and adaptability of the model. Specifically, it includes:

[0122] S41: Fuzzify the graph attention features and define the fuzzy sets and membership functions of the input / output variables;

[0123] Input: Model feature representation H1 (spatiotemporal feature representation H, graph attention feature ); where H is the lower-level spatiotemporal feature, and Ĥ is the high-level feature weighted by temporal and spatial attention.

[0124] Processing: Operate the input features independently at each node-time step, and define fuzzy sets of input and output variables. Input variables can be node features, such as power production, consumption, load, etc. Output variables can be deployment strategies, such as charging capacity scheduling. Specifically:

[0125] Input variable set ,in The elements contained in the set are specifically characteristics such as power production / load; the input variable set V contains M Elements: , , ..., . M Indicates the total number of input variables. For example, M = 2, in the fuzzy control module, the input variables include = normalized electrical energy production, = Normalized grid load.

[0126] Define the fuzzy sets of each variable: Electric energy production: {very low, low, medium, high, very high}; Grid load: {light load, normal, heavy load}; Charging capacity scheduling: {substantially reduce, reduce, maintain, increase, substantially increase}.

[0127] The input variables are fuzzified using triangular and trapezoidal membership functions. Each input variable is assigned to a number of fuzzy sets. After fuzzification, the fuzzified input variables are output.

[0128] In some optional embodiments,

[0129]

[0130] The parameters ( a,b,c,d ) determined by clustering historical data; x Represents input variables A specific, clear (non-fuzzy) value of , which is the input of the fuzzification process.

[0131] S42: Define fuzzy rules based on expert knowledge and system characteristics, and apply the fuzzy rule base to perform Mamdani reasoning;

[0132] Fuzzy rule base: contains at least 15 rules in the form of:

[0133] Based on expert knowledge and system characteristics, fuzzy rules are defined. Fuzzy rules are in the form of "IF-THEN":

[0134] R i :IF v 1ISA p1 AND v2ISA i2 THENuIS B i

[0135] in: v 1 is the normalized electric energy production ([0,1]); v 2 is the normalized grid load ([0,1]).

[0136] For example: IF Electricity Production IS High AND Grid Load IS Low THEN Charge Scheduling IS Increased

[0137] IF Electricity production IS Low AND Grid load IS High THEN Charge scheduling IS Reduced.

[0138] More specifically, for example: IF v 1 ≥ 0.8 AND v 2≤ 0.3 THEN u = "substantially increased"; IF v 1≤ 0.3 AND v 2 ≥ 0.7 THEN u = “substantially reduced”.

[0139] Normalized benchmark: v 1 (PV output) takes the installed capacity as 1.0, v 2 (Grid load) The rated capacity of the transformer is 1.0.

[0140] The fuzzy rule base contains multiple such rules to cover different combinations of input variables.

[0141] Fuzzy reasoning is performed on the fuzzified input variables, including the following processing:

[0142] (1) Apply the fuzzified input variables to the fuzzy rule base and calculate the compliance of each rule.

[0143] (2) Use the fuzzy reasoning method (Mamdani reasoning) to calculate the fuzzy output. For each rule, determine its contribution to the output variable and output the fuzzified output variable.

[0144] For example, using Mamdani reasoning: R q Calculating activation strength ; R q express q Fuzzy rules q ), Indicates the i’The firing strength of a rule, which is used to measure the degree to which the current input satisfies this rule; :Input variable The value for i’ The rule is Specified fuzzy sets The degree of membership. Representation Rules i’ About variables j’ Conditions (such as IS High in High).

[0145] Output fuzzy set ,in, is the membership function of the aggregated output fuzzy set, which is the synthesis of all rule reasoning results; [] is the maximum value operator, which means that the output fuzzy set is calculated. B' At the point u When the membership degree is i’ The maximum value of the contribution membership at this point. u It is a specific and clear numerical point of the output variable (such as charging capacity scheduling) in its domain.

[0146] Converting continuous variables into fuzzy language rules through the fuzzy rule engine can solve the decision oscillation problem of traditional models at threshold boundaries.

[0147] S43: Defuzzification is performed through the center of gravity method or the maximum membership method to generate control decisions, including charging scheduling instructions.

[0148] Defuzzification is performed on the fuzzified output variables:

[0149] (1) Convert the fuzzy output variables into precise values. Methods include the centroid method and the maximum membership method.

