Energy storage system power grid allocation method based on dynamic space-time diagram convolutional neural network

Abstract

The invention provides an energy storage system power grid allocation method based on a dynamic space-time diagram convolutional neural network. The energy storage management method comprises the following steps: S1, acquiring time sequence data of a power grid node; s2, sequentially executing a graph convolution layer operation and a time convolution operation by using a dynamic space-time graph convolution neural network, and extracting space-time features of the time sequence data; s3, calculating attention weighted features of the spatio-temporal features; s4, processing the new energy output uncertainty based on a fuzzy rule by using the spatial-temporal characteristics and the attention weighted characteristics; s5, realizing distributed model training based on a federated learning framework, and dynamically updating model parameters through an attention mechanism; s6, integrating the geographic positions and the connection relations of the nodes, and outputting enhanced features; and S7, the enhanced features pass through a full connection layer and an activation function, a power grid load prediction result is output, and a charging and discharging strategy of the energy storage system is dynamically regulated and controlled. Power supply and demand dynamic balance in a new energy high-permeability scene is realized, and the operation cost and the environmental influence are reduced.

Description

Technical Field

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

[0002] An energy system dominated by new energy is a future development feature, but at the same time, it also brings huge pressure to power regulation. Insufficient peak shaving capacity will lead to curtailment of wind and solar power. In a power system dominated by new energy, it will be very difficult to meet the system regulation requirements solely relying on controllable units. The dispatchability of a certain flexibility resource is mainly restricted by its own ramping rate, and the dispatchability of these resources further constitutes the responsive flexibility demand within the system.

[0003] Under the scenario of high penetration of new energy, the traditional power dispatch system exposes the following deficiencies: The output of new energy has typical spatio-temporal imbalance characteristics. The minute-level fluctuation of photovoltaic output exceeds 50% of the installed capacity, resulting in a serious shortage of traditional peak shaving capacity and a curtailment rate of wind and solar power as high as 15%-20%, causing waste of resources; Difficulty in multi-agent collaboration: The data of flexibility resources such as distributed charging stations and energy storage power stations are isolated as data islands, and it is difficult for a centralized dispatch model to balance privacy protection and global optimization. The centralized optimization method relying on global data aggregation has a risk of privacy leakage. The traditional graph neural network uses a fixed adjacency matrix and cannot capture the spatio-temporal dynamic correlation of new energy output; In addition, when dealing with new energy fluctuations using traditional PID control or rule engines, the number of equipment starts and stops increases by 3-5 times, significantly shortening the service life of the energy storage system.

[0004] To solve the above problems, the present invention proposes an intelligent deployment method that integrates dynamic spatio-temporal graph modeling, federated learning privacy protection, and fuzzy uncertainty processing, constructs a three-layer collaborative architecture of data-driven, knowledge-guided, and physical constraints, and realizes the dynamic balance of power supply and demand under the scenario of high penetration of new energy. Summary of the Invention

[0005] To address the above problems, the present invention proposes a dynamic spatio-temporal graph convolutional neural network algorithm FC-FL-AM-DST-GCN (Dynamic spatial-temporal graph convolutional neural networks based on fuzzy control, federated learning and attention mechanism). Combining with an energy storage system example for photovoltaic renewable energy charging, the FC-FL-AM-DST-GCN algorithm can monitor and analyze the supply situation of various energy sources and grid loads in real time, especially applicable to the dynamic balance of power supply and demand in scenarios with a high proportion of new energy access. It can intelligently allocate charging resources to ensure charging with green electricity during peak periods, reducing operating costs and environmental impacts.

[0006] To achieve the above object, it is realized through the following technical solutions: In a first aspect, the present invention provides a method for grid allocation of an energy storage system based on a dynamic spatio-temporal graph convolutional neural network, including: Step S1: The data input layer obtains the time series data of grid nodes; Step S2: The dynamic spatio-temporal graph convolutional layer uses a dynamic spatio-temporal graph convolutional neural network (DST-GCN, Dynamic Spatial-Temporal Graph Convolutional Network) to sequentially perform graph convolutional layer operations and time convolutional operations to extract the spatio-temporal features of the time series data; Step S3: The attention mechanism layer calculates the attention weighted features of the spatio-temporal features; Step S4: Input the spatio-temporal features and the attention weighted features into the fuzzy control module to handle the uncertainty of new energy output based on fuzzy rules; Step S5: Implement distributed model training based on the federated learning framework and dynamically update the model parameters through the attention mechanism; Step S6: Integrate the node geographical location and connection relationship through the physical information fusion layer to output enhanced features; Step S7: Based on the enhanced features, through the fully connected layer and activation function, output the grid load prediction result and dynamically adjust the charge and discharge strategy of the energy storage system.

[0007] Furthermore, in step S1, it further includes: Construct a three-dimensional tensor (N, T, F) based on the time series data of the power 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 power grid access points, and the features at least include power production, consumption, and load.

[0008] Further, the operations performed by the dynamic spatio-temporal graph convolutional layer in step S2 include: Dynamic graph construction: Construct a dynamic graph structure based on the physical connection relationship and spatio-temporal dependence relationship between nodes, where: the physical connection relationship determines the physical connectivity between nodes through the power grid topology; the spatio-temporal dependence relationship is determined by the Pearson correlation coefficient. Graph convolution operation: For the dynamic graph structure and the three-dimensional tensor (N, T, F), use graph convolution operations to extract the spatial features at each time step. Temporal convolution operation: Perform a one-dimensional temporal convolution operation on the spatial features along the time dimension to output the spatio-temporal feature representation.

