A method and system for power supply restoration after distribution network faults based on graph attention mechanism

Abstract

The present invention discloses a method and system for power supply restoration in the event of a distribution network fault based on a graph attention mechanism, and belongs to the field of intelligent operation and maintenance technology for distribution networks. The method comprises: performing Thiessen polygon division on the distribution network area according to the fault point to obtain the initial fault division area; performing island division according to the initial fault division area; combining the graph attention mechanism with a bidirectional deep recurrent Q network according to the island division result to solve the optimal distribution network power supply restoration strategy; executing the optimal distribution network power supply restoration strategy to realize power supply restoration in the event of a distribution network fault based on the graph attention mechanism. The present invention improves the convergence efficiency of the reinforcement learning strategy, reduces the action search space, and reduces the computational complexity of the global search strategy; the present invention integrates the graph attention mechanism with the bidirectional deep recurrent Q network, breaks through the defect of insufficient modeling of topological correlation in traditional algorithms, accurately captures the implicit coupling relationship of power interaction between islands, and improves the robustness of the strategy.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent operation and maintenance of distribution networks, and specifically relates to a distribution network fault power supply recovery method and system based on a graph attention mechanism. Background Art

[0002] Existing distribution network restoration technologies still have significant limitations in fault zoning, islanding, and restoration strategy optimization. Traditional islanding models employ random search or global traversal strategies for islanding and optimization, generating numerous invalid solutions (such as disconnected islands or out-of-limit solutions) and requiring computational time of minutes. Traditional methods require traversing the entire interconnector switch operation domain, resulting in a highly complex action space and computational efficiency that cannot meet the minute-by-minute restoration requirements.

[0003] At the dynamic decision-making level, existing deep reinforcement learning algorithms have inherent flaws in modeling distribution network topology dependencies. Traditional Q-learning or unidirectional LSTM networks struggle to capture the dynamic coupling between node electrical properties and neighborhood states, causing the strategy to ignore implicit constraints on inter-island power interactions. Furthermore, unidirectional temporal dependency modeling fails to simultaneously account for the impact of preceding operations and the potential for subsequent recovery, resulting in poor coordination of tie-breaker action sequences and low strategy robustness.

[0004] Although existing research has attempted to introduce graph neural networks, it has not combined bidirectional temporal feature extraction, and cannot solve the local optimal trap of action decision-making, resulting in a high strategy failure rate in complex fault scenarios. Summary of the Invention

[0005] To address the deficiencies in the prior art, the present invention provides a method and system for power supply restoration in distribution network failures based on a graph attention mechanism. By considering the impact on power supply restoration during the distribution network emergency repair process and the impact on the next emergency repair work during the power restoration process, a comprehensive overall optimal solution that takes into account both emergency repair and restoration is obtained, thereby guiding the emergency repair and restoration work of the distribution network.

[0006] The present invention adopts the following technical solutions.

[0007] A first aspect of the present invention provides a method for power supply restoration after a distribution network fault based on a graph attention mechanism, comprising:

[0008] The distribution network area is divided into Thiessen polygons according to the fault point to obtain the initial fault division area;

[0009] Island division is performed based on the initial fault area;

[0010] Based on the islanding results, the graph attention mechanism is combined with a bidirectional deep recurrent Q network to solve the optimal distribution network power restoration strategy.

[0011] Execute the optimal distribution network power restoration strategy and realize distribution network fault power restoration based on graph attention mechanism.

[0012] Preferably, the method of performing Thiessen polygon division on the distribution network area according to the fault point to obtain the initial fault division area specifically includes:

[0013] Set the fault point, line end, and tie switch node as the vertices of the Thiessen polygon;

[0014] Input the vertices into the Delaunay algorithm to generate a triangulated network;

[0015] Connect the centers of the circumscribed circles of the triangles in the generated triangulated network to form the Thiessen polygon boundary and obtain the Thiessen polygon sub-region;

[0016] The Thiessen polygon sub-regions at both ends of the tie switch are merged to obtain the initial fault partition region.

