A Dynamic Region Partitioning and Distributed Optimization Method for Distribution Networks Based on Graph Neural Networks

Abstract

This invention discloses a dynamic regional partitioning and distributed optimization method for distribution networks based on graph neural networks to address the challenges of high-penetration distributed energy access. The method first constructs a weighted adjacency matrix and node feature matrix based on distribution network topology parameters and source-load prediction data, and then uses a graph neural network-based partitioning model to generate a partitioning scheme that balances electrical coupling and power self-balancing. Second, candidate partitions and the currently implemented partition are evaluated during rolling operation, and partition updates are dynamically triggered when the overall benefit exceeds a preset threshold. Finally, a multi-region decoupling coordination optimization model is established, and a synchronous alternating direction multiplier method with variable penalty parameters is used to coordinate boundary variables for consistency, outputting the scheduling results. This invention effectively reduces inter-regional coupling and tie-line power exchange requirements, improves regional supply-demand self-balancing capabilities and distributed solution efficiency, and has certain engineering application value.

Description

Technical Field

[0001] This invention relates to a method for dynamic regional division and distributed optimization of power distribution networks based on graph neural networks, belonging to the field of power system optimization and operation technology. Background Technology

[0002] With the integration of massive distributed controllable resources into the distribution network, traditional centralized control relies heavily on global information and communication links, resulting in complex operation and regulation that struggles to balance flexibility, real-time performance, and coordination. Distributed collaborative control based on regional autonomy and coordination is gradually replacing centralized control. To support distributed optimal scheduling, the distribution network typically needs to be rationally divided into regions, adhering to the principle of strong coupling within regions and weak coupling between regions. This reduces the number of cross-regional power flow exchanges, boundary variable consistency iterations, and communication overhead, while improving the convergence performance of collaborative optimization. Existing distribution network partitioning methods mainly include clustering, community detection, and heuristic optimization. Clustering and community detection focus primarily on topology, with insufficient modeling of operational constraints, supply-demand self-balancing, and electrical coupling. Heuristic methods are computationally complex and prone to getting trapped in local optima. Furthermore, current distribution network partitioning methods are mostly for static partitioning, making it difficult to adapt to the time-varying fluctuations in renewable energy output and load. Existing dynamic partitioning schemes suffer from fixed update cycles, insufficient foresight, and frequent repartitioning, making it difficult to simultaneously balance partitioning quality, operational stability, and distributed coordination efficiency. Summary of the Invention

[0003] The technical problem to be solved by this invention is to provide a dynamic regional division and distributed optimization method for power distribution networks based on graph neural networks. By using graph neural networks to integrate electrical coupling and power supply and demand balance characteristics, combined with threshold-triggered update mechanism and variable penalty parameter coordinated solution, the rationality of regional division, dynamic adaptability and cross-regional coordination efficiency can be improved.

[0004] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: A method for dynamic region partitioning and distributed optimization of a power distribution network based on graph neural networks includes the following steps: Step 1: Obtain the distribution network topology parameters and source-load prediction data. Calculate the electrical distance between nodes in the distribution network based on voltage-power sensitivity to construct a weighted adjacency matrix and a node feature matrix. Input the weighted adjacency matrix and node feature matrix into the end-to-end region partitioning model to generate a partitioning scheme that takes into account both electrical coupling and power self-balancing. Step 2: During the rolling operation, the candidate partitions for the current period are solved based on the source load prediction data of the prediction time window. The prediction time window includes all periods with a preset window length starting from the current period. The comprehensive evaluation index of the candidate partitions and the actual implementation partitions of the previous period is calculated and compared. When the difference between the comprehensive evaluation index of the candidate partitions and the actual implementation partitions of the previous period is greater than the preset threshold, the partition update is triggered; otherwise, the original partitions are maintained. Step 3: Based on the current time period's partitioning results, the global optimization problem of the distribution network is decomposed into multiple regional sub-problems, and tie-line reference variables are introduced to establish consistency constraints; the synchronous alternating direction multiplier method with variable penalty parameters is used for parallel coordination iteration, and the multi-region coordinated optimization scheduling results are output after convergence.