[0009] Further, the dynamic graph construction includes: Determine the physical connection relationship between nodes based on the power grid topology. Quantify the spatio-temporal dependence between nodes through the Pearson correlation coefficient. Generate a time-varying sparse adjacency matrix, where non-zero elements represent valid connections.

[0010] Further, in step S3: Calculate the attention-weighted features of the spatio-temporal features through the attention mechanism layer, including: For the spatio-temporal feature representation, calculate the attention weights of each node feature in the time dimension based on the self-attention mechanism, and output the time attention-weighted features. For the time attention-weighted features, calculate the attention weights of each node and other nodes in the graph structure based on the graph attention mechanism, and output the graph attention features.

[0011] Further, in step S4: Input the spatio-temporal features and the graph attention features into the fuzzy control module, and process the uncertainty of new energy output based on fuzzy rules to generate a dynamic dispatching strategy, including: Fuzzify the graph attention features, and define the fuzzy sets and membership functions of the input / output variables. Define fuzzy rules according to expert knowledge and system characteristics, and apply the fuzzy rule base to perform Mamdani reasoning. Through the centroid method or the maximum membership degree method, defuzzify to generate control decisions, including charging scheduling instructions.

[0012] Further, in step S5: Implement distributed model training based on the federated learning framework, and dynamically update the 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 each node 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. The global model parameters are sent to each node to synchronously update the local model, and the spatio-temporal features after federated learning update are output. Among them, in the federated learning module: Differential privacy technology is used in local model training, and Gaussian noise is added to the gradient update process; Weighted average strategy is used in model aggregation, and the weights are dynamically allocated according to the node data volume.

[0013] Further, in step S6: Integrate the node geographical location and connection relationship through the physical information fusion layer, and output enhanced features, including: Encode the physical information of the node into a physical constraint matrix, where the physical information includes: power grid impedance parameters, line capacity limits. Fuse the spatio-temporal features after federated learning update and the physical constraint matrix through feature cross-operation to obtain fused features. For the fused features, enhance the relationship and dependence between nodes through physical information, and output enhanced features.

[0014] Further, in step S7: Based on the enhanced features, through the fully connected layer and activation function, output the power grid load prediction result, and dynamically regulate the charge and discharge strategy of the energy storage system, including: Input the enhanced feature representation output by the physical information fusion layer into the fully connected layer, and generate a three-channel prediction output through linear transformation, including charge and discharge power instructions, energy storage unit selection instructions, and power grid interaction confidence. The physical constraint activation function includes charge and discharge instruction activation function, energy storage selection instruction activation, and confidence activation; and, the charge and discharge instructions are corrected in real time based on safety constraints.

[0015] In a second aspect, the present invention also provides an energy storage system, including: A photovoltaic power generation module, including photovoltaic panels and an inverter, for collecting solar energy, converting it into electrical energy, and converting direct current into alternating current for system use. An energy storage module, including a lithium-ion battery pack and a bidirectional charge and discharge controller, for storing excess photovoltaic electrical energy and supporting peak power output. A charging station equipped with a multi - protocol intelligent charging interface and a battery management module, which are respectively used to support fast charging of multiple electric vehicle battery specifications and monitor the status of each lithium - ion battery pack; An energy management module, which is used to monitor the production of photovoltaic electric energy, grid load, energy storage status and charging demand of the charging station in real time, and execute the energy storage system grid deployment method based on the dynamic spatio - temporal graph convolutional neural network as described in the first aspect, optimize the allocation of charging resources, preferentially use photovoltaic electric energy for charging, and intelligently allocate grid electric energy during peak periods.

[0016] Compared with the prior art, the present invention has the following beneficial effects: 1. Maximize the utilization of green energy: In a high - proportion new - energy scenario covered by the ramp - up scenario, by using fuzzy control to process uncertain information, effectively capture spatio - temporal correlation, improve the adaptability to the dynamic environment, preferentially use photovoltaic electric energy for battery charging, and reduce the dependence on fossil energy.

[0017] 2. Reduce peak power demand: Comprehensively consider physical constraints and improve the real - world adaptability of the model. Through the prediction and allocation of intelligent power, minimize power demand during grid peak periods, and reduce power costs and load pressure.

[0018] 3. Improve system efficiency and synergy: Through the federated learning framework, realize distributed model training and collaborative optimization on the premise of protecting the data privacy of each participating party (photovoltaic power station, charging station, grid); by integrating physical constraints and fuzzy control, significantly improve the accuracy of power prediction and the practical feasibility of dispatching instructions, thereby overall improving energy utilization efficiency and reducing system operation costs.

