Multi-period stochastic dynamic reconfiguration method for power distribution network with wind power based on graph neural network guided improved switch breaking and making
Patent Information
- Application Number
- CN202611188293.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-06
- Publication Date
- 2026-10-27
Smart Images
Figure CN122992748A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the intersection of distribution network optimization and artificial intelligence, and in particular to a multi-period stochastic dynamic reconfiguration method for distribution networks containing wind power based on graph neural network-guided improved switch switching. Background Technology
[0002] Distribution networks typically operate in a radial topology, with each node connected to a substation via a unique path, and loops are not permitted within the network. The switches within the network mainly include normally open tie switches and normally closed sectionalizing switches, whose states determine the operational structure of the distribution network. Distribution network reconfiguration, by altering the opening and closing states of these switches, creates a new network topology. This aims to reduce network losses, improve voltage quality, and enhance system reliability and economy while satisfying constraints such as radial topology, power flow, and voltage. Due to the involvement of discrete switch state variables and nonlinear AC power flow constraints, distribution network reconfiguration is an NP-hard mixed-integer nonlinear programming (MINLP) problem.
[0003] With the large-scale integration of distributed wind power, the randomness and volatility of wind power output bring significant uncertainties to the operation of distribution networks. Wind power output exhibits non-stationary fluctuations due to factors such as wind speed and direction, leading to instability in network voltage distribution and branch power flow, which in turn affects the reliability of reconfiguration decisions. Traditional deterministic or single-period load reconfiguration methods are prone to local optima and voltage overruns in high wind power penetration scenarios, making it difficult to guarantee long-term operational reliability and economy. To address wind power fluctuations, robust optimization or stochastic distribution network reconfiguration (SDNR) models are often used in engineering to comprehensively consider the spatiotemporal relationships and probability distributions of wind power output throughout the day. However, large-scale SDNR models involve a large number of binary decision variables and nonlinear power flow constraints, and traditional mathematical programming (MP) methods are computationally too time-consuming, making it difficult to meet the real-time requirements of online scheduling.
[0004] Heuristic algorithms (such as genetic algorithms and particle swarm optimization) are favored in engineering practice due to their fast solution speed and simple implementation. Among them, the switch opening and exchange (SOE) method, by opening loops and performing local branch exchanges, has shown significant advantages in balancing reconstruction accuracy and computational efficiency. However, in high-dimensional stochastic scenarios, even with heuristic algorithms, topology searches are still required for each time segment and a large number of stochastic scenarios. The computational burden increases dramatically with the network size and the number of scenarios, and there are a large number of invalid traversals in the search space.
[0005] In recent years, artificial intelligence-based methods have provided new approaches for optimizing the operation of power distribution networks. Graph Neural Networks (GNNs) possess spatial feature extraction and nonlinear mapping capabilities, enabling them to directly operate on non-Euclidean power grid topologies and uncover the intrinsic correlation between node-injected features and the optimal radial topology. However, purely data-driven methods struggle to guarantee the radiality and physical feasibility of the topology, potentially leading to islanding or loops under highly nonlinear power flows and stochastic operating conditions, thus impacting system security.
[0006] In summary, the reconfiguration of distribution networks including wind power still faces the following key issues: 1) Large-scale random reconfiguration leads to a significant expansion of the search space, and traditional physical optimization methods are difficult to meet the computational efficiency requirements of online scheduling; 2) Although most existing intelligent algorithms solve problems quickly, they cannot strictly guarantee the radial topology and physical feasibility of the distribution network under complex operating conditions. Summary of the Invention
[0007] The purpose of this invention is to provide a multi-period random dynamic reconfiguration method for wind power distribution networks based on graph neural network-guided improved switch switching, which enhances reconfiguration accuracy and computational efficiency.
[0008] The objective of this invention can be achieved through the following technical solutions: A multi-period stochastic dynamic reconfiguration method for wind power distribution networks based on graph neural networks-guided improved switchgear switching includes the following steps: Construct a multi-period stochastic reconfiguration model with the objective of minimizing total energy loss; Abstracting the power distribution network as an undirected network topology graph And based on the undirected graph Construct the cross-sectional feature matrix for each time period ,in Represents a set of nodes. Represents the set of branches. Indicates time period, This indicates the scenario, with each branch corresponding to a handover switch; Based on the cross-sectional feature matrix The pre-trained graph convolutional neural network prediction model is used to output the probability of each branch being disconnected as the connection switch. Based on the probability that all communication switches are disconnected, the switch with the highest probability is selected. For each branch, a candidate set of interconnecting switches is constructed, and based on the candidate set of interconnecting switches, an initial radial topology is generated using a sequential switching strategy, denoted as the decision vector. ; Based on the decision vector The model employs a multi-period random reconstruction method, using an improved switch-on-off switching method to perform local topology optimization and global topology reconstruction, thereby obtaining the globally optimal reconstructed topology and completing the random dynamic reconstruction process for each period.
[0009] Furthermore, the multi-time-period stochastic reconstruction model includes an objective function aimed at minimizing the total energy loss and corresponding constraints. The objective function is: In the formula, This represents the total number of time periods. For scene collection, For the set of all switch branches, for Time Period Scene probability of occurrence branch road The switch status is represented by a binary variable, where a value of 1 indicates the switch is closed and a value of 0 indicates it is open. and They are nodes During the period Scene The voltage amplitude and phase angle under these conditions, and These are the resistance and reactance of the branch circuit, respectively. The unit time interval; The constraints include: Nodal active power balance constraints: In the formula, and They are nodes Total active power output and active load demand, For nodes A set of connected nodes. For nodes During the period Scene Voltage amplitude under, and These are the real and imaginary parts of the admittance matrix, respectively. The voltage angle difference between nodes. For nodes During the period Scene The phase angle below; Nodal reactive power balance constraints: In the formula, and They are nodes Total reactive power output and reactive load demand; Branch capacity constraints: In the formula, branch road Maximum permissible apparent power limit; Node voltage range constraints: In the formula, and node Permissible lower and upper limits for voltage amplitude; Constraints are formed by the active and reactive power output of the power source: In the formula, and They are nodes The active power output and reactive power output of conventional distributed power sources. and They are nodes The active power output and reactive power output of wind power; Scenario probability constraints: In the formula, For the scene Probability of occurrence; Wind power output constraints: In the formula, The preset wind power operating power factor; Radial network topology and connectivity constraints: In the formula, This represents the total number of nodes in the network topology. This represents the total number of branches that are in a closed state after reconstruction. Represents a node With the root node of the substation The topological distance between them.