[0017] Preferably, the island division according to the initial fault division area specifically includes:

[0018] The loads within each initial fault partition area form an initial island set;

[0019] Based on the initial island set, an island partitioning model is constructed with the goal of maximizing power restoration.

[0020] Solve the island partition model to obtain the island partition result.

[0021] Preferably, the method combines the graph attention mechanism with the bidirectional deep recurrent Q network based on the islanding results to solve the optimal distribution network power restoration strategy, specifically including:

[0022] Extract the island adjacency matrix and island load characteristics based on the island division results, and construct the island graph data;

[0023] According to the island graph data, the global state vector of each island is extracted;

[0024] The global state vector of each island is used to make sequential decisions through a deep recurrent Q network based on a bidirectional long-short-term neural network to obtain the action instructions of each island.

[0025] Execute the action instructions of each island and calculate the reward function to obtain the next state, immediate reward and termination flag;

[0026] The target Q value is calculated for the new state, immediate reward and termination flag, and the TD error is minimized. The calculation is repeated until the termination condition is reached, and the optimal distribution network power restoration strategy is obtained.

[0027] Preferably, extracting the island adjacency matrix and island load characteristics according to the island division result and constructing the island graph data specifically includes:

[0028] Extract each isolated node set and edge set;

[0029] Construct an adjacency matrix based on the node set and edge set;

[0030] Extract the initial node features in each island;

[0031] Based on the adjacency matrix and initial node features, the graph structure data of each island is constructed.

[0032] Preferably, extracting the global state vector of each island based on the island graph data specifically includes:

[0033] Calculate attention weights based on island graph data;

[0034] According to the normalized attention weight of neighbor node j on node i, combined with the linear transformation of the neighbor features, a nonlinear activation function is input, and the aggregated features of all attention heads are repeatedly calculated to obtain the aggregated result of the neighbor features of the i-th node;

[0035] Based on the aggregation results of the neighbor features of the i-th node, the aggregation results of the neighbor features of all nodes in each island are globally pooled to obtain the global state vector of each island.

[0036] Preferably, the calculating of the attention weight according to the island map data specifically includes:

[0037] When there is a power line between node i and node j in the adjacency matrix, the corresponding node is determined to be a neighbor node of node i. The features of node i and node j in the initial node features are mapped to a high-dimensional space through a linear transformation matrix, and then the two mapping results are spliced;

[0038] Transpose the attention parameter vector and do a dot product with the concatenated result to get a scalar value;

[0039] Input the scalar value into the activation function to get the raw attention score of neighbor node j to node i;

[0040] According to the original attention score of neighbor node j to node i, the normalized attention weight of neighbor node j to node i is solved.

[0041] Preferably, solving the normalized attention weight of neighbor node j to node i based on the original attention score of neighbor node j to node i specifically includes:

[0042] The attention scores of all neighboring nodes directly connected to node i through power lines are summed by mapping the exponential results of the exponential function exp(.) to the positive domain, and the summed result is set as the denominator;

[0043] The exponential result of the original attention score of neighbor node j to node i is set as the numerator to obtain the normalized attention weight of neighbor node j to node i.

[0044] Preferably, the global state vector of each island is subjected to a time sequence decision by a deep recurrent Q network based on a bidirectional long short-term neural network to obtain an action instruction for each island, specifically including:

[0045] A bidirectional long-term and short-term neural network is independently constructed for each island, and the global state vector of each island is input into the corresponding bidirectional long-term and short-term neural network to obtain the time series state vector of each island;

[0046] Calculate the Q value based on the time series state vector of each island to obtain the expected cumulative reward value of each action;

[0047] According to the expected cumulative reward value of each action, the action is selected through the ε-greedy strategy to obtain the action instructions of each island.