[0150] For example, the center of gravity method is used to calculate the exact value:

[0151]

[0152] (2) Generate specific control decisions, such as charging capacity scheduling, based on the defuzzified output. Output the defuzzified feature representation (N, T, F'')

[0153] Output constraint: charging capacity scheduling value , negative values ​​indicate discharge, and positive values ​​indicate charge.

[0154] Multi-objective constraint injection: Introducing hard constraints during defuzzification:

[0155]

[0156] in, is the remaining capacity of the transformer, is the current output power of the energy storage unit, and SOC is the current state of charge of the energy storage unit; is the rated capacity of the battery.

[0157] Furthermore, step S5: implements distributed model training based on the federated learning framework, and dynamically updates model parameters through the attention mechanism.

[0158] The federated learning module in step S5 collaboratively trains the model on multiple distributed nodes (such as charging stations and photovoltaic power plants in different regions) to protect data privacy and improve the generalization ability of the model. The specific processing steps are as follows:

[0159] S51: Each distributed node k' uses local data to train a local model. Each node calculates its own loss function and updates the local model parameters according to the gradient of the loss function.

[0160]

[0161] in, , ) is based on the node k' Graph attention weighted features prediction models; is the learning rate, r is the local training round and r ≥3; For nodes k' The local dataset of , contains its own feature sequence and corresponding labels. k' Only use Self-Festival point The feature sequence is trained locally.

[0162] The purpose of step S51 is local model training. Each node trains the model independently, extracts features using local data and updates model parameters. The graph attention weighted features of each distributed node (such as charging stations and photovoltaic power stations in different areas) are processed. Local data training: Each node uses its own local data for model training. Assume that each node has a local model , the model is trained based on graph attention weighted features. Each node calculates its own loss function. Common loss functions include mean squared error (MSE) and cross entropy loss. Each node updates the local model parameters according to the gradient of the loss function; the final output is the local model parameters .

[0163] S52: The federated learning server aggregates the local model parameters of each node to generate global model parameters.

[0164] The federated learning server aggregates the model parameters of each node to generate a global model, ensuring the consistency and collaboration of the model on each node. Specifically, each node will aggregate its trained local model parameters. Upload to the federated learning server. The federated learning server aggregates the model parameters uploaded by all nodes to generate global model parameters. The aggregation method is federated averaging, which is formulated as follows:

[0165]

[0166] in, K is the number of nodes participating in training; R The total number of preset local training rounds; are the global model parameters for output.

[0167] When nodes upload model parameter updates, Paillier homomorphic encryption is used. That is, the server aggregates the updates in an encrypted state, satisfying E(Δθ1)+E(Δθ2)=E(Δθ1+Δθ2). Here, E(·) is the Paillier encryption function, and Δθ1 and Δθ2 are the model parameter gradients uploaded by different nodes.

[0168] S53: Send the global model parameters to each node to synchronously update the local model, and output the spatiotemporal features after federated learning update;

[0169] This step S53 is used to distribute the global model parameters to each node, ensuring that the model of each node is updated synchronously, thereby improving the generalization ability and performance of the model.

[0170] Specifically, the federated learning server sets the global model parameters Distributed to each node, each node uses the global model parameters Update its local model ,Right now:

[0171]

[0172] In the federated learning module: local model training uses differential privacy technology and adds Gaussian noise to the gradient update process; model aggregation uses a weighted average strategy, with weights dynamically allocated based on the amount of node data. Specifically:

[0173] Differential privacy protection: adding noise to local training gradients:

[0174]

[0175] in, is the clipping threshold, the differential privacy formula , δ is the privacy budget and δ∈(0.1,1.0], the smaller the value, the stronger the privacy protection.

[0176] After the model parameters are dynamically updated in step S5, the updated feature representation of the model is output.

[0177] Furthermore, the step S6: integrating the node geographical location and connection relationship through the physical information fusion layer to output enhanced features.