[0010] Furthermore, the cross-sectional feature matrix The construction steps include: Considering the dual uncertainties of load and wind power output, the construction period is... Scene Next node The disturbance load and wind power output model is expressed as: In the formula, , For nodes During the period Scene The active and reactive load demands are as follows: For nodes The active power output of the wind power source , For nodes During the period Baseline predicted values for active and reactive loads. These are the random disturbance variables for load and wind power, respectively. For nodes During the period Baseline predicted value of active power output of wind power; To characterize the actual power demand of nodes, time periods are defined based on the aforementioned disturbance load and wind power output model. Scene Next node Net injected active power With net injected reactive power , respectively represented as: In the formula, For nodes The reactive power output of the wind power source; Based on the net injected active power With net injected reactive power Construct the cross-sectional feature matrix , is represented as: in: , In the formula, For nodes The cross-sectional feature matrix, subscript This represents the total number of nodes in the network.
[0011] Furthermore, the graph convolutional neural network prediction model includes a preprocessing layer, multiple graph convolutional layers, and an MLP-based edge classifier. The execution steps of the graph convolutional neural network prediction model include: In the preprocessing layer, the cross-sectional feature matrix is... Z-score standardization is performed to obtain the standardized feature matrix of the entire network. , is represented as: in: , In the formula, For nodes The standardized cross-sectional feature matrix, For nodes The cross-sectional feature matrix, , Let be the vector of mean and standard deviation for all cross-sections. For element-wise division, It is a very small constant; Based on the network-wide standardized feature matrix The multi-layer graph convolutional layer is used to aggregate neighborhood node information to obtain the final representation matrix that integrates power grid topology correlation and power injection information. ,in This represents the total number of convolutional layers in the graph. Based on the final representation matrix The MLP-based edge classifier is used to infer each branch as a tie switch. Probability of being disconnected .
[0012] Furthermore, the final representation matrix The steps to obtain it include: The network-wide standardized feature matrix As input to a multi-layer graph convolutional layer, it utilizes an inter-layer feature evolution mechanism to aggregate neighborhood node information, thereby enabling features to evolve from the first layer... layer to the first Layer update propagation, wherein the inter-layer feature evolution mechanism follows: In the formula, , The first Layer, First The hidden feature matrix of the nodes in the layer, for Activation function For symmetric normalized adjacency matrix operators, for The degree matrix, To introduce the adjacency matrix of the self-loop structure, Given an adjacency matrix defined based on an undirected network topology graph, when the branch... , Otherwise, it is 0. For the set of all switch branches, for The Line 1 Column elements, for The Line 1 Column elements, It is the identity matrix. For the first The learnable weight matrix of the layer; go through The aggregation operation of the layered graph convolutional layer ultimately outputs a final representation matrix that integrates grid topology correlation and power injection information. The final representation matrix The first in Row vectors represent nodes The high-dimensional embedding features are denoted as: In the formula, For nodes High-dimensional embedding features, for The first in Row vectors.
[0013] Furthermore, the inference process of the MLP-based edge classifier is expressed as follows: In the formula, To connect branch roads The concatenated vector of features from both ends of the node. For nodes High-dimensional embedding features, , , , These are the weight matrices and bias terms for each layer of the edge classifier, respectively.
[0014] Furthermore, the decision vector The steps to obtain it include: The candidate contact switch set is set as follows: ,in For candidate tie switch indexes, their corresponding branches , For all switch branch sets, This is a function for sorting in descending order. For the first The probability of a candidate contact switch being disconnected; For the candidate set of handshake switches Each candidate contact switch in In serial or parallel topology iteration across different threads, an initial radial topology satisfying the network's radial topology and connectivity constraints is generated using a sequential switching strategy, denoted as the decision vector. One of the topological iteration processes includes: Force disconnect the candidate contact switch As the initial constraint for this topology iteration, namely , For candidate handover switch State variables, Indicates a disconnected state; The initial constraints, the radial topology of the network, and the connectivity constraints are taken as input, and the sequential switch on / off strategy is called to generate the current initial radial topology, which is used as the decision vector for this operation. The decision vector Represented as: In the formula, This is a sequential switch on / off mapping operator.
[0015] Furthermore, the local topology optimization step includes: Based on the decision vector The end-segment switches are selected as Type-2 switches, wherein the end-segment switches are candidate tie switches, are in a closed state, and the number of segmented branches between their downstream nodes and the topology end nodes does not exceed a preset threshold. ; For each of the Type-2 switches, disconnect it to make its downstream area an island; To restore network connectivity and satisfy the network radial topology constraints, only one backup contact switch connected to the island is closed to obtain the initial radial topology generated by all single O1C1 operations. For each single O1C1 operation, the total energy loss is recalculated based on the initial radial topology. If the recalculated total energy loss is lower than the total energy loss before the operation, the O1C1 operation is considered an effective energy loss reduction action. Extract independent and effective loss reduction actions that do not belong to the same feeder region. Simultaneously execute two or more effective loss reduction actions in combination. Based on the single O1C1 operation and the combined O1C1 operation, a final radial topology set is constructed, denoted as... ; In the radial topology set Within this process, the multi-time-period random reconstruction model is used for screening to obtain candidate contact switches. Local optimal topology , is represented as: In the formula, Radial topological set Any legal topology within, To the current cross-sectional feature matrix The total energy loss evaluation function is as follows: This is the objective function in the multi-period stochastic reconstruction model, which aims to minimize total energy loss.
[0016] Furthermore, the step of obtaining the globally optimal reconstructed topology includes: For all candidate contact switches Corresponding local optimal topology Under the constraints of radial topology and connectivity, compare the corresponding total energy loss values and find the local optimal topology with the minimum total energy loss value. As the globally optimal topology reconstruction , is represented as: In the formula, For the candidate handshake switch set, This is the total energy loss evaluation function. This is the objective function in the multi-period stochastic reconstruction model, which aims to minimize total energy loss.