[0048] The second aspect of the present invention provides a distribution network fault power supply restoration system based on a graph attention mechanism, which runs the distribution network fault power supply restoration method based on a graph attention mechanism described in the first aspect, including:

[0049] The area division module is used to divide the distribution network area into Thiessen polygons according to the fault point to obtain the initial fault division area;

[0050] The island division module is used to divide the islands according to the initial fault division area;

[0051] The power restoration strategy solving module is used to solve the optimal distribution network power restoration strategy based on the islanding results by combining the graph attention mechanism with the bidirectional deep recurrent Q network;

[0052] The output module is used to execute the optimal distribution network power supply restoration strategy and realize distribution network fault power supply restoration based on the graph attention mechanism.

[0053] The beneficial effects of the present invention are as follows:

[0054] The partitioning mechanism of the present invention is based on Thiessen polygons and Delaunay triangulation, which naturally meets the "power-load spatial proximity" requirement, breaks through the limitations of traditional empirical partitioning, provides a physically feasible initial solution set for subsequent island optimization, avoids the generation of invalid solutions caused by random search in traditional island partitioning, and reduces the invalid solution generation rate.

[0055] The adjacency matrix constructed based on the island partitioning results of the Thiessen polygon boundary can accurately reflect the actual electrical connections between nodes, provide highly reliable topological correlation features for the subsequent weight calculation of the graph attention mechanism, improve the convergence efficiency of the reinforcement learning strategy, and only need to make action decisions for the tie switches at the boundaries of the Thiessen sub-region, reducing the action search space and the computational complexity of the global search strategy.

[0056] The present invention integrates the graph attention mechanism and the bidirectional deep recurrent Q network, overcoming the defect of traditional algorithms in insufficient modeling of topological correlation. The graph attention mechanism accurately captures the implicit coupling relationship of power interaction between islands by aggregating node electrical properties and neighborhood correlation characteristics through multiple attention heads; the bidirectional LSTM network synchronously captures the impact of previous operations and the potential for subsequent recovery, realizing the coordinated optimization of the timing of interconnection switch actions. The decision-making process considers the global optimality of state timing dependency, thereby improving the robustness of the strategy. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 This is a flow chart of the distribution network restoration process proposed in an embodiment of the present invention. DETAILED DESCRIPTION

[0058] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. The described embodiments are only part of the embodiments of the present invention, not all of them. Based on the spirit of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts are within the scope of protection of the present invention.

[0059] like Figure 1 As shown, embodiment 1 of the present invention provides a method for power supply restoration after a distribution network fault based on a graph attention mechanism, comprising the following steps:

[0060] Step 1: divide the distribution network area into Thiessen polygons according to the fault points and tie switches to obtain Thiessen polygon sub-areas.

[0061] In a preferred but non-limiting embodiment of the present invention, step 1 comprises:

[0062] Step 1.1, set the fault point, line end, and tie switch node as the vertices of the Thiessen polygon;

[0063] Step 1.2: Input the vertices described in step 1.1 into the Delaunay algorithm to generate a triangulated network.

[0064] Make sure that the circumcircle of each triangle contains no other vertices.

[0065] Step 1.3, connecting the centers of the circumscribed circles of the triangles in the triangulated network generated in step 1.2 to form a Thiessen polygon boundary, thereby obtaining a Thiessen polygon subregion;

[0066] In step 1.4, merge the Thiessen polygon sub-regions at both ends of the tie switch to obtain the initial fault partition area.

[0067] Step 2: divide the islands based on the initial fault area.

[0068] In a preferred but non-limiting embodiment of the present invention, step 2 comprises:

[0069] Step 2.1: The loads in each initial fault partition area form an initial island set;

[0070] In step 2.2, based on the initial island set, an island partitioning model is constructed with the goal of maximizing power restoration.

[0071] Further preferably, step 2.2 includes:

[0072] Step 2.2.1: Based on the initial island set, with the goal of maximizing power restoration, construct the objective function for island partitioning, which can be expressed as the following formula:

[0073]

[0074] Where,

[0075] Represents the set of all load nodes in the distribution network that need to restore power supply;

[0076] Represents a load node The grading weight of

[0077] Represents a load node Whether power supply has been restored;

[0078] Represents a load node power requirements.