[0178] This step S6 combines the physical information of the nodes to further enhance the feature representation and improve the prediction accuracy of the model. The updated feature representation and the physical information of the nodes (such as geographic location, connection relationship, etc.) are input into the physical information fusion layer. By fusing the dynamic spatiotemporal features with the physical information, the node feature representation is further enriched and the fused features are output. The fused features are then enhanced with the physical information to enhance the relationships and dependencies between nodes and output the enhanced features. Specifically, it includes:

[0179] S61: Encode the physical information of the node into a physical constraint matrix ,Wherein, the physical information includes: grid impedance parameters, line capacity ,limitations;

[0180] ;

[0181] ,in is the line impedance;

[0182] / Max , normalized line capacity;

[0183] The dimension of the physical constraint matrix P is D=3 (it must include distance / impedance / capacity at the same time). P is a three-dimensional tensor, P[ i, j, p ]Store p Physical constraint features at nodes i and j Values ​​between ( p =0, 1, 2 correspond to the inverse of distance, impedance index, and normalized capacity, respectively. , , .

[0184] S62: Fusing the spatiotemporal features updated by the federated learning with the physical constraint matrix through a feature crossover operation to obtain a fused feature;

[0185]

[0186] in: Updated features for federated learning; is the trainable projection matrix; ⊕ is the Hadamard product (element-wise multiplication); fusion features .

[0187] S63: For the fusion feature, enhance the relationship and dependency between nodes through physical information, and output enhanced features.

[0188] Physical relationship propagation:

[0189] in, is the set of neighbor nodes, is the value transformation matrix. The final output is the enhanced feature .

[0190] Furthermore, the step S7: outputting the grid load forecast result based on the enhanced features through the fully connected layer and activation function, and dynamically regulating the energy storage system charging and discharging strategy, includes:

[0191] The enhanced feature representation output by the physical information fusion layer is input into the fully connected layer, and a three-channel prediction output is generated through linear transformation, including charging and discharging power instructions, energy storage unit selection instructions, and grid interaction confidence.

[0192] Enhanced Features Perform the following operations:

[0193]

[0194] in: represents the enhanced feature input; is the weight matrix, Flatten(⋅) represents the flattening operation, N×T×F′′′ Flattened to N ×( T × F′′′ ), Represents the bias vector; Reshape(⋅) represents reconstruction , used to achieve multi-target prediction, converting three-dimensional features into linear decoding of three-channel instructions, corresponding to N nodes in T Three prediction targets for each time step to achieve three-channel prediction:

[0195] Channel 1: Charge and discharge power setting value of each node ; Negative values ​​indicate discharge power, and positive values ​​indicate charge power.

[0196] Channel 2: Energy storage unit selection command , M The total number of energy storage units; One-Hot coding, 1 means the energy storage unit is activated, and 0 means it is closed.

[0197] Channel 3: Grid interaction confidence , used to quantify the reliability of scheduling instructions, for example, 0.8 represents high reliability and 0.3 represents low reliability.

[0198] Each energy storage unit is an independent node (the number of nodes N includes the energy storage unit node). For the energy storage unit node, channel 1 outputs its charge and discharge power ( ), channel 2 outputs its own activation state ( Scalar 0 / 1, which can be the switch instruction of the node itself), channel 3 outputs confidence.

[0199] The physical constraint activation function includes the charge and discharge instruction activation function, the energy storage selection instruction activation and the confidence activation; and the charge and discharge instructions are corrected in real time based on the safety constraints.

[0200] Among them, the charge and discharge instruction activation (using a five-segment activation function):

[0201]

[0202] Among them, | x When |<0.2, the output is 0 to avoid frequent start and stop of the equipment; the slope of 2.5 is used to achieve smooth power regulation. x |>0.8 when full power charging and discharging to protect the device from overload.

[0203] Energy storage selection instruction activation (using Sigmoid-Argmax-OneHot cascade activation):

[0204]

[0205] Among them, Sigmoid(x) is used to convert to a probability value, argmax is used to select the energy storage unit with the highest probability, and OneHot is used to generate a device control signal (for example, [0,1,0] indicates starting energy storage unit No. 2), ensuring that only one energy storage unit is activated at the same time.

[0206] Confidence activation (using sinusoidal periodic activation):

[0207]

[0208] The three-channel separation design can avoid the instruction conflict of the traditional single output channel (such as requiring charging and discharging at the same time).