[0017] Furthermore, the graph convolutional neural network prediction model employs a weighted binary cross-entropy loss function to guide offline training during the training process. The weighted binary cross-entropy loss function is expressed as follows: Among them, the penalty coefficient Represented as: In the formula, For the total loss, For the set of all switch branches, For the contact switch The true label, For the contact switch Predicted labels, For the set constant, The real labels are generated using a sequential switching strategy, and the generation process of the real labels includes: Assuming all branches in the distribution network are in a closed state, construct a fully closed network topology, denoted as . , branch road The closed state, where a value of 1 indicates the closed state. For the set of all switch branches; Based on the aforementioned fully closed network topology, a sequential switch-on / off strategy is used to iteratively determine the current set of loop switch branches. Search for the optimal switching device The number of iterations is the number of loops in the fully closed network topology, and the branch that minimizes the increase in energy loss is selected in each iteration. To break, set to =0, until the radial topology constraint of the network is satisfied, the optimal on / off switch. Represented as: In the formula, For loop switch branch collection The switch in the middle, To disconnect the switch Then in the current cross-sectional feature matrix Energy loss; After the iteration ends, a loop-free network topology is obtained, and each branch is generated using the mapping relationship. The corresponding real labels form the real label vector. , is represented as: In the formula, For sequential switch on / off mapping operators, These are real labels; a value of 1 indicates a closed state, and a value of 0 indicates an open state. This is the set of all switch branches.
[0018] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention abstracts the distribution network into an undirected network topology graph, uses the spatial feature extraction and nonlinear mapping capabilities of the graph convolutional neural network prediction model to predict the probability distribution of each branch becoming a tie switch in order to intelligently reduce the search space, and combines the improved SOE algorithm to correct the local optimal error of sequential greedy switching, and further selects the globally optimal reconstructed topology to improve the reconstruction accuracy and computational efficiency, and effectively improve voltage quality.
[0019] (2) For the multi-period stochastic reconfiguration model with total energy loss as the optimization objective, this invention considers AC power flow constraints, topology constraints and wind power output uncertainty. It establishes a nonlinear mapping from the node net power section to the optimal topology probability distribution through GNN. By identifying the set of key interconnection switches in advance, it reduces invalid searches in the reconfiguration.
[0020] (3) Based on the prediction results, this invention introduces an improved SOE local exchange operator, which improves the problem of easily getting trapped in local optima with single sequential discontinuity. Under the premise of ensuring the absolute radial direction of the topology and physical constraints, the reconstruction accuracy in complex scenarios is improved.
[0021] (4) This invention utilizes the spatial feature extraction capability of graph neural networks to establish a nonlinear mapping from the node net power section to the optimal radial topology probability distribution. By intelligently reducing the search space through early identification of key interconnection switch sets, it significantly reduces invalid searches. In simulation verification, the Top-K hit rate reached 96.27% on the IEEE 119-node system, and the reconstruction computation time was reduced by 4.36 times compared to the original SOE method. On the 136-node system, the speedup reached 3.98 times, meeting the real-time requirements of online scheduling.
[0022] (5) This invention introduces an improved SOE local switching operator, focusing on performing O1C1 local switching on the end segmented switches, and combining and synchronously executing independent loss reduction actions of different feeders. This effectively overcomes the problem of easily getting trapped in local optima with a single sequential interruption, and improves reconstruction accuracy while strictly ensuring radial topology and physical constraints. In simulation verification, the network loss error of the 33-node system is 0%, and the errors of the 119-node and 136-node systems are only 0.05% and 0.37%, respectively.
[0023] (6) The present invention exhibits strong robustness in scenarios with multiple loads and uncertain wind power output. All test scenarios achieved a loss reduction effect of 4.85%~5.21%, and effectively improved voltage quality (the lowest node voltage in the heavy load scenario was increased from 0.9434pu to 0.9528pu). Multi-time block reconfiguration only requires 30 switching actions to approach the loss reduction level of hourly reconfiguration, taking into account both system economy and the operational safety of switching equipment. Attached Figure Description
[0024] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a diagram of the graph convolutional neural network prediction model architecture of the present invention; Figure 3 This is a schematic diagram of the O1C1 local exchange operator of the present invention; Figure 4 This is the convergence curve of the training loss of the GNN model in the IEEE 119-node system of this invention; Figure 5 This invention illustrates the evolution trend of Top-K hit rate in the topology prediction of the IEEE 119-node system. Figure 6 This invention presents a comparison of voltage profiles and loss reduction effects before and after reconfiguration of a 136-node system under different scenarios. Detailed Implementation
[0025] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.
[0026] This embodiment provides a multi-period stochastic dynamic reconfiguration method for wind power distribution networks based on graph neural network-guided improved switch-on switching, aiming to overcome the problems of large-scale stochastic reconfiguration search space, low computational efficiency, and difficulty in guaranteeing the physical feasibility of radial topology in pure data-driven approaches in existing technologies. Specifically, as shown... Figure 1 As shown, the method includes the following steps: Step 1: Constructing a multi-period random reconstruction model.
[0027] This step primarily establishes a multi-period stochastic reconfiguration model with the goal of minimizing the total energy loss of the system throughout the day. It comprehensively considers AC power flow balance constraints, branch capacity constraints, node voltage range constraints, scenario probability constraints, wind power fixed power factor output model, and network radial topology and connectivity constraints. The dual uncertainties of load and wind power output are modeled as stochastic disturbances in each time period and scenario. In this embodiment, the branch corresponds to the tie switch.
[0028] The objective function of the multi-period stochastic reconstruction model is to minimize the total energy loss of the system. (1) in, This represents the total number of time periods, assuming each time period is 1 hour. A collection of scenes; For the set of all switch branches; for Time Period Scene Probability of occurrence; branch road The switch status is a binary variable, where 1 represents closed and 0 represents open; and They are nodes During the period Scene The voltage amplitude and phase angle under these conditions; and These are the resistance and reactance of the branch circuit, respectively. The unit time interval.
[0029] The objective is to target energy loss rather than instantaneous power because multi-time period optimization requires accumulating the losses across all time periods; given that Δt is constant, minimizing energy loss is equivalent to minimizing the sum of power losses across all time periods. The constraints to be satisfied include: Active and reactive power balance constraints at nodes -- AC power flow equations: Active power balance constraints: (2) Active power balance constraints: (3) in, and They are nodes Total active power output and active load demand; For nodes A set of connected nodes; and These are the real and imaginary parts of the admittance matrix, respectively. The voltage angle difference between nodes; and They are nodes Total reactive power output and reactive load demand.