[0019] It should be understood that the content described in the summary of the invention section is not intended to limit the key or important features of the embodiments of the present invention, nor is it used to limit the scope of the present invention. Other features of the present invention will become easily understood through the following description. Brief Description of the Drawings

[0020] Combined with the drawings and referring to the following detailed description, the above - mentioned and other features, advantages and aspects of the embodiments of the present invention will become more obvious. The drawings are used to better understand the solution and do not constitute a limitation to the present invention. In the drawings, the same or similar reference numerals represent the same or similar elements, where: Figure 1 is a schematic flowchart of the energy storage system grid deployment method based on the dynamic spatio - temporal graph convolutional neural network according to the embodiment of the present invention; Figure 2 is a flowchart of the network dynamic spatio - temporal graph convolutional layer according to the embodiment of the present invention; Figure 3 is a schematic diagram of the modules of the energy storage system according to the embodiment of the present invention; Figure 4 It is a schematic diagram of the grid deployment process executed by the energy management module of the energy storage system according to an embodiment of the present invention. Specific Embodiments

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Apparently, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0022] In addition, the term "and / or" in this article is merely a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist simultaneously, and B exists alone. In addition, the character " / " in this article generally represents an "or" relationship between the associated objects before and after.

[0023] Embodiment 1: Figure 1 It shows a schematic diagram of the grid deployment method of the energy storage system based on the dynamic spatio-temporal graph convolutional neural network according to an embodiment of the present invention. As Figure 1 shown, the grid deployment method of the energy storage system based on the dynamic spatio-temporal graph convolutional neural network is applied to the energy storage system 200 as Figure 3 shown, and includes the following steps: Step S1: The data input layer obtains the time series data of the grid nodes; Step S2: The dynamic spatio-temporal graph convolutional layer uses the dynamic spatio-temporal graph convolutional neural network (DST-GCN) to sequentially perform graph convolutional layer operations and time convolutional operations to extract the spatio-temporal features of the time series data; Step S3: Calculate the attention weighted features of the spatio-temporal features through the attention mechanism layer; Step S4: Input the spatio-temporal features and the attention weighted features into the fuzzy control module, and process the uncertainty of new energy output based on fuzzy rules to generate a dynamic deployment strategy; Step S5: Implement distributed model training based on the federated learning framework, and dynamically update the model parameters through the attention mechanism; Step S6: Integrate the node geographical locations and connection relationships through the physics-information fusion layer to output enhanced features; Step S7: Based on the enhanced features, pass through the fully connected layer and the activation function to output the grid load prediction result, and dynamically adjust the charge and discharge strategy of the energy storage system.

[0024] Figure 3It is a schematic diagram of the system architecture of an energy storage system 200 according to an embodiment of the present invention. As Figure 3 shown, an energy storage system 200 according to an embodiment of the present invention is used to implement a method for dispatching an energy storage system power grid based on a dynamic spatio-temporal graph convolutional neural network. The energy storage system 200 includes: A photovoltaic power generation module 210, including photovoltaic panels and an inverter, is used to collect solar energy, convert it into electrical energy, and convert direct current into alternating current for system use; Further, the photovoltaic panels are installed in or near the charging station to collect solar energy and convert it into electrical energy; the inverter converts the direct current generated by the photovoltaic panels into alternating current for system use.

[0025] An energy storage module 220, including a lithium-ion battery pack and a bidirectional charge and discharge controller, is used to store excess photovoltaic electrical energy and support peak power output; Further, the battery pack is used to store excess photovoltaic electrical energy; the charge and discharge controller ensures that green electrical energy can also be used at night or on cloudy days.

[0026] A charging station 230, equipped with a multi-protocol intelligent charging interface and a battery management module; Further, the intelligent charging interface supports the quick replacement and charging of various electric vehicle battery specifications; the battery management module monitors the status of each battery, including the charging status, health status, and temperature, to ensure safe and efficient charging.

[0027] An energy management module 240: The energy management module is used to monitor the production of photovoltaic electrical energy, grid load, energy storage status, and charging demand of the charging station in real time, and has an FC-FL-AM-DST-GCN algorithm scheduling core built in to execute a method for dispatching an energy storage system power grid based on a dynamic spatio-temporal graph convolutional neural network, optimize the allocation of charging resources, preferentially use photovoltaic electrical energy for charging, and intelligently dispatch grid electrical energy during peak periods.

[0028] Further, step S1: The data input layer obtains the time series data of the grid nodes, including: Input data: including the time series data of different nodes (such as photovoltaic power generation stations, charging stations, power grids, etc.). The features of each node include electrical energy production, consumption, load, etc. 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. 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. The nodes include photovoltaic power generation stations, charging stations, and grid access points, and the features at least include electrical energy production, consumption, and load.

[0029] Further, as Figure 2As shown in the figure, it is the flowchart of the network dynamic spatio-temporal graph convolution layer of the embodiment of the present invention. Among them, Figure 2 the numbers 1, 2, 3, ……, 34, 35 in it respectively represent different nodes of the power grid topological structure. The processing of the dynamic spatio-temporal graph convolution layer in step S2 includes: S21: Dynamic graph construction: Construct a dynamic graph structure (N, N, T) based on the physical connection relationship and spatio-temporal dependence relationship between nodes; This can be achieved by constructing an adjacency matrix, where the connection between each node and its adjacent nodes (based on physical location or correlation) is represented as 1, otherwise 0. The adjacency matrix A is (N, N), representing the connection relationship between nodes.