[0079] Step 2.2.2: Construct the constraints for island partitioning.

[0080] More preferably, step 2.2.2 comprises:

[0081] Set voltage limit constraints so that the voltage amplitudes of all nodes in the island do not exceed the lower and upper voltage limits.

[0082] Line load rate constraint: the power transmission of each line must meet the requirement that the active and reactive power of the line does not exceed the maximum carrying capacity of the line;

[0083] Output upper and lower limit constraints: the active output of distributed generation (DG) must operate within the safe range;

[0084] All load nodes in the island must be connected to at least one power node through a closed line.

[0085] Step 3: Based on the islanding results of step 2, the graph attention mechanism is combined with the bidirectional deep recurrent Q network to solve the optimal distribution network power restoration strategy;

[0086] In a preferred but non-limiting embodiment of the present invention, step 3 comprises:

[0087] Step 3.1: Extract the island adjacency matrix and island load characteristics based on the island partition results and construct the island graph data.

[0088] Further preferably, step 3.1 includes:

[0089] Step 3.1.1, extract each island A collection of nodes V k and edge sets E k , where the edge set E k For the k The set of all power lines within an island, and each edge represents the physical connection between two nodes.

[0090] Step 3.1.2: Construct an adjacency matrix based on the node set and edge set ,in, A k [i][j] Indicates the k Whether the i-th row node and the j-th column node in the node set of the isolated island have power lines, N k is the number of nodes of the kth island.

[0091] Step 3.1.3: Extract the initial node features within each island ,in, Indicates the load type, Indicates the load level, Indicates the recovery status of the node.

[0092] Step 3.1.4: Construct the graph structure data of each island based on the adjacency matrix of step 3.1.2 and the initial node features of step 3.1.3 .

[0093] Step 3.2: Extract the global state vector of each island based on the island graph data in step 3.1.

[0094] Further preferably, step 3.2 includes:

[0095] Step 3.2.1, calculate the attention weight based on the island map data.

[0096] More preferably, step 3.2.1 comprises:

[0097] When there is a power line between node i and node j in the adjacency matrix, the corresponding node is determined to be the neighbor node of node i, and the feature of node i in the initial node feature is h i and the characteristics of node j h j Through the linear transformation matrix W m Map to a high-dimensional space, and then concatenate the two mapping results;

[0098] Transpose the attention parameter vector and do a dot product with the concatenated result to get a scalar value;

[0099] The scalar value is input into the activation function to obtain the raw attention score of neighbor node j to node i, which is expressed as follows:

[0100]

[0101] Where,

[0102] represents the original attention score of the current neighbor node j to node i;

[0103] W m represents the linear transformation matrix;

[0104] h i represents the characteristics of node i, h j represents the characteristics of node j;

[0105] represents the activation function;

[0106] According to the original attention score of neighbor node j to node i, the normalized attention weight of neighbor node j to node i is solved, which includes:

[0107] The attention scores of all neighbor nodes directly connected to node i through power lines are calculated by exponential function The exponential results after being mapped to the positive number domain are summed, and the summed result is set as the denominator;

[0108] The exponential result of the original attention score of neighbor node j to node i is set as the numerator to obtain the normalized attention weight of neighbor node j to node i, which is expressed as the following formula:

[0109]

[0110] Where,

[0111] represents the normalized attention weight of neighbor node j to node i;

[0112] represents the original attention score of neighbor node j to node i, which is not normalized;

[0113] represents the attention score of the lth neighbor node directly connected to node i through the power line;

[0114] represents the set of all neighbor nodes directly connected to node i through power lines;

[0115] In step 3.2.2, based on the normalized attention weight of neighbor node j to node i, combined with the linear transformation of the neighbor features, a nonlinear activation function is input, and the aggregated features of all attention heads are repeatedly calculated to obtain the aggregated results of the neighbor features of the i-th node.