[0005] Compared with the prior art, the present invention, employing the above technical solution, has the following technical effects: The method proposed in this invention can accurately capture the matching relationship between the distribution network topology and time-varying source loads. It achieves system regional autonomy through a dynamic partitioning strategy, which significantly reduces tie-line power exchange and distributed computing complexity while improving operational reliability and has certain engineering application value. Attached Figure Description

[0006] Figure 1 This is an architecture diagram of the dynamic region partitioning and distributed optimization method for power distribution networks based on graph neural networks, as proposed in this invention. Figure 2 This is a typical daily power curve of photovoltaic, wind power and load in the distribution network in the embodiments of the present invention; Figure 3 This is a training loss convergence curve of the partitioning method based on graph neural networks in an embodiment of the present invention; Figure 4 This is a graph showing the changes in indicators and the optimal number of partitions under different numbers of partitions in this embodiment of the invention; Figure 5 This is a comparison chart of the distribution network partitioning results under different network partitioning methods in the embodiments of the present invention; Figure 6 This is a comparison chart of 24-hour power self-balancing indices under different division methods in this embodiment of the invention; Figure 7 This is a schematic diagram of the dynamic network partitioning results in an embodiment of the present invention; Figure 8 This is a comparison chart of modularity and self-balancing index under static partitioning and dynamic partitioning in this embodiment of the invention; Figure 9 This is a convergence curve of the distributed coordination optimization algorithm in an embodiment of the present invention. Detailed Implementation

[0007] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0008] like Figure 1 As shown, this invention proposes a method for dynamic region partitioning and distributed collaborative optimization of power distribution networks based on graph neural networks, comprising the following steps: Step 1: Obtain distribution network topology parameters and source-load prediction data; calculate node electrical distances based on voltage-power sensitivity to construct a weighted adjacency matrix and node feature matrix; then input the data into a graph neural network clustering model to generate a partitioning scheme that balances electrical coupling and power self-balancing; details are as follows: Step 101: Obtain distribution network topology parameters and source-load prediction data; calculate the electrical distance between nodes based on voltage-power sensitivity; and then construct a weighted adjacency matrix and a node feature matrix. Specifically, this includes: First, the influence of node j on node i is calculated based on voltage-power sensitivity. The electrical distance between node i and node j is calculated using Euclidean distance. : , , In the formula, These represent the sensitivities of active and reactive power injection at node j to the voltage at node i, respectively. These represent the sensitivities of active and reactive power injection at node j to the voltage at node j, respectively; N is the total number of nodes in the distribution network.

[0009] Secondly, based on the calculated electrical distance, a weighted adjacency matrix A is constructed using a Gaussian kernel function, whose elements... The calculation formula is: , In the formula, For bandwidth parameters, Let the set of edges of the distribution network topology be denoted as . This represents the edge formed by node i and node j;

[0010] Finally, it will be by The resulting electrical distance matrix is ​​denoted as L. Principal component analysis (PCA) is used to reduce the dimensionality of matrix L, and the first p principal components are extracted to construct the node feature matrix X. , In the formula, PCA represents principal component analysis, and p is the number of principal components retained. Let represent the N×p-dimensional real matrix space.

[0011] Step 102: Construct an end-to-end region partitioning model based on graph convolutional network (GCN) and deep modularity network (DMoN), jointly optimize structural clustering and power self-balancing loss, and output the final distribution network partitioning scheme. First, using the weighted adjacency matrix A and the node feature matrix X obtained in step 101 as input, a GCN with L layers is constructed for node feature learning. The layer update formula is as follows: , In the formula, It is a symmetric normalized weighted adjacency matrix. The hidden state of layer l (input layer) ), Let be the learnable weight matrix of the l-th layer. Here is the learnable weight matrix for the l-th layer skip connection. It is a non-linear activation function; After L layers of convolution, the final node embedding representation is obtained: , In the formula, Let h be the hidden state of the Lth layer, and h be the embedding dimension;

[0012] Secondly, based on the node embedding H, the soft allocation matrix S from the node to K preset regions is calculated by introducing a region allocation head with a softmax activation function: , In the formula, Assign a weight matrix to the learnable region; Next, a joint loss function is constructed to perform end-to-end training of the model. Joint Loss Function Clustering loss based on DMoN structure and regional power self-balancing loss composition: , In the formula, represents the weighting coefficients. Wherein, the DMoN structural clustering loss... The calculation formula is: , In the formula, For matrix trace operations, Let B be the total edge weight, and B be the modularity matrix. Let i be the vector of the i-th row of matrix S. It is the Frobenius norm. Regional power self-balancing loss. Defined in the prediction time window The calculation formula is as follows: , , In the formula, To predict the time window, , The preset window length; Let k be the power self-balancing level of region k in time period t; , or , It is a very small positive number; , , These are the net active load, adjustable power capacity, and adjustable power capacity of region k, obtained based on the soft allocation matrix S, respectively. Let K be the total active power load demand for region k.