[0030] Abstract photovoltaic power stations, charging stations, and power grid nodes as graph nodes, and the adjacency matrix A(t) dynamically encodes two types of relationships: Physical connection (such as cable topology), where the physical connection relationship determines the physical connectivity between nodes through the power grid topological structure; Spatio-temporal correlation (such as regional lighting similarity), quantifying the spatio-temporal dependence between nodes through the Pearson correlation coefficient:

[0031] represents the node i and the node j of the Pearson correlation coefficient, used to quantify the node i and the node j of the time series similarity, W represents the sliding time window (corresponding to 3 hours of historical data, step size 15 minutes, covering the new energy fluctuation period), and the value can be W = 12; represents the node i at the moment τ of the eigenvalue (such as photovoltaic output), represents the node i in the time window of the feature mean, represents the node j in the time window of the feature mean.

[0032] The adjacency matrix A of the element :

[0033] Among them, when it is 1, it means that the node i and the node j at time t there is a connection, and when it is 0, it means that the node i and the node j at timet There is no connection; Represents an element of the physical connection adjacency matrix, where, Represents a node i And node j Establish a physical connection through direct cable connection; when > 0.7, a connection is established; The sliding time window is 15 minutes, that is, it is calculated every 15 minutes , and update the adjacency matrix A.

[0034] Based on the physical connection (hard connection) of the power grid topology and the dynamic establishment of spatio-temporal dependence (soft connection) through the correlation coefficient threshold (0.7), the connection relationship is quantified, thereby generating a time-varying sparse adjacency matrix, and non-zero elements represent effective connections.

[0035] S22: Graph convolution operation: For the dynamic graph structure (N, N, T) and the three-dimensional tensor (N, T, F), use the graph convolution operation to extract the spatial features at each time step;

[0036] Where: Represents after the l th layer of graph convolution operation, the l +1 layer at time step t Output feature matrix; Represents the feature representation of the l th layer, and its 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 optionally used; Perform graph convolution operations on each time step, and output the spatial features (N, T, F') at each time step. F' is the feature dimension after graph convolution. Graph convolution is used to capture the spatial dependence of nodes, such as the coordinated change of the output power of adjacent photovoltaic stations.

[0037] S23: Temporal convolution operation: Perform a one-dimensional temporal convolution operation on the spatial features along the time dimension, and output the spatio-temporal feature representation.

[0038] Perform a one-dimensional convolution operation on the spatial features in the time dimension to extract temporal features.

[0039] Temporal convolution operation formula:

[0040] Where: Represents after thel After the temporal convolution operation of the layer, at the l output of the +1 layer, at time step t of all node feature representations, which is a matrix (number of nodes N × feature dimension F''); represents the feature representation at time step t-k after the graph convolution operation; is the weight matrix of the temporal convolution, and the temporal convolution kernel K takes 3.

[0041] Perform a temporal convolution operation on each node to output a spatio-temporal feature representation , F'' is the feature dimension after temporal convolution. For example, F" = 64 is selected. Temporal convolution is used to extract ramp scenario features, such as a 50% sudden drop in PV output in 15 minutes.

[0042] Furthermore, in step S3: Calculate the attention-weighted features of the spatio-temporal features through the attention mechanism layer.

[0043] This step S3 dynamically adjusts the importance of each node and time step through self-attention and graph attention mechanisms, enhancing the robustness of the feature representation. Specifically, it includes: S31: For the spatio-temporal feature representation H, calculate the attention weights of each node feature in the time dimension based on the self-attention mechanism, and output the time attention-weighted features; Self-attention mechanism: Calculate the time attention weight matrix for the spatio-temporal feature H :

[0044] Among them, is a trainable weight vector; represents the concatenation operation, is a two-layer perceptron structure for the concatenated features; , are the feature vectors of all nodes at the source time step , respectively; LeakyReLU is the activation function; Preferably, the time attention weight calculation adopts the multi-head mechanism, and the number of multi-heads M ≥ 4;

[0045] After dimensional transformation, output the time attention-weighted features , (F''' is the attention output dimension), in the time attention mechanism, it is the calculated time attention-weighted features of all nodes at the target time step . is the result of the time attention module weighted and aggregated on the original spatio-temporal feature H in the time dimension, at the moment The output. The dimension is the number of nodes N × feature dimension F'''; T is the total number of time steps. For example, T = 12 represents 3 hours of data with a step size of 15 minutes.

[0046] This step S31 is based on the self-attention mechanism, focusing on important time-step features, amplifying the weights of key time periods, such as the moment when the photovoltaic output drops suddenly.

[0047] S32: For the time-attention weighted features, calculate the attention weights of each node with other nodes in the graph structure based on the graph attention mechanism, and output the graph attention features.

[0048] Graph attention mechanism: Calculate the spatial attention weights between nodes for the time-weighted feature H'. :

[0049] is the projection matrix; is the trainable vector; represents the node i 's neighbor set, defined as , and the node j and the node i Euclidean distance d ij ≤ 5km; , are the time-attention weighted feature vectors of nodes i and j respectively.

[0050] Output the graph-attention weighted features: , with the dimension remaining: . Optionally, σ in the spatial attention adopts the ELU activation function.

[0051] This step S32 focuses on other nodes with high relevance to this node, focusing on key nodes, such as overloaded power grid nodes.

[0052] Furthermore, in step S4: Input the spatio-temporal features and the graph attention features into the fuzzy control module, and process the uncertainty of new energy output based on fuzzy rules to generate a dynamic allocation strategy.