[0116] More preferably, step 3.2.2 comprises:

[0117] For each neighbor node j, use the linear transformation matrix W m Pair Features h j Perform linear transformation;

[0118] Weight the transformation result according to the normalized attention weight of the corresponding neighbor node j on node i;

[0119] Sum the weighted results of all neighbor nodes and input a nonlinear activation function to obtain the aggregated neighbor features of a single attention head;

[0120] Compute the aggregated neighbor features of all attention heads in parallel to obtain the first The aggregation result of the neighbor features of a node is expressed as follows:

[0121]

[0122] Where,

[0123] Represents the aggregation result of the neighbor features of the i-th node;

[0124] m represents the mth attention head;

[0125] K represents the number of attention heads.

[0126] In step 3.2.3, based on the aggregation results of the neighbor features of the i-th node, the aggregation results of the neighbor features of all nodes in each island are globally pooled to obtain the global state vector of each island.

[0127] Further preferably, step 3.2.3 includes:

[0128] Based on the aggregation results of the neighbor features of the i-th node, the aggregation results of the neighbor features of all nodes in the k-th island are weighted summed and averaged to obtain the global state vector of the k-th island, which is expressed as the following formula:

[0129]

[0130] Where,

[0131] represents the global state vector of the island;

[0132] is the number of nodes of the kth island;

[0133] Then, based on the global state vector of the kth island, the global state vectors of all islands are solved.

[0134] In step 3.3, the global state vector of each island is subjected to a time-series decision-making based on a bidirectional long-term short-term neural network (DRQN) to obtain the action instructions of each island.

[0135] Further preferably, step 3.3 includes:

[0136] In step 3.3.1, a bidirectional long-term short-term neural network is independently constructed for each island, and the global state vector of each island is input into the corresponding bidirectional long-term short-term neural network to obtain the time series state vector of each island.

[0137] More preferably, step 3.3.1 comprises:

[0138] Input the global state vector of each island into the corresponding forward LSTM network to obtain the forward hidden state;

[0139] Input the global state vector of each island into the corresponding backward LSTM network to obtain the backward hidden state;

[0140] The forward hidden state and the backward hidden state are combined to obtain the temporal state vector of each island, which is expressed as follows:

[0141]

[0142] Where,

[0143] is the time series state vector of the kth island at time step t;

[0144] is the forward hidden state, indicating the state dependency before t time steps;

[0145] is the backward hidden state, which represents the potential trend after t time steps.

[0146] Step 3.3.2: Calculate the Q value based on the time series state vector of each island in step 3.3.1 to obtain the expected cumulative reward value of each action;

[0147] More preferably, step 3.3.2 comprises:

[0148] Set the bias vector according to load priority;

[0149] The temporal state vector of each island is mapped to the action space through the weight matrix and summed with the bias vector to obtain the expected cumulative reward value of each action, which is expressed as the following formula:

[0150]

[0151] Where,

[0152] Indicates action In state The expected cumulative reward value under

[0153] represents the weight matrix;

[0154] Represents the bias vector used to adjust the action preference based on the high priority load.

[0155] In step 3.3.3, based on the expected cumulative reward value of each action in step 3.3.2, the action selection is performed using the ε-greedy strategy to obtain the action instructions for each island, which can be expressed as the following formula:

[0156]

[0157] Where,

[0158] represents the action chosen by the kth island at time step t;

[0159] Indicates choosing action a that maximizes the Q value;

[0160] represents the probability of randomly selecting an action; represents the probability of choosing the current optimal action.

[0161] Step 3.4: Execute the action instructions of each island in step 3.3 and calculate the reward function to obtain the next state, immediate reward and termination flag.

[0162] Further preferably, step 3.4 includes:

[0163] Step 3.4.1, execute the action instructions of each island, update the global state vector, and set the new global state vector as the next state.

[0164] More preferably, step 3.4.1 comprises:

[0165] Execute the action instructions of each island, update the adjacency matrix and node features, generate new island graph data based on the updated adjacency matrix and node features, and then update the global state vector based on the new island graph data to obtain the next state.