[0030] Equations (2) and (3) describe the active and reactive power balance of each node in the system under any scenario.
[0031] Branch capacity (apparent power) limitation, i.e., branch Maximum capacity limit: (4) in, branch road Maximum apparent power limit allowed.
[0032] Node voltage range constraints: (5) in, and node The lower and upper limits of the permissible voltage amplitude.
[0033] Constraints are formed by the active and reactive power output of the power source: (6) (7) Among them, superscript and These represent conventional distributed power sources and wind power sources, respectively. The above formula shows that the active and reactive power generation of a node is composed of both conventional generators and wind power sources.
[0034] Sum of scenario probabilities constraint: (8) This constraint guarantees that the sum of the probabilities of all considered random scenarios is 1.
[0035] Wind power output characteristic modeling: This embodiment uses a fixed power factor model to describe the output characteristics of wind power, and its mathematical expression is as follows: (9) This is the preset wind power factor. In power flow calculations, nodes connected to wind power are considered PQ nodes. Radial network topology and connectivity constraints—after reconstruction, the network must always maintain a radial topology and avoid islanding: According to graph theory, to ensure that the distribution network maintains radial operation and avoids islanding after reconfiguration, the following radial topology and connectivity constraints must be satisfied: (10) (11) in, This represents the total number of nodes in the network. This represents the total number of branches that are in a closed state after reconstruction. Represents a node With the root node of the substation The topological distance between them. Equation (10) is a necessary condition for the formation of a radial structure in the network. Equation (11) ensures that there is a path of finite length between all nodes and the substation, ensuring the connectivity of the entire network and avoiding the generation of isolated nodes.
[0036] In summary, equations (1)-(11) construct a multi-period stochastic reconfiguration model for distribution network reconfiguration. This model aims to optimize the economic efficiency of system operation by optimizing binary decision variables, while satisfying complex AC power flow physical constraints and topological constraints.
[0037] Step 2: Construct the power distribution network diagram model and monitoring labels.
[0038] This step abstracts the distribution network into an undirected network topology graph, constructing a cross-sectional feature matrix based on the net active and reactive power injected into each node. Starting from the fully closed network topology, a Sequential Switching Operation (SSO) strategy iteratively selects branches from the set of loop switch branches that minimize network loss increments without violating connectivity and voltage constraints, and disconnects them until the radial condition is met. The generated optimal topology is then used as the true label to construct a training dataset. This step specifically includes: To use graph neural networks to address the topology optimization problem of power distribution networks, it is necessary to first model the physical system as a graph data structure and construct a supervised learning dataset.
[0039] Abstracting the power distribution network as an undirected network topology graph Node set The scale is Branch set correspond The nodes in this topology correspond to the busbars in the distribution network, including electrical nodes such as substation busbars, distribution transformer busbars, load access busbars, and distributed wind power grid-connected busbars. The substation node serves as the root node of the undirected network topology. Considering the dual uncertainties of load and wind power output, the time period... Scene Next node The disturbance load and wind power output model is as follows: (12) in, , This is the baseline load forecast value. This is the predicted benchmark output value for wind power. These are the random disturbance variables for load and wind power, respectively.
[0040] To characterize the actual power requirements of nodes, time periods are defined. Scene Next node Net injected active power With net injected reactive power They are respectively: (13) Based on this, a net load input characteristic matrix for the entire network under a specific cross-section is constructed. : (14) Among them, node feature vectors The equivalent power distribution under this cross section was recorded.
[0041] This embodiment defines a heuristic mapping function. Its input is the feature matrix. The tag generation process employs a Sequential Switching Operation (SSO) strategy, with the core optimization criterion as follows: starting from a fully closed mesh topology, i.e. Initially, the algorithm iteratively checks the current set of loop switch branches. Search for the optimal switching device : (15) To disconnect the switch The energy loss of the system at the current cross-section is given by equation (1). If disconnected... If the connectivity constraint of equation (11) is violated or the voltage limit is exceeded, then the following conditions are imposed: Maximum penalty value. In each round, the algorithm selects the branch that minimizes the increase in network loss. To cut off, that is, to place Until the radial operating conditions of equation (10) are met.
[0042] Note: The number of iterations in SSO is equal to the number of loops in the original fully switched closed network. In each iteration, a switch is selected and disconnected, i.e., a tie switch is selected. The selected switch must be located within a loop so that disconnecting the switch opens a loop but does not isolate any bus.
[0043] Using mapping relationships The topological states generated by SSO are combined in sequence to obtain the true label vectors for supervised learning by the model. : (16) The final constructed training dataset is represented as .
[0044] Step 3: Construct a graph convolutional neural network prediction model.
[0045] like Figure 2 As shown, this step involves Z-score standardization of the input features, followed by stacking multiple graph convolutional layers to aggregate neighborhood node information and extract spatial topological features. Then, a multilayer perceptron edge classifier outputs the probability of each branch being disconnected as a connection switch. A weighted binary cross-entropy loss function is used for offline training to address class imbalance caused by the sparsity of connection switches. Specifically, this step is as follows: This step establishes a parametric model (GNN model). The aim is to learn from the characteristics of high-dimensional load sections. To the branch switch opening probability vector The architecture consists of three parts: a feature preprocessing layer, a multi-layer graph convolutional layer, and an edge state classifier (or edge classifier for short).
[0046] Considering the significant differences in load amplitude between different buses, in order to eliminate the influence of dimensions and improve the convergence speed of the neural network, the input feature matrix is first... Perform Z-score standardization. For nodes... original features Features after processing for: (17) in, For nodes The standardized cross-sectional feature matrix; and These are the mean and standard deviation vectors of all cross-sections in the training set, respectively. This represents element-wise division; It is a very small constant.