[0013] Finally, the joint loss function is minimized through backpropagation. After the model converges, the region number corresponding to the highest probability is extracted based on the soft allocation matrix S, resulting in the hard partitioning result. : , In the formula, This represents the probability that node i is assigned to region k, thus outputting a partitioning scheme that balances tight electrical coupling with the region's power self-balancing capability.

[0014] Step 2: During the rolling operation, candidate partitions for the current time period are calculated based on the source load prediction data within the prediction time window. The prediction time window includes all time periods of a preset window length starting from the current time period. The comprehensive evaluation index of the candidate partitions and the actual implemented partitions of the previous time period is calculated and compared. When the difference between the comprehensive evaluation index of the candidate partitions and the actual implemented partitions of the previous time period is greater than a preset threshold, a partition update is triggered; otherwise, the original partitions are maintained. Details are as follows: Step 201: Construct a comprehensive evaluation index for assessing candidate partitioning schemes, specifically including: First, calculate the modularity index for time period t. : , In the formula, The weighted edge between nodes i and j in time period t; , , ; and These represent the regions to which node i and node j belong, respectively. This is an indicator function; its value is 1 when nodes i and j belong to the same region, and 0 otherwise.

[0015] Secondly, calculate the power self-balancing index for time period t. : , , , , In the formula, Let be the power self-balancing index for region k, and truncate it to the range [0-1]. , K is a very small positive number; K is the number of regions. , , These represent the set of nodes within region k, the set of micro gas turbines, and the set of electric energy storage, respectively. , , These represent the load power, photovoltaic output, and wind power output of node i in time period t, respectively. , These represent the net active power load and total load demand for region k, respectively. , These represent the maximum up-adjustment and down-adjustment capabilities of region k, respectively. This represents the upper limit of the maximum usable output of the micro gas turbine g. , These represent the maximum discharge and maximum charging power limits of energy storage e, respectively.

[0016] Finally, a comprehensive evaluation index was constructed. : , In the formula, These are the weighting coefficients.

[0017] Step 202: Update the implementation partition based on a threshold-triggered dynamic partitioning mechanism, specifically including: For time period t, the partitioning actually implemented in the previous time period Based on the prediction time window Solving candidate partitions from source load prediction data , , The preset window length; Calculate the implementation partitions separately With candidate partitions Corresponding comprehensive evaluation indicators and And calculate the gain. : , Determine gain With preset threshold Size relationship: If This triggers a partition update, and candidate partitions are used in time period t. As a new implementation partition; otherwise, the original partition remains unchanged.

[0018] Step 3: Based on the current partitioning results, decompose the global optimization problem into multiple regional sub-problems and establish consistency constraints for the connection variables; use synchronous ADMM (Alternating Direction Multiplier Method) with variable penalty parameters for parallel coordinated iteration, and output the multi-region coordinated optimization scheduling results after convergence; details are as follows: Step 301: Distribution Network Area Decoupling and Variable Initialization Based on the division of multiple distribution network areas, electrical decoupling is performed at the tie lines between adjacent areas. The nodes at both ends of the tie line are defined as transmitting node i and receiving node j, respectively, and their respective areas correspond to the transmitting area and the receiving area. A tie line state variable vector is defined. for: , In the formula, and These are the active power and reactive power of the tie line, respectively. and These are the voltage amplitudes at the transmitting and receiving nodes, respectively. Initialize the number of iterations Initialize the Lagrange multipliers for each region. Initialize penalty parameters And set convergence tolerance .

[0019] Step 302: Parallel solution of local subproblems and global variable update Because distribution networks in multiple areas are electrically interconnected, the calculated tie-line power flow states in adjacent areas may be inconsistent. To gradually converge the boundary states on both sides to a uniform state, a tie-line reference variable is introduced. .