[0209] In some embodiments, the method further includes: modifying the charge and discharge instructions in real time based on the safety constraints:

[0210]

[0211] Among them, the first It is a charging safety constraint used to implement transformer overload protection. The charging upper limit is dynamically bound to the grid status: ,in is the remaining capacity of the transformer, i.e. the upper limit of the power that the transformer can currently carry. is the real-time total load of the power grid; the second It is a discharge safety constraint and a deep protection for battery discharge. SOC is the current state of charge of the energy storage unit. is the rated capacity of the battery; It is the first channel tensor output by the model, representing the charge and discharge power instructions of each node (range [-1, 1]); It is the actual charge and discharge power instruction (mapped to physical units) after being processed by the activation function. For example, the remaining capacity of the transformer =100kW, current load =80kW, the maximum allowed charging power is min(command value, 100-80=20kW).

[0212] Figure 4 This is a schematic diagram of the grid dispatching process executed by the energy management module of the energy storage system according to an embodiment of the present invention. The output layer also serves as the execution layer, and the execution layer specifically executes the operation process as follows Figure 4 As shown, it is used to optimize the allocation of charging resources, give priority to the use of photovoltaic power for charging, and intelligently allocate grid power during peak periods, thereby maximizing green energy and minimizing operating costs.

[0213] The above-mentioned embodiment of the present invention adopts a three-layer architecture of data-driven, knowledge-guided and physical constraints. The data layer constructs a spatiotemporal graph structure; the algorithm layer integrates federated learning and fuzzy control; and the decision layer embeds safety constraint rules to form a closed-loop control system. By processing uncertain information through fuzzy control, spatiotemporal correlations are effectively captured, and adaptability to dynamic environments is improved. Photovoltaic power is preferentially used for battery charging, reducing dependence on fossil energy and maximizing green energy utilization. Physical constraints are comprehensively considered to improve the real-world adaptability of the model. Through intelligent power forecasting and allocation, power demand is minimized during peak power grid periods, reducing power costs and load pressure. Data privacy is protected through federated learning, and distributed training is achieved. By optimizing the use of charging stations and energy storage systems, the accuracy of the power forecasting algorithm is improved, and the energy utilization efficiency of the entire system is improved.

[0214] The experimental equipment configuration table used in this embodiment is shown in Table 1, and the experimental parameter configuration table used in this embodiment is shown in Table 2.

[0215] Table 1: Experimental equipment configuration table

[0216]

[0217] Table 2: Experimental parameter configuration table

[0218]

[0219] Based on the above experimental verification and this experimental configuration, the technical solution provided by the embodiments of this invention achieves: spatiotemporal prediction accuracy of MAE ≤ 5%; privacy protection strength: meeting differential privacy of ϵ ≤ 1.0; scheduling real-time: control instruction generation < 200ms; and green energy utilization: > 80%. This invention provides a solution for accurate prediction, privacy protection, and secure scheduling in power grids with a high proportion of new energy.

[0220] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the description of the methods.

[0221] It should also be noted that, in the embodiments of the present application, relational terms such as first and second, etc. are merely used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply the existence of any such actual relationship or order between these entities or operations. Moreover, the terms "comprise", "include" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or device comprising 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, article or device. In the absence of further restrictions, the elements defined by the statement "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device comprising the elements.

[0222] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined in the embodiments of the present application may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown in the embodiments of the present application, but rather will conform to the widest scope consistent with the principles and novel features disclosed in the embodiments of the present application.

Claims

1. A method for grid deployment of energy storage system based on dynamic spatiotemporal graph convolutional neural network, characterized in that: include: Step S1: The data input layer obtains time series data of the power grid nodes, including: Constructing a three-dimensional tensor (N, T, F) based on the time series data of the grid nodes, where N is the number of nodes, T is the number of time steps, and F is the number of features. The nodes include photovoltaic power stations, charging stations, and grid access points, and the features include at least power production, consumption, and load; Step S2: The dynamic spatiotemporal graph convolution layer uses a dynamic spatiotemporal graph convolutional neural network to sequentially perform graph convolution layer operations and temporal convolution operations to extract the spatiotemporal features of the time series data; Step S3: Calculating the attention-weighted features of the spatiotemporal features through the attention mechanism layer; Step S4: Input the spatiotemporal features and the attention weighted features into a fuzzy control module, process the uncertainty of new energy output based on fuzzy rules, and generate a dynamic allocation strategy, including: Fuzzify the graph attention features and define the fuzzy sets and membership functions of the input / output variables; Define fuzzy rules based on expert knowledge and system characteristics, and apply the fuzzy rule base to perform Mamdani reasoning; Defuzzification is performed to generate control decisions, including charging scheduling instructions, using the center of gravity method or the maximum membership method; Step S5: Implement distributed model training based on the federated learning framework and dynamically update model parameters through the attention mechanism; Step S6: Integrate the node geographic location and connection relationship through the physical information fusion layer and output enhanced features; Step S7: Based on the enhanced features, through the fully connected layer and activation function, output the grid load forecast result, and dynamically adjust the charging and discharging strategy of the energy storage system, including: The enhanced feature representation output by the physical information fusion layer is input into the fully connected layer, and a three-channel prediction output is generated through linear transformation, including charging and discharging power instructions, energy storage unit selection instructions, and grid interaction confidence. The physical constraint activation function includes the charge and discharge instruction activation function, the energy storage selection instruction activation and the confidence activation; and the charge and discharge instructions are corrected in real time based on the safety constraints.