[0047] Therefore, a standardized feature matrix for the entire network is constructed. As the initial input to the graph neural network: (18) The model aggregates neighborhood node information through stacked multi-layer graph convolutional networks to extract the spatial topology features of the power distribution network. In this embodiment, the number of layers is set to 3, denoted as the... The hidden feature matrix of the nodes in the layer is ,in This represents the feature dimension of the neurons in this layer. The inter-layer feature evolution mechanism follows: (19) in, Here, the adjacency matrix operator is symmetric and normalized, where To introduce the adjacency matrix of the self-loop structure, it is defined according to the initial topology of the distribution network. If the branch road , Otherwise, it is 0; Its corresponding degree matrix; For the first The learnable weight parameter matrix of the layer; for Activation function; It is an identity matrix.
[0048] After three layers of feature aggregation, the model outputs a final representation matrix that deeply integrates grid topology correlation and power injection information. The first in the matrix Row vectors represent nodes The high-dimensional embedding features are denoted as: (20) After obtaining refined node feature representations, this embodiment designs a multi-layer perceptron (MLP) module as an edge classifier for inferring switches. break probability This process can be described as follows: Design a multilayer perceptron (MLP) edge classifier by splicing features from the nodes at both ends of a branch: (twenty one) (twenty two) in, To connect branch roads The concatenated vector of features from both ends of the node; and These are the weight matrices and bias terms for each layer of the classifier, respectively. The function ensures the output probability The larger the value, the higher the probability that the branch will be disconnected as a tie switch under the current cross section.
[0049] Because the number of tie switches in a radial distribution network is far less than that of sectionalizing switches, the training samples exhibit extreme class imbalance. To guide the model in accurately capturing the sparse features of tie switches, a weighted binary cross-entropy loss function is used to guide offline training. (twenty three) Among them, the penalty coefficient Defined as: (twenty four) It is worth noting that the setting For larger numbers, the model will impose a larger gradient penalty on the behavior of false alarms or missed alarms of contact switches during backpropagation, maintaining high reconstruction accuracy under complex load scenarios. The specific architecture diagram is shown below. Figure 1 As shown.
[0050] Step 4: Probability-guided candidate generation and improved SOE local exchange optimization and global reconstruction.
[0051] This step first sorts the branch disconnection probabilities in descending order for the current time period and scenario, and selects the top K branches with the highest probabilities to construct a set of high-potential candidate interconnection switches. For each candidate switch, a sequential disconnection mapping is called to generate an initial radial topology. Then, local switching and global optimization are performed. The improved SOE O1C1 local switching operator is executed on the initial topology, focusing on the opening and closing linkage of the end segment switches (Type-2 switches). Independent loss reduction actions that do not belong to the same feeder are combined and executed synchronously to explore a better structure. A local optimum is searched within the legal topology set, and then the losses of each candidate branch are combined to obtain the globally optimal reconstructed topology. Specifically, this step is as follows: First, the current scene is obtained based on the pre-built GNN side classifier. Next period The predicted probability of each branch being interrupted. For simplicity, let the switch index... Corresponding node pairs Its prediction probability is denoted as .right All branches are arranged according to Sort the values in descending order and truncate the first value with the highest probability. Branch lines, constructing a set of high-potential candidate interconnection switches : in, For candidate switch indices, their corresponding node pairs , This is a function for sorting in descending order. (Set) The system was selected at the current power injection section. The candidate branch most likely to be disconnected to reduce active power loss is then identified. Subsequently, regarding... Each candidate switch in Topology iteration can be performed serially or in parallel across different threads. Force disconnect the switch. As an initial constraint, let Call the sequential switch on / off mapping operator defined earlier. Under the premise of satisfying equations (10) and (11), this process generates the corresponding initial radial topology (used to describe the switching state of each candidate switch), denoted as the decision vector. : (26) in, It includes the open / closed status of all switches in the distribution network under this branch.
[0052] To correct The local optimum error of sequential greedy disconnection, for Implement a local branch switching strategy. In a radial distribution network, the switching states of nodes near the root node of the substation usually have a significant impact on the global power flow, and their opening and closing states are relatively easy to determine accurately in the initial opening phase; while the switching states of nodes near the end of the network only have a minor adjustment effect on the local power flow, and conventional heuristic rules are very likely to get stuck in local optima at this point.
[0053] Therefore, the focus of local switching is on the end-section sectionalizing switch (which is a candidate switch). For example... Figure 3 Specifically, it is defined as a Type-2 switch, which is a closed switch that meets the following physical conditions: the number of segmented branches between its location and the downstream topology end node does not exceed a preset value. For this type of switch, this embodiment executes the O1C1 optimization action of the SOE algorithm. First, a certain end segment switch is disconnected. At this time, the feeder section downstream of the switch will be disconnected from the main network, forming an island. In order to restore the overall network connectivity of the system and strictly maintain the radial topology constraint specified by formula (10), only one backup tie switch connected to the island section can be closed to obtain the initial radial topology generated by all single O1C1 operations. After performing the one-on-one linkage operation, the total energy loss of the system defined by formula (1) is recalculated. If the loss after the operation is lower than before the operation, the group of operations is recorded as an effective loss reduction action.
[0054] Furthermore, considering that large distribution networks typically contain multiple main feeder branches directly derived from substations, and given the tree-like structure of the distribution network, O1C1 actions occurring on different main feeders (i.e., independent feeder regions) have minimal cross-influence on their respective power flow distributions. Therefore, to improve the optimization depth without causing combinatorial explosion, the algorithm extracts independent loss reduction actions that do not belong to the same feeder region and performs two or more independent actions synchronously and in combination to explore a better network structure.
[0055] Initial topology The radial topology set that can be generated after all legal O1C1 actions and their individual combinations is denoted as . Based on this, regarding candidate switches The calculated local optimal topology It can be solved by the following formula: (27) in, For set Any legal topology in; Inject power feature matrices into nodes within the current time period and scenario; The total energy loss evaluation function of the system is given by equation (1). , The objective function in the multi-period stochastic reconstruction model is to minimize the total energy loss, i.e., Equation (1).
[0056] Ultimately, combining The results are as follows. Under the premise of satisfying the constraints of equations (10) and (11), the minimum loss values of each path are compared. Current time period. With scene Global optimal topology reconstruction as follows: (28) In the formula, The objective function in the multi-period stochastic reconstruction model is to minimize the total energy loss, i.e., Equation (1).
[0057] Step 5: Multi-time segmented reconstruction.