[0053] In this step S4, the fuzzy control module processes the uncertainty information through fuzzy rules to improve the robustness and adaptability of the model. Specifically, it includes: S41: Fuzzify the graph attention features, and define the fuzzy sets and membership functions of the input / output variables; Input: The feature representation H1 of the model (the spatio-temporal feature representation H, the graph attention feature ); where H is the lower-level spatio-temporal feature, and Ĥ is the high-level feature weighted by temporal and spatial attention.

[0054] Processing: Independently operate on the input features by node-time step, and define the fuzzy sets of the input variables and output variables. The input variables can be node features such as electricity production volume, consumption volume, load, etc. The output variable can be a deployment strategy such as charging volume scheduling. Specifically: Set of input variables , where The elements included in the set are specifically features such as electricity production volume / load, etc.; this set of input variables V contains M elements: , , ..., . M represents the total number of input variables. For example, take M = 2. In the fuzzy control module, the input variables include = normalized electricity production volume, = normalized grid load.

[0055] Define the fuzzy sets of each variable: Electricity production volume: {extremely low, low, medium, high, extremely high}; Grid load: {light load, normal, heavy load}; Charging volume scheduling: {substantially reduce, reduce, maintain, increase, substantially increase}.

[0056] Fuzzify the input variables using membership functions such as triangular and trapezoidal. Each input variable is assigned to several fuzzy sets. After the fuzzification process, the fuzzified input variables are output.

[0057] In some alternative embodiments,

[0058] where the parameter ( a, b, c, d ) is determined by historical data clustering; x represents a specific, clear (non-fuzzy) value of the input variable , which is the input of the fuzzification process.

[0059] S42: Define fuzzy rules based on expert knowledge and system characteristics, and apply the fuzzy rule base to perform Mamdani inference; Fuzzy rule base: Contains at least 15 rules in the form of: Define fuzzy rules according to expert knowledge and system characteristics. The fuzzy rules are in the form of "IF-THEN": R i : IF v 1 ISA p1 AND v2ISA i2 THEN uIS B i Wherein: v 1 is the normalized electric energy production ([0, 1]); v 2 is the normalized grid load ([0, 1]).

[0060] For example: IF the electric energy production IS High AND the grid load IS Low THEN the charging amount scheduling IS increased IF the electric energy production IS Low AND the grid load IS High THEN the charging amount scheduling IS decreased.

[0061] 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 decreased".

[0062] Normalization reference: v 1 (photovoltaic output) is based on an installed capacity of 1.0, v 2 (grid load) is based on a transformer rated capacity of 1.0.

[0063] The fuzzy rule base contains multiple such rules to cover different combinations of input variables.

[0064] For the fuzzified input variables, fuzzy inference is performed, including the following processing: (1) Apply the fuzzified input variables to the fuzzy rule base and calculate the degree of compliance of each rule.

[0065] (2) Use the fuzzy inference method (Mamdani inference) to calculate the fuzzy output. For each rule, determine its contribution degree to the output variable and output the fuzzified output variable.

[0066] For example, using Mamdani inference: for the rule R q Calculate the activation strength ; R q Indicates q the number of fuzzy rules (Rule 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; : The input variable 's value for the i’ th rule for the specified fuzzy set . Represents the condition in rule i’ regarding variable j’ (such as High in IS High).

[0067] Output fuzzy set , where is the membership function of the aggregated output fuzzy set, which is the synthesis of the inference results of all rules; [ ] is the maximum operator, indicating that when calculating the membership degree of the output fuzzy set B' at point u , take the maximum value of the membership degrees contributed by all rules i’ at this point. u is a specific and clear numerical point of the output variable (such as charge amount scheduling) in its universe of discourse.

[0068] Converting continuous variables into fuzzy language rules through a fuzzy rule engine can solve the problem of decision-making oscillation at the threshold boundary in traditional models.

[0069] S43: Through the centroid method or the maximum membership degree method, defuzzification is performed to generate control decisions, including charging scheduling instructions.

[0070] Perform defuzzification processing on the fuzzified output variable: (1) Convert the fuzzy output variable into an exact value. The methods include the centroid method and the maximum membership degree method.

[0071] For example, use the centroid method to calculate the exact value:

[0072] (2) Generate specific control decisions according to the defuzzified output, such as charge amount scheduling. Output the feature representation (N, T, F'') after defuzzification Output constraint: Charge amount scheduling value , a negative value indicates discharging, and a positive value indicates charging.

[0073] Multi-objective constraint injection: Introduce hard constraints during defuzzification:

[0074] Among them, 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.

[0075] Further, in step S5: Implement distributed model training based on the federated learning framework, and dynamically update the model parameters through the attention mechanism.

[0076] The federated learning module in step S5 collaboratively trains the model on multiple distributed nodes (such as charging stations and photovoltaic power stations in different regions), protecting data privacy and improving the generalization ability of the model. The specific processing steps are as follows: S51: Each distributed node k' uses local data to train a local model. Each node calculates its own loss function, and each node updates the local model parameters according to the gradient of the loss function;

[0077] Among them, , ) is the prediction model based on the graph attention weighted feature k' belonging to node ; is the learning rate, r is the number of local training rounds and r ≥ 3; is the local dataset of node k' , including its own feature sequence and corresponding labels. Node k' only uses self - segment point 's feature sequence for local training.