2. The energy storage system grid deployment method based on dynamic spatiotemporal graph convolutional neural network according to claim 1 is characterized in that: in, The operations performed by the dynamic spatiotemporal graph convolution layer in step S2 include: Dynamic graph construction: A dynamic graph structure is constructed based on the physical connection relationships and spatiotemporal dependencies between nodes. The dynamic graph construction includes: determining the physical connection relationships between nodes based on the power grid topology; quantifying the spatiotemporal dependencies between nodes using the Pearson correlation coefficient; and generating a sparse adjacency matrix that changes over time, with non-zero elements indicating valid connections. Graph convolution operation: for the dynamic graph structure and the three-dimensional tensor (N, T, F), the spatial features of each time step are extracted using graph convolution operation; Temporal convolution operation: Perform a one-dimensional temporal convolution operation on the spatial features along the time dimension and output a spatiotemporal feature representation.

3. The energy storage system grid deployment method based on dynamic spatiotemporal graph convolutional neural network according to claim 2 is characterized in that: The step S3: calculating the attention-weighted features of the spatiotemporal features through the attention mechanism layer, includes: For the spatiotemporal feature representation, the attention weight of each node feature in the time dimension is calculated based on the self-attention mechanism, and the time attention weighted feature is output; For the temporal attention weighted feature, the attention weights of each node and other nodes in the graph structure are calculated based on the graph attention mechanism, and the graph attention feature is output.

4. The energy storage system grid deployment method based on dynamic spatiotemporal graph convolutional neural network according to claim 3 is characterized in that: Step S5: implementing distributed model training based on the federated learning framework and dynamically updating model parameters through the attention mechanism, including: Each distributed node uses local data to train a local model. Each node calculates its own loss function and updates the local model parameters according to the gradient of the loss function. The federated learning server aggregates the local model parameters of each node to generate global model parameters; Send the global model parameters to each node to synchronously update the local model, and output the updated spatiotemporal features of federated learning; Among them, in the federated learning module: local model training adopts differential privacy technology and adds Gaussian noise to the gradient update process; model aggregation adopts a weighted average strategy, and weights are dynamically allocated according to the amount of node data.

5. The energy storage system grid deployment method based on dynamic spatiotemporal graph convolutional neural network according to claim 4 is characterized in that: Step S6: integrating the node geographic location and connection relationship through the physical information fusion layer to output enhanced features, including: Encoding the physical information of the node into a physical constraint matrix, wherein the physical information includes: grid impedance parameters and line capacity limitations; fusing the spatiotemporal features updated by the federated learning with the physical constraint matrix through a feature crossover operation to obtain a fused feature; With respect to the fusion features, the relationships and dependencies between nodes are enhanced through physical information, and enhanced features are output.

6. An energy storage system, characterized in that: include: Photovoltaic power generation module, including photovoltaic panels and inverters, used to collect solar energy and convert it into electrical energy and convert direct current into alternating current for system use; Energy storage module, including lithium-ion battery pack and bidirectional charge and discharge controller, used to store excess photovoltaic power and support peak power output; Charging stations equipped with a multi-protocol intelligent charging interface and a battery management module to support fast charging of multiple electric vehicle battery specifications and monitor the status of each lithium-ion battery pack; An energy management module is used to monitor the production of photovoltaic power, grid load, energy storage status, and charging needs of charging stations in real time, and to execute the energy storage system grid dispatching method based on dynamic spatiotemporal graph convolutional neural network as described in any one of claims 1-5, optimize the allocation of charging resources, give priority to the use of photovoltaic power for charging, and intelligently dispatch grid power during peak periods.

Images