[0058] The above completes the optimal topology search under a single time period and single scenario. In multi-time period operation, the impact of the number of switching operations on equipment lifespan also needs to be considered. This invention utilizes a GNN to process the cross-sectional features of each time period throughout the day, dividing the 24 hours into low-wind-power periods and high-wind-power periods based on wind power output levels. The same topology is maintained within each time period, requiring only two switching operations throughout the day. Compared to hourly reconstruction (188 switching operations), this significantly reduces the number of operations, and compared to a fixed topology throughout the day, it significantly reduces cumulative network losses, achieving a balance between economic efficiency and equipment safety.
[0059] The above methods can be summarized into the following algorithm: To verify the effectiveness of the above method, the following simulation analysis is performed in this embodiment: (1) Simulation settings To verify the effectiveness of the proposed GNN-SOE-based stochastic dynamic reconfiguration algorithm for distribution networks, this example was tested on standard IEEE 33-bus, 119-bus, and 136-bus systems. The simulation environment consisted of an Intel Core i7-9750h processor with 8GB of RAM, and the software platform was MATLAB R2022b with PyTorch 1.12.
[0060] In the data preparation phase, 1500 sets of random load scenarios were generated using the Matpower toolbox in MATLAB in conjunction with a random perturbation model as the training set. The load perturbations follow a uniform distribution. .
[0061] The specific structural parameters of the GNN model are shown in Table 1.
[0062] Table 1. GNN Model and Algorithm Parameter Settings Search width The value is set to 2 in the 33-node system, and to 5 and 10 in the 119 and 136-node systems, respectively. Comparison methods include: MP as the baseline optimal solution proposed in the paper "Optimal reconfiguration of electrical distribution systems using mathematical programming," and the SOE method proposed in the paper "Switch opening and exchange method for stochastic distribution network reconfiguration."
[0063] (2) Analysis of model training and prediction accuracy This embodiment sets "Top- The accuracy of predictions is evaluated using "hit rate," meaning that the handshake switches in the optimal reconstruction scheme are included before the predicted probability. The proportion of bits. Taking the IEEE 119-node system as an example, the model training convergence curve is as follows: Figure 4 As shown in the figure, after approximately 60 iterations, the system enters a convergence plateau, with the optimal loss value on the test set being 0.0881. Figure 5 This demonstrates the prediction accuracy trend, and the system's Top- The highest hit rate reached 96.27%, indicating that GNN can deeply explore the nonlinear mapping relationship between the power flow section of the distribution network and the optimal topology, and provide a high-quality candidate space.
[0064] (3) Reconstruction performance comparison The GNN-SOE algorithm was compared with the MP and SOE algorithms, and the results are shown in Table 2. Experiments revealed that in the 33-node system, GNN-SOE found the same globally optimal solution as MP, with a network loss error of 0%, outperforming the SOE algorithm. In the 119 and 136-node systems, the computational errors of nN-SOE were 0.05% and 0.37%, respectively. At this point, GNN predicts the branch breakage probability and sets the search width. It can cover the optimal solution space with a high probability, avoiding the problem of heuristic algorithms getting trapped in local optima while ensuring accuracy. The computation time of the reconstruction algorithm is an indicator of its real-time performance. As the system scale increases, the search space of the traditional SOE algorithm increases exponentially. At this time, GNN-SOE, guided by probability, eliminates most invalid search paths, reducing the computation time from 1345.8s to 308.35s, a speedup of 4.36 times; in the 136-node example, it also achieved a speedup of 3.98 times, indicating that the proposed algorithm can efficiently complete large-scale distribution network reconstruction.
[0065] Table 2 shows the different systems Value. How many to select? It depends on the reconstruction accuracy and computational cost. For a 33-node system, the calculated value of the optimal solution is set to... For a 136-node system, take It can effectively balance computation time and network loss optimization. The "TOP-" of the GNN model "Hit rate" makes it possible to achieve a relatively low hit rate. This includes a critical communication switch, thereby reducing the computational load.
[0066] Table 2. Performance comparison of the proposed method with the original SOE and benchmark methods (fully serial computation) Note: Errors are calculated relative to the MP benchmark; speedup is calculated relative to the original SOE calculation under the same environment. All methods employ single-core serial computation, and the hardware environment remains consistent.
[0067] Table 3 shows the combination of interconnecting switches for each example system under the optimal loss target.
[0068] Table 3 Optimal Handling Switch Combinations After Reconstruction for Each Example (4) Robustness analysis under multiple operating scenarios To verify the robustness of the algorithm under conditions of dual uncertainties in load and wind power output, this embodiment designed three typical operating scenarios for the 136-node system: Scenario S1 (load multiplier 1.00, PV multiplier 1.00), Scenario S2 (load multiplier 1.25, PV multiplier 0.60), and Scenario S3 (load multiplier 0.80, PV multiplier 1.40). Specific parameters are shown in Table 4 below. Scenario S2 represents the most severe voltage drop condition across the entire system, while Scenario S3 represents the potential risks of voltage exceeding limits and power backfeed. The comparison of reconstruction loss reduction effects and voltage optimization is as follows: Figure 6 As shown.
[0069] Table 4 Performance indicators of the 136-node system under different PV and load levels Experimental results show that the GNN-SOE algorithm can achieve a loss reduction effect of 4.85% to 5.21% in all test scenarios. Among them, the original network loss of the system in the heavy-load scenario S2 is 325.35kW, and the network loss is reduced to 16.94kW after reconstruction, which is the most significant loss reduction effect. This indicates that the algorithm has a good ability to optimize power distribution under high voltage reduction conditions.
[0070] Furthermore, refactoring also improves system voltage quality. Taking scenario S2 as an example, the node voltage profiles before and after refactoring are as follows: Figure 6 As shown. Before the reconfiguration, due to the heavy load, the voltage of some nodes at the end of the system (such as near 40 and 100) was lower than the safe threshold of 0.95 pu, posing a risk of exceeding the limit. After the reconfiguration, the topology was changed and the power flow was rearranged, and the voltage of all nodes in the network rose above the safe threshold. The voltage of the lowest node rose from 0.9434 pu to 0.9528 pu.
[0071] Based on the above simulation analysis, the algorithm can stably seek the economically optimal operating topology under the physical safety constraints of different loads and different wind power output sections, and has strong practicality and robustness.