[0020] First, in the m-th iteration, each distribution network region solves its own optimization subproblem in parallel, and the transmitting and receiving regions obtain the locally calculated values ​​of the tie-line boundary state variables under the current iteration. and ; Secondly, the local state variable values ​​of the sending and receiving areas are averaged to update the tie-line reference variable in the (m+1)th iteration. : , In the formula, , These are the locally calculated values ​​of the tie-line boundary state variables for the transmitting and receiving areas at the m-th iteration, respectively. Finally, based on the obtained local state variable values ​​and the latest reference variables, update the Lagrange multipliers in the sending and receiving areas: , In the formula, , These are the Lagrange multipliers for the sending or receiving region at the (m+1)th and mth iterations, respectively. This is the penalty parameter for the m-th iteration.

[0021] Step 303: Residual calculation, adaptive parameter adjustment and convergence judgment First, calculate the original residual for the m-th iteration. and dual residual : , , In the formula, , These are the connection line reference variables for the m-th and (m-1)-th iterations, respectively;

[0022] Secondly, the penalty parameters are updated based on the standard residual balancing rules. : , In the formula, This is the residual imbalance factor. and These are the increment and decrement factors for the penalty parameter, respectively; and the updated... The value is limited to preset upper and lower limits. Within the range;

[0023] Finally, determine the original residuals. and dual residual Do they simultaneously meet the convergence conditions? and If the conditions are met, the algorithm terminates and outputs the collaborative optimization scheduling results for each region; otherwise, it sets... Return to step 302 to continue the next iteration.

[0024] To verify the effectiveness of the distribution network zoning and distributed scheduling method proposed in this invention, this embodiment uses an IEEE 33-node distribution network as the test system, which is equipped with distributed energy sources such as photovoltaic, wind power, energy storage, and micro gas turbines. Figure 2 Typical daily power curves for photovoltaic, wind power, and load are displayed, reflecting the 24-hour source-load fluctuation characteristics of the system.

[0025] First, the graph partitioning model is built on the PyTorch platform, using the Adam optimizer with a learning rate of 0.001 and 400 training epochs. Figure 3 As can be seen, the total loss decreases rapidly in the early stages of training and stabilizes after about 50 rounds, indicating that the model has good convergence. Among them, the modularity loss decreases significantly, indicating that the model effectively learns the topology of compact regions; the collapse regularization term remains at a low level, suppressing the degeneracy caused by excessive node concentration; and the power self-balancing loss changes smoothly, providing continuous self-balancing guidance while ensuring clustering quality.

[0026] Secondly, different numbers of partitions are selected. Figure 4 Modularity, average self-balancing index, and overall score are presented for different numbers of partitions. As the number of partitions increases, modularity rises while the average self-balancing index decreases, indicating that excessive partitioning weakens the source-load matching ability within a region. After comprehensive comparison, K=4 is selected as the static partitioning result. Figure 5 and Figure 6 This demonstrates that, compared with spectral clustering, the Louvain method, and the DMoN Only method, the present invention can significantly improve the regional power supply and demand self-balance level while maintaining a high degree of modularity.

[0027] Since static partitioning is difficult to adjust with the time-varying characteristics of source load, this embodiment introduces a dynamic network partitioning mechanism. Figure 7 The system's evolution was demonstrated: during the period from 0:00 to 4:59, wind power was high and load was low, and the system maintained a four-zone structure; during the period from 5:00 to 16:59, the distribution of wind and solar power output and load changed, and the zone boundaries were fine-tuned; during the period from 17:00 to 23:59, photovoltaic power returned to zero and the evening peak load increased, and some adjacent zones were merged into three zones to expand the coverage of regulation resources. Figure 8 This indicates that although dynamic partitioning slightly reduces modularity, it achieves a higher level of self-balancing across multiple time periods, enhancing short-term operational adaptability.

[0028] Finally, a multi-region distributed coordination and scheduling model is established based on dynamic partitioning, and solved using the synchronous ADMM algorithm with variable penalty parameters. This part is implemented in MATLAB with eight parallel threads, and each region's subproblem is solved independently by the Gurobi solver. Figure 9 The results show that both the original residual and the dual residual can be reduced to the tolerance range within a few iterations. Simulation results demonstrate that, under the premise of comparable total cost, this invention effectively reduces tie-line switching power and shortens computation time, thereby improving distributed scheduling efficiency. The above simulation results verify the effectiveness of this invention.

[0029] Based on the same inventive concept, this application provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the aforementioned method for dynamic regional division and distributed optimization of power distribution networks based on graph neural networks.