[0078] The purpose of step S51 is local model training. Each node independently trains the model, extracts features using local data, and updates the model parameters. The graph attention weighted features of each distributed node (such as charging stations and photovoltaic power stations in different regions) are processed for local data training: Each node uses its own local data for model training. Assume that each node has a local model , and this model is trained based on the graph attention weighted feature. Each node calculates its own loss function. Common loss functions include mean square error (MSE), cross-entropy loss, etc. Each node updates the local model parameters according to the gradient of the loss function; finally, the local model parameters are output.

[0079] S52: The federated learning server aggregates the local model parameters of each node to generate global model parameters.

[0080] The federated learning server aggregates the model parameters of each node to generate a global model, ensuring the consistency and collaboration of the model across nodes. Specifically, each node uploads its locally trained model parameters 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, and its formula is as follows:

[0081] where, K is the number of nodes participating in training; R is the preset total number of local training rounds; is the output global model parameters.

[0082] Among them, when nodes upload the updated amounts of model parameters, Paillier homomorphic encryption is used, that is, the server aggregates the updated amounts in the encrypted state, satisfying E(Δθ1)+E(Δθ2)=E(Δθ1+Δθ2). Where, E(·) is the Paillier encryption function, and Δθ1, Δθ2 are the gradients of model parameters uploaded by different nodes.

[0083] S53: Send the global model parameters to each node to synchronously update the local model, and output the spatio-temporal features after federated learning update; This step S53 is used to distribute the global model parameters to each node, ensure the synchronous update of the model for each node, and improve the generalization ability and performance of the model.

[0084] Specifically, the federated learning server distributes the global model parameters to each node, and each node uses the global model parameters to update its local model , that is:

[0085] In the federated learning module: differential privacy technology is used in local model training, and Gaussian noise is added to the gradient update process; weighted average strategy is used in model aggregation, and the weights are dynamically allocated according to the amount of node data. Specifically: Differential privacy protection: Adding noise to local training gradients:

[0086] where, is the clipping threshold, and the differential privacy formula , δ is the privacy budget and δ∈(0.1,1.0], and the smaller the value, the stronger the privacy protection.

[0087] After dynamically updating the model parameters in step S5, the feature representation after model update is output.

[0088] Further, in step S6: Integrate the node geographical location and connection relationship through the physical information fusion layer, and output enhanced features.

[0089] This step S6 combines the physical information of the nodes to further enhance the feature representation and improve the prediction accuracy of the model. Input the updated feature representation and the physical information of the nodes (such as geographical location, connection relationship, etc.) into the physical information fusion layer. By fusing the dynamic spatio-temporal features with the physical information, further enrich the node feature representation and output the fused features; then enhance the relationship and dependence between nodes through the physical information for the fused features, and output the enhanced features. Specifically, it includes: S61: Encode the physical information of the nodes into a physical constraint matrix , where the physical information includes: power grid impedance parameters, line capacity limits; ; , where is the line impedance; / Max , normalize the line capacity; Among them, the dimension D of the physical constraint matrix P = 3 (must include distance / impedance / capacity at the same time), P is a three-dimensional tensor, P i, j, p stores the value of the p th physical constraint feature between nodes i and j ( p = 0, 1, 2 correspond to the reciprocal of distance, impedance exponent, and normalized capacity respectively), , , .

[0090] S62: Fuse the spatio-temporal features updated by the federated learning and the physical constraint matrix through feature cross-operation to obtain fused features;

[0091] Among them: is the feature updated by the federated learning; is a trainable projection matrix; ⊕ is the Hadamard product (element-wise multiplication); the fused feature .

[0092] S63: For the fused features, enhance the relationship and dependence between nodes through the physical information, and output enhanced features.

[0093] Physical relationship propagation:

[0094] Among them, is the set of neighbor nodes, is the value transformation matrix. The final output is the enhanced feature .

[0095] Furthermore, in step S7: based on the enhanced feature, through the fully connected layer and the activation function, the power grid load prediction result is output, and the charge and discharge strategy of the energy storage system is dynamically regulated, including: The enhanced feature representation output by the physical information fusion layer is input into the fully connected layer, and three-channel prediction outputs are generated through linear transformation, including charge and discharge power commands, energy storage unit selection commands, and power grid interaction confidence levels; For the enhanced feature the following operations are performed:

[0096] Among them: represents the enhanced feature input; is the weight matrix, Flatten(⋅) represents the flattening operation, which flattens N×T×F′′′ into N ×( T × F′′′ ), represents the bias vector; Reshape(⋅) represents restructuring into , which is used to achieve multi-objective prediction and perform linear decoding of converting three-dimensional features into three-channel instructions, corresponding to three prediction objectives of N nodes at T time steps, to achieve three-channel prediction: Channel 1: The set value of the charge and discharge power of each node ; a negative value represents the discharge power, and a positive value represents the charge power.

[0097] Channel 2: The energy storage unit selection instruction , M is the total number of energy storage units; One-Hot encoding, 1 means activating the energy storage unit, and 0 means turning it off.

[0098] Channel 3: The power grid interaction confidence level , which is used to quantify the reliability of the scheduling instruction. For example, 0.8 represents high reliability, and 0.3 represents low reliability.

[0099] Among them, each energy storage unit is an independent node (the number of nodes N includes the energy storage unit nodes). For the energy storage unit nodes, 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), and Channel 3 outputs the confidence level.