[0072] In summary, this embodiment proposes a GNN-SOE-based solution method for the multi-time-scale stochastic dynamic reconfiguration problem of distribution networks with distributed wind power access. Leveraging the spatial feature extraction advantages of GNNs in processing graph-structured data, a nonlinear mapping is established between the node net power cross-section and the optimal radial topology. Intelligent reduction of the search space is achieved by accurately predicting tie-off switching probabilities. Furthermore, an improved SOE operator is coupled to ensure that the reconfiguration results satisfy physical constraints such as AC power flow and network topology. Simulation results show that on standard IEEE 33, 119, and 136-node systems, the proposed method significantly improves computational efficiency compared to SOE while maintaining high reconfiguration accuracy, and effectively improves voltage quality, demonstrating strong robustness and engineering application value. Future research will further explore the generalization and transfer performance of graph neural networks in heterogeneous topologies and strive to construct a multi-time-scale dynamic reconfiguration framework for complex distribution networks with the collaborative participation of energy storage and flexible loads.
[0073] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0074] 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 implemented 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. The solutions in the embodiments of the present invention can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.
[0075] 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.
[0076] 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.
[0077] 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 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0078] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.
[0079] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.
Claims
1. A method for multi-time-period stochastic dynamic reconfiguration of a wind power distribution network based on graph neural network-guided improved switchgear switching, characterized in that, Includes the following steps: Construct a multi-period stochastic reconfiguration model with the objective of minimizing total energy loss; Abstracting the power distribution network as an undirected network topology graph And based on the undirected graph Construct the cross-sectional feature matrix for each time period ,in Represents a set of nodes. Represents the set of branches. Indicates time period, This indicates the scenario, with each branch corresponding to a handover switch; Based on the cross-sectional feature matrix The pre-trained graph convolutional neural network prediction model is used to output the probability of each branch being disconnected as the connection switch. Based on the probability that all communication switches are disconnected, the switch with the highest probability is selected. For each branch, a candidate set of interconnecting switches is constructed, and based on the candidate set of interconnecting switches, an initial radial topology is generated using a sequential switching strategy, denoted as the decision vector. ; Based on the decision vector The model employs a multi-period random reconstruction method, using an improved switch-on-off switching method to perform local topology optimization and global topology reconstruction, thereby obtaining the globally optimal reconstructed topology and completing the random dynamic reconstruction process for each period.
2. The method for multi-time-period stochastic dynamic reconfiguration of a wind power distribution network based on graph neural network-guided improved switchgear switching as described in claim 1, characterized in that, The multi-time stochastic reconstruction model includes an objective function and corresponding constraints aimed at minimizing the total energy loss. The objective function is: In the formula, This represents the total number of time periods. For scene collection, For the set of all switch branches, for Time Period Scene probability of occurrence branch road The switch status is represented by a binary variable, where a value of 1 indicates the switch is closed and a value of 0 indicates it is open. and They are nodes During the period Scene The voltage amplitude and phase angle under these conditions, and These are the resistance and reactance of the branch circuit, respectively. The unit time interval; The constraints include: Nodal active power balance constraints: In the formula, and They are nodes Total active power output and active load demand, For nodes A set of connected nodes. For nodes During the period Scene Voltage amplitude under, and These are the real and imaginary parts of the admittance matrix, respectively. The voltage angle difference between nodes. For nodes During the period Scene The phase angle below; Nodal reactive power balance constraints: In the formula, and They are nodes Total reactive power output and reactive load demand; Branch capacity constraints: In the formula, branch road Maximum permissible apparent power limit; Node voltage range constraints: In the formula, and node Permissible lower and upper limits for voltage amplitude; Constraints are formed by the active and reactive power output of the power source: In the formula, and They are nodes The active power output and reactive power output of conventional distributed power sources. and They are nodes The active power output and reactive power output of wind power; Scenario probability constraints: In the formula, For the scene Probability of occurrence; Wind power output constraints: In the formula, The preset wind power operating power factor; Radial network topology and connectivity constraints: In the formula, This represents the total number of nodes in the network topology. This represents the total number of branches that are in a closed state after reconstruction. Represents a node With the root node of the substation The topological distance between them.
3. The method for multi-time-period stochastic dynamic reconfiguration of a wind power distribution network based on graph neural network-guided improved switchgear switching as described in claim 1, characterized in that, The cross-sectional feature matrix The construction steps include: Considering the dual uncertainties of load and wind power output, the construction period is... Scene Next node The disturbance load and wind power output model is expressed as: In the formula, , For nodes During the period Scene The active and reactive load demands are as follows: For nodes The active power output of the wind power source , For nodes During the period Baseline predicted values for active and reactive loads. These are the random disturbance variables for load and wind power, respectively. For nodes During the period Baseline predicted value of active power output of wind power; To characterize the actual power demand of nodes, time periods are defined based on the aforementioned disturbance load and wind power output model. Scene Next node Net injected active power With net injected reactive power , respectively represented as: In the formula, For nodes The reactive power output of the wind power source; Based on the net injected active power With net injected reactive power Construct the cross-sectional feature matrix , is represented as: in: , In the formula, For nodes The cross-sectional feature matrix, subscript This represents the total number of nodes in the network.
4. The method for multi-time-period stochastic dynamic reconfiguration of a wind power distribution network based on graph neural network-guided improved switchgear switching as described in claim 1, characterized in that, The graph convolutional neural network prediction model includes a preprocessing layer, multiple graph convolutional layers, and an MLP-based edge classifier. The execution steps of the graph convolutional neural network prediction model include: In the preprocessing layer, the cross-sectional feature matrix is... Z-score standardization is performed to obtain the standardized feature matrix of the entire network. , is represented as: in: , In the formula, For nodes The standardized cross-sectional feature matrix, For nodes The cross-sectional feature matrix, , Let be the vector of mean and standard deviation for all cross-sections. For element-wise division, It is a very small constant; Based on the network-wide standardized feature matrix The multi-layer graph convolutional layer is used to aggregate neighborhood node information to obtain the final representation matrix that integrates power grid topology correlation and power injection information. ,in This represents the total number of convolutional layers in the graph. Based on the final representation matrix The MLP-based edge classifier is used to infer each branch as a tie switch. Probability of being disconnected .