[0030] Based on the same inventive concept, embodiments of this application provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned steps of the dynamic regional division and distributed optimization method for power distribution networks based on graph neural networks.

[0031] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0032] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0033] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0034] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0035] The above embodiments are merely illustrative of the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solutions based on the technical concept proposed in this invention shall fall within the scope of protection of this invention.

Claims

1. A method for dynamic region partitioning and distributed optimization of a distribution network based on graph neural networks, characterized in that, Includes the following steps: Step 1: Obtain the distribution network topology parameters and source-load prediction data, and calculate the electrical distance between nodes in the distribution network based on voltage-power sensitivity to construct a weighted adjacency matrix and a node feature matrix; The weighted adjacency matrix and node feature matrix are input into the end-to-end region partitioning model to generate a partitioning scheme that takes into account both electrical coupling and power self-balancing. Step 2: During the rolling operation, the candidate partitions for the current time period are solved based on the source load prediction data of the prediction time window. The prediction time window includes all time periods with a preset window length starting from the current time period. Calculate and compare the comprehensive evaluation index of the candidate partition with the actual implemented partition of the previous period. When the difference between the comprehensive evaluation index of the candidate partition and the actual implemented partition of the previous period is greater than the preset threshold, the partition is triggered to update; otherwise, the original partition is maintained. Step 3: Based on the zoning results of the current time period, the global optimization problem of the distribution network is decomposed into multiple regional sub-problems, and tie line reference variables are introduced to establish consistency constraints; A synchronous alternating direction multiplier algorithm with variable penalty parameters is used for parallel coordination iteration, and the multi-region coordinated optimization scheduling result is output after convergence.

2. The method for dynamic region partitioning and distributed optimization of distribution networks based on graph neural networks according to claim 1, characterized in that, The specific process of step 1 is as follows: Step 1.1: Calculate the influence of node j on node i in the distribution network based on voltage-power sensitivity. The electrical distance between node i and node j is calculated using Euclidean distance. : ， ， In the formula, These represent the sensitivities of active and reactive power injection at node j to the voltage at node i, respectively. These represent the sensitivities of active and reactive power injection at node j to the voltage at node j, respectively, where N is the total number of nodes in the distribution network. Based on the calculated electrical distance A weighted adjacency matrix A is constructed using a Gaussian kernel function, whose elements... The calculation formula is: ， In the formula, For bandwidth parameters, Let the set of edges of the distribution network topology be denoted as . This represents the edge formed by node i and node j; Will be The resulting electrical distance matrix is ​​denoted as L. Principal component analysis is used to reduce the dimensionality of L, and the first p principal components are extracted to construct the node feature matrix X. ， In the formula, PCA represents principal component analysis, and p is the number of principal components retained. Represents an N×p dimensional real matrix space; Step 1.2: Construct an end-to-end region partitioning model based on a Graph Convolutional Network (GCN) and a Deep Modular Network (DMoN). The GCN consists of L convolutional layers and a region assignment head. The weighted adjacency matrix A and the node feature matrix X are used as inputs to the GCN. The layer update formula for each convolutional layer is as follows: ， In the formula, It is a symmetric normalized weighted adjacency matrix. The first Hidden state of the layer, hidden state of the input layer , Let be the learnable weight matrix of the l-th layer. Here is the learnable weight matrix for the l-th layer skip connection. It is a non-linear activation function; The final node embedding is obtained after L convolutional layers. Represented as: ， In the formula, Let h be the hidden state of the Lth layer, and h be the embedding dimension; Based on node embedding H, a soft allocation matrix S from all nodes to K preset regions is calculated by introducing a region allocation head using a softmax activation function: ， In the formula, Assign a weight matrix to the learnable region; Constructing a joint loss function The region partitioning model is trained end-to-end using the following formula: ， ， ， ， In the formula, This represents the clustering loss of the DMoN structure. This indicates the regional power self-balancing loss. These are the weighting coefficients. For total side weight, This is a matrix trace operation, where B is the modularity matrix. Let i be the vector of the i-th row of the soft-assignment matrix S. It is the Frobenius norm. Let k be the power self-balancing level in region k during time period t. To predict the time window, , The preset window length; , or , It is a very small positive number. , , These represent the net active power load, adjustable power capacity, and adjustable power capacity of region k, obtained based on the soft allocation matrix S. Let K be the total active power load demand for region k. Minimize the joint loss function through backpropagation. After the model converges, the region number corresponding to the highest probability is extracted based on the soft allocation matrix S to obtain the hard partitioning result. : ， In the formula, This represents the probability that node i is assigned to region k, thus outputting a partitioning scheme that balances tight electrical coupling with the region's power self-balancing capability.