[0100] The physical constraint activation function includes the charge and discharge command activation function, the energy storage selection command activation, and the confidence activation; and, the charge and discharge commands are corrected in real time based on safety constraints.

[0101] Among them, the charge and discharge command activation (using a five-segment activation function):

[0102] Among them, | x | When < 0.2, the output is 0, which can avoid frequent start and stop of the device; the slope of 2.5 is used to achieve smooth power regulation, | x | When > 0.8, full-power charge and discharge can be performed to protect the device from overload.

[0103] The energy storage selection command activation (using a cascaded activation of Sigmoid-Argmax-OneHot):

[0104] 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 (such as [0,1,0] indicating to start the 2nd energy storage) to ensure that only one energy storage unit is activated at the same time.

[0105] The confidence activation (using a sine period activation):

[0106] Through a three-channel separation design, command conflicts in traditional single-output channels (such as simultaneously requiring charging and discharging) can be avoided.

[0107] And in some embodiments, the method further includes: correcting the charge and discharge commands in real time based on safety constraints:

[0108] Among them, the first item is the charging safety constraint, which is used to achieve overload protection of the transformer, and the charging upper limit is dynamically bound to the grid state: , where is the remaining capacity of the transformer, that is, the upper limit of the power that the transformer can currently carry, is the real-time total load of the grid; the second item is the discharge safety constraint, which is a deep protection for battery discharge, and SOC is the current state of charge of the energy storage unit; is the rated capacity of the battery; is the first-channel tensor output by the model, representing the charge and discharge power commands 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 = 100 kW, and the current load = 80 kW, then the maximum allowable charging power is min(instruction value, 100 - 80 = 20 kW).

[0109] Figure 4 It is a schematic diagram of the power grid dispatching process executed by the energy management module of the energy storage system in the embodiment of the present invention. The output layer also serves as the execution layer, and the specific operation process executed by the execution layer is as Figure 4 shown, which is used to optimize the allocation of charging resources, preferentially use photovoltaic electric energy for charging, and intelligently dispatch grid electric energy during peak periods, so as to maximize green energy and minimize operating costs.

[0110] The above embodiment of the present invention adopts a three-layer architecture of data-driven, knowledge-guided, and physical constraints. The data layer constructs a spatio-temporal graph structure; the algorithm layer integrates federated learning and fuzzy control; the decision layer embeds safety constraint rules to form a closed-loop control system. By processing uncertain information through fuzzy control, it can effectively capture spatio-temporal correlation, improve the adaptability to dynamic environments, preferentially use photovoltaic electric energy for battery charging, reduce the dependence on fossil energy, and achieve the maximum utilization of green energy. Considering physical constraints comprehensively, it improves the real-world adaptability of the model. Through the prediction and dispatching of intelligent power, the power demand is minimized during the peak period of the power grid, reducing the power cost and load pressure. By using federated learning to protect data privacy, distributed training is achieved. By optimizing the use of charging stations and energy storage systems, the accuracy of power prediction algorithms is improved, and the energy utilization efficiency of the entire system is enhanced.

[0111] The experimental equipment configuration table adopted in this embodiment is shown in Table 1, and the experimental parameter configuration table adopted in this embodiment is shown in Table 2.

[0112] Table 1: Experimental Equipment Configuration Table

[0113] Table 2: Experimental Parameter Configuration Table

[0114] According to the above experimental verification, this experimental configuration proves that: in the 256-node power grid scenario, the technical solution provided by the embodiment of the present invention can achieve: the spatio-temporal prediction accuracy is MAE ≤ 5%; the privacy protection strength: satisfies ϵ ≤ 1.0 differential privacy; the scheduling real-time performance: the control instruction generation < 200 ms; the green energy utilization rate: > 80%. The present invention provides a set of solutions for high-proportion new energy power grids to achieve accurate prediction, privacy protection, and safe dispatching.

[0115] It should be noted that the embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. For the same or similar parts among the embodiments, reference can be made to each other. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple. For the relevant parts, reference can be made to the description in the method part.

[0116] It should also be noted that in the embodiments of the present application, relative terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variation thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not expressly listed, or also includes elements inherent to such process, method, article or device. Without further limitation, an element defined by the statement "including one..." does not exclude the existence of additional identical elements in the process, method, article or device including the said element.

[0117] 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 obvious to those skilled in the art, and the general principles defined in the embodiments of the present application can 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 these embodiments shown in the embodiments of the present application, but will conform to the widest scope consistent with the principles and novel features disclosed in the embodiments of the present application.

Claims

1. A power grid deployment method for an energy storage system based on a dynamic spatio-temporal graph convolutional neural network, characterized in that, Including: Step S1: The data input layer acquires the time series data of grid nodes; Step S2: The dynamic spatio-temporal graph convolution layer uses a dynamic spatio-temporal graph convolutional neural network to sequentially perform graph convolution layer operations and time convolution operations to extract the spatio-temporal features of the time series data; Step S3: The attention mechanism layer calculates the attention weighted features of the spatio-temporal features; Step S4: Input the spatio-temporal features and the attention weighted features into the fuzzy control module to handle the uncertainty of new energy output based on fuzzy rules; 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 geographical locations and connection relationships through the physical information fusion layer to output enhanced features; Step S7: Based on the enhanced features, pass through the fully connected layer and activation function to output the grid load prediction result and dynamically adjust the charge and discharge strategies of the energy storage system.