5. The method for multi-time-period stochastic dynamic reconfiguration of a wind power distribution network based on graph neural network-guided improved switchgear switching, as described in claim 4, is characterized in that... The final representation matrix The steps to obtain it include: The network-wide standardized feature matrix As input to a multi-layer graph convolutional layer, it utilizes an inter-layer feature evolution mechanism to aggregate neighborhood node information, thereby enabling features to evolve from the first layer... layer to the first Layer update propagation, wherein the inter-layer feature evolution mechanism follows: In the formula, , The first Layer, First The hidden feature matrix of the nodes in the layer, for Activation function For symmetric normalized adjacency matrix operators, for The degree matrix, To introduce the adjacency matrix of the self-loop structure, Given an adjacency matrix defined based on an undirected network topology graph, when the branch , Otherwise, it is 0. For the set of all switch branches, for The Line 1 Column elements, for The Line 1 Column elements, It is the identity matrix. For the first The learnable weight matrix of the layer; go through The aggregation operation of the layered graph convolutional layer ultimately outputs a final representation matrix that integrates grid topology correlation and power injection information. The final representation matrix The first in Row vectors represent nodes The high-dimensional embedding features are denoted as: In the formula, For nodes High-dimensional embedding features, for The first in Row vectors.
6. The method for multi-time-period stochastic dynamic reconfiguration of a wind power distribution network based on graph neural network-guided improved switchgear switching as described in claim 5, characterized in that, The inference process of the MLP-based edge classifier is represented as follows: In the formula, To connect branch roads The concatenated vector of features from both ends of the node. For nodes High-dimensional embedding features, , , , These are the weight matrices and bias terms for each layer of the edge classifier, respectively.
7. The method for multi-time-period stochastic dynamic reconfiguration of a wind power distribution network based on graph neural network-guided improved switchgear switching as described in claim 1, characterized in that, The decision vector The steps to obtain it include: The candidate contact switch set is set as follows: ,in For candidate tie switch indexes, their corresponding branches , For all switch branch sets, This is a function for sorting in descending order. For the first The probability of a candidate contact switch being disconnected; For the candidate set of handshake switches Each candidate contact switch in In serial or parallel topology iteration across different threads, an initial radial topology satisfying the network's radial topology and connectivity constraints is generated using a sequential switching strategy, denoted as the decision vector. One of the topological iteration processes includes: Force disconnect the candidate contact switch As the initial constraint for this topology iteration, namely , For candidate handover switch State variables, Indicates a disconnected state; The initial constraints, the radial topology of the network, and the connectivity constraints are taken as input, and the sequential switch on / off strategy is called to generate the current initial radial topology, which is used as the decision vector for this operation. The decision vector Represented as: In the formula, This is a sequential switch on / off mapping operator.
8. The method for multi-time-period stochastic dynamic reconfiguration of a wind power distribution network based on graph neural network-guided improved switchgear switching as described in claim 1, characterized in that, The steps of the local topology optimization include: Based on the decision vector The end-segment switches are selected as Type-2 switches, wherein the end-segment switches are candidate tie switches, are in a closed state, and the number of segmented branches between their downstream nodes and the topology end nodes does not exceed a preset threshold. ; For each of the Type-2 switches, disconnect it to make its downstream area an island; To restore network connectivity and satisfy the network radial topology constraints, only one backup contact switch connected to the island is closed to obtain the initial radial topology generated by all single O1C1 operations. For each single O1C1 operation, the total energy loss is recalculated based on the initial radial topology. If the recalculated total energy loss is lower than the total energy loss before the operation, the O1C1 operation is considered an effective energy loss reduction action. Extract independent and effective loss reduction actions that do not belong to the same feeder region. Simultaneously execute two or more effective loss reduction actions in combination. Based on the single O1C1 operation and the combined O1C1 operation, a final radial topology set is formed, denoted as . ; In the radial topology set Within this process, the multi-time-period random reconstruction model is used for screening to obtain candidate contact switches. Local optimal topology , is represented as: In the formula, Radial topological set Any legal topology within, To the current cross-sectional feature matrix The total energy loss evaluation function is as follows: This is the objective function in the multi-period stochastic reconstruction model, which aims to minimize total energy loss.
9. The method for multi-time-period stochastic dynamic reconfiguration of a wind power distribution network based on graph neural network-guided improved switchgear switching as described in claim 1, characterized in that, The steps for obtaining the globally optimal reconstructed topology include: For all candidate contact switches Corresponding local optimal topology Under the constraints of radial topology and connectivity, compare the corresponding total energy loss values and find the local optimal topology with the minimum total energy loss value. As the globally optimal topology reconstruction , is represented as: In the formula, For the candidate set of handover switches, This is the total energy loss evaluation function. This is the objective function in the multi-period stochastic reconstruction model, which aims to minimize total energy loss.
10. The method for multi-time-period stochastic dynamic reconfiguration of a wind power distribution network based on graph neural network-guided improved switchgear switching as described in claim 1, characterized in that, The graph convolutional neural network prediction model uses a weighted binary cross-entropy loss function to guide offline training during the training process. The weighted binary cross-entropy loss function is expressed as follows: Among them, the penalty coefficient Represented as: In the formula, For the total loss, For the set of all switch branches, For the contact switch The true label, For the contact switch Predicted labels, For the set constant, The real labels are generated using a sequential switching strategy, and the generation process of the real labels includes: Assuming all branches in the distribution network are in a closed state, construct a fully closed network topology, denoted as . , branch road The closed state, where a value of 1 indicates the closed state. For the set of all switch branches; Based on the fully closed network topology, a sequential switch-on / off strategy is used to iteratively determine the current set of loop switch branches. Search for the optimal switching device The number of iterations is the number of loops in the fully closed network topology, and the branch that minimizes the increase in energy loss is selected in each iteration. To break, set to =0, until the radial topology constraint of the network is satisfied, the optimal on / off switch. Represented as: In the formula, For loop switch branch collection The switch in the middle, To disconnect the switch Then in the current cross-sectional feature matrix Energy loss; After the iteration ends, a loop-free network topology is obtained, and each branch is generated using the mapping relationship. The corresponding real labels form the real label vector. , is represented as: In the formula, For sequential switch on / off mapping operators, These are real labels; a value of 1 indicates a closed state, and a value of 0 indicates an open state. This is the set of all switch branches.