3. The method for dynamic region partitioning and distributed optimization of distribution networks based on graph neural networks according to claim 1, characterized in that, The specific process of step 2 is as follows: Step 2.1: Construct a comprehensive evaluation index for assessing candidate partitioning schemes. : ， In the formula, These are the weighting coefficients. The modularity index for time period t. The power self-balancing index for time period t; ， In the formula, The weighted edge between nodes i and j in time period t. , , ; and These represent the regions to which node i and node j belong, respectively. This is an indicator function; its value is 1 when nodes i and j belong to the same region, and 0 otherwise. ， ， ， ， In the formula, Let K be the power self-balancing index for region k, truncated to the range [0-1]; K is the number of regions. , or , , These represent the maximum up-adjustment and down-adjustment capabilities of region k, respectively. , These represent the net active power load and total load demand for region k, respectively. It is a very small positive number; , , These represent the set of nodes within region k, the set of micro gas turbines, and the set of electric energy storage, respectively. , , Let represent the load power, photovoltaic output, and wind power output of node i in time period t, respectively. This represents the upper limit of the maximum usable output of the micro gas turbine g. , These represent the maximum discharge and maximum charging power limits of the energy storage e, respectively. Step 2.2, for the current time period t, use the prediction time window Solving candidate partitions from source load prediction data , , The preset window length; Calculate the actual partitions implemented in the previous time period respectively. With candidate partitions Corresponding comprehensive evaluation indicators and And calculate the gain. : ， Determine gain With preset threshold Size relationship: If This triggers a partition update, and candidate partitions are used in time period t. As a new implementation partition; otherwise, the original partition remains unchanged.

4. The method for dynamic region partitioning and distributed optimization of distribution networks based on graph neural networks according to claim 1, characterized in that, The specific process of step 3 is as follows: Step 3.1: Based on the divided distribution network areas, perform electrical decoupling at the tie lines between adjacent areas, defining the nodes at both ends of the tie line as transmitting node i and receiving node j, respectively, with their respective areas corresponding to the transmitting area and the receiving area; define the tie line state variable vector. for: ， In the formula, and These are the active power and reactive power of the tie line, respectively. and These are the voltage amplitudes of the transmitting and receiving nodes, respectively. Initialize the number of iterations Initialize the Lagrange multipliers for each region. Initialize penalty parameters And set convergence tolerance ; Step 3.2 introduces tie-line reference variables. In the m-th iteration, each distribution network region solves its corresponding sub-problems in parallel, and the sending and receiving regions obtain the locally calculated values ​​of the tie-line boundary state variables under the current iteration. and And calculate the average update of the connection reference variable in the (m+1)th iteration. : ， Based on the obtained local state variable values ​​and the latest reference variables, update the Lagrange multipliers in the sending and receiving areas: ， In the formula, These are the Lagrange multipliers of the receiver region at the (m+1)th and mth iterations, respectively. These are the Lagrange multipliers of the sending region at the (m+1)th and mth iterations, respectively. Let be the penalty parameter for the m-th iteration; Step 3.3, calculate the original residual for the m-th iteration. and dual residual : ， ， In the formula, , These are the connection line reference variables for the m-th and (m-1)-th iterations, respectively; Update penalty parameters based on standard residual balancing rules: ， In the formula, Let be the penalty parameter for the (m+1)th iteration. This is the residual imbalance factor. and These are the increment and decrement factors for the penalty parameter, respectively; and the updated... The value is limited to preset upper and lower limits. Within the range; Determine the original residual and dual residual Do they simultaneously meet the convergence conditions? and , If the condition is met, the algorithm terminates and outputs the collaborative optimization scheduling results for each region; otherwise, it sets the value to a minimum positive number. Return to step 302 to continue the next iteration.

5. A computer device comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the dynamic regional division and distributed optimization method for power distribution networks based on graph neural networks as described in any one of claims 1 to 4.

6. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the dynamic regional division and distributed optimization method for power distribution networks based on graph neural networks as described in any one of claims 1 to 4.

Images