2. The power grid deployment method for an energy storage system based on a dynamic spatio-temporal graph convolutional neural network according to claim 1, wherein The step S1 further includes: Construct 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 at least include electricity production, consumption, and load.

3. The energy storage system grid deployment method based on the dynamic spatio-temporal graph convolutional neural network according to claim 2, wherein, Wherein, The operations performed by the dynamic spatio-temporal graph convolution layer in step S2 include: Dynamic graph construction: Construct a dynamic graph structure based on the physical connection relationship and spatio-temporal dependence relationship between nodes. Among them, the physical connection relationship determines the physical connectivity between nodes through the grid topology structure; the spatio-temporal dependence relationship is determined by the Pearson correlation coefficient; Graph convolution operation: For the dynamic graph structure and the three-dimensional tensor (N, T, F), use the graph convolution operation to extract the spatial features of each time step; Time convolution operation: Perform a one-dimensional time convolution operation on the spatial features along the time dimension to output the spatio-temporal feature representation.

4. The method for dispatching an energy storage system power grid based on a dynamic spatio-temporal graph convolutional neural network according to claim 3, wherein The dynamic graph construction includes: Determine the node physical connection relationship based on the grid topology structure; Quantify the spatio-temporal dependence between nodes through the Pearson correlation coefficient; Generate a sparse adjacency matrix that changes with time, and the non-zero elements represent effective connections.

5. The power grid deployment method for an energy storage system based on a dynamic spatio-temporal graph convolutional neural network according to claim 4, wherein The step S3: Calculating the attention weighted features of the spatio-temporal features through the attention mechanism layer includes: For the spatio-temporal feature representation, calculate the attention weights of each node feature in the time dimension based on the self-attention mechanism and output the time attention weighted features; For the time attention weighted features, calculate the attention weights of each node and other nodes in the graph structure based on the graph attention mechanism and output the graph attention features.

6. The method for dispatching an energy storage system power grid according to claim 5 based on a dynamic spatio-temporal graph convolutional neural network, characterized in that, The step S4: Input the spatio-temporal features and the graph attention features into the fuzzy control module to handle the uncertainty of new energy output based on fuzzy rules and generate a dynamic allocation strategy, including: Fuzzify the graph attention features, define the fuzzy sets and membership functions of the input / output variables; Define fuzzy rules according to expert knowledge and system characteristics, and apply the fuzzy rule base to perform Mamdani reasoning; Through the centroid method or the maximum membership degree method, defuzzify to generate control decisions, including charging scheduling instructions.

7. The energy storage system grid deployment method based on the dynamic spatio-temporal graph convolutional neural network according to claim 6, characterized in that Step S5: Implement distributed model training based on the federated learning framework and dynamically update 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 each node 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; The global model parameters are sent to each node to synchronously update the local model, and the spatio-temporal features after federated learning update are output; Among them, in the federated learning module: differential privacy technology is used in local model training, and Gaussian noise is added to the gradient update process; weighted average strategy is used for model aggregation, and the weights are dynamically allocated according to the node data volume.

8. The method for dispatching an energy storage system power grid based on a dynamic spatio-temporal graph convolutional neural network according to claim 7, wherein Step S6: Integrate the node geographical location and connection relationship through the physical information fusion layer to output enhanced features, including: Encode the physical information of the node into a physical constraint matrix, where the physical information includes: power grid impedance parameters, line capacity limits; Fuse the spatio-temporal features after the federated learning update and the physical constraint matrix through feature cross-operation to obtain fused features; For the fused features, enhance the relationship and dependence between nodes through physical information to output enhanced features.

9. The energy storage system grid deployment method based on a dynamic spatio-temporal graph convolutional neural network according to claim 6, characterized in that, Step S7: Based on the enhanced features, through the fully connected layer and activation function, output the power grid load prediction result and dynamically regulate the charge and discharge strategy of the energy storage system, including: Input the enhanced feature representation output by the physical information fusion layer into the fully connected layer, and generate a three-channel prediction output through linear transformation, including charge and discharge power instructions, energy storage unit selection instructions, and power grid interaction confidence; The physical constraint activation function includes charge and discharge instruction activation function, energy storage selection instruction activation, and confidence activation; and the charge and discharge instructions are corrected in real time based on safety constraints.

10. A energy storage system, characterized in that, Including: The photovoltaic power generation module, including photovoltaic panels and inverters, is used to collect solar energy, convert it into electrical energy, and convert direct current into alternating current for system use; The energy storage module, including lithium-ion battery packs and bidirectional charge and discharge controllers, is used to store excess photovoltaic electrical energy and support peak power output; The charging station is equipped with multi-protocol intelligent charging interfaces and battery management modules, which are used to support fast charging of various electric vehicle battery specifications and monitor the status of each lithium-ion battery pack respectively; The energy management module is used to monitor the production of photovoltaic electrical energy, power grid load, energy storage status, and charging demand of the charging station in real time, and execute the energy storage system power grid deployment method based on the dynamic spatio-temporal graph convolutional neural network as described in any one of claims 1-9, optimize the allocation of charging resources, preferentially use photovoltaic electrical energy for charging, and intelligently allocate power grid electrical energy during peak periods.

Images