[0166] For example, but not limited to, closing the tie switch to connect different islands. If the action causes the islands to merge, re-execute the Thiessen polygon partitioning in step 1.4, update the adjacency matrix and node features, and generate new island graph data. The adjacency matrix and node features in the new island graph data will be used as input for the next time step (t+1).

[0167] Step 3.4.2, calculate the reward function parameters based on the voltage constraint and line load rate, and get the value of each island. The instantaneous reward at a time step is expressed as follows:

[0168]

[0169] Where,

[0170] represents the immediate reward of the kth island at time step t;

[0171] represents the load priority weight vector, normal load: , important loads: , critical load ;

[0172] Indicates the recovery status of the i-th node, 0 = not recovered, 1 = recovered;

[0173] represents the load power penalty coefficient, the microgrid output of the i-th node Exceeding rated capacity 10 points will be penalized for every 1kW over the limit;

[0174] Indicates excess power;

[0175] represents the node voltage penalty coefficient;

[0176] Represents the indicator function, the node voltage of the i-th node Over the time limit ( ) is 1, otherwise it is 0, is the lower limit of the node voltage, is the upper limit of the node voltage;

[0177] is the number of nodes in the kth island.

[0178] In step 3.4.3, when the total amount of restored load is greater than the set threshold, it is set as a termination flag.

[0179] In step 3.5, the target Q value is calculated for the new state, immediate reward, and termination flag obtained in step 3.4, and the TD error is minimized. The calculation is repeated until the termination condition is reached, and the optimal distribution network power restoration strategy is obtained.

[0180] Further preferably, step 3.5 includes:

[0181] Step 3.5.1: Calculate the target Q value for the new state, immediate reward, and termination flag obtained in step 3.4, expressed as the following formula:

[0182]

[0183] Where,

[0184] is the target Q value, which represents the ideal Q value of island k at time step t,

[0185] represents the discount factor,

[0186] Indicates the termination flag, 1 = end of round, 0 = continue,

[0187] represents the maximum Q value of all possible actions in the next state predicted by the target network,

[0188] Represents the global state vector after the kth island is updated, Set to the next state,

[0189] Indicates the action for the next state,

[0190] In step 3.5.2, the loss function is calculated for the target Q value of step 3.5.1, the online network parameters are updated, and steps 3.1-3.5 are repeated until the termination condition is met. The optimal distribution network power restoration strategy is obtained, which is expressed as the following formula:

[0191]

[0192] Where,

[0193] represents the loss function;

[0194] B represents the batch size;

[0195] b represents the batch index, indicating that the bth sequence in the batch is currently being processed;

[0196] L is the length of the time series, which indicates the number of consecutive time steps contained in each sequence;

[0197] t is the time step index, indicating the tth time step in the sequence;

[0198] Indicates the current network status Next action Q value prediction, the online network parameters are .

[0199] Step 4: Execute the load power supply restoration in step 3 and the emergency repair in step 4, and finally obtain the distribution network fault emergency repair and restoration coordination plan to achieve distribution network fault recovery.

[0200] The beneficial effects of the present invention include at least: utilizing the functions of the distribution network itself, considering the goals of restoring important loads and restoring as many loads as possible and reducing the number of switch operations, changing the topology of the distribution network, and restoring power supply to some non-fault points; the emergency repair of the distribution network fault is to first restore the power supply to the distribution network after the fault, and then consider arranging the repair sequence more reasonably so that the fault of the distribution network can be repaired faster and more economically, and the normal structure of the distribution network can be restored as soon as possible; by considering the impact on power supply restoration during the distribution network emergency repair process and the impact on the next emergency repair work during the power supply restoration process, a comprehensive overall optimal solution that takes into account both emergency repair and restoration is obtained, thereby minimizing various economic and social losses caused by the fault.

[0201] Example 2

[0202] Embodiment 2 of the present invention provides a distribution network fault power supply restoration system based on a graph attention mechanism, which is used to execute the distribution network fault power supply restoration method based on a graph attention mechanism described in embodiment 1, including:

[0203] The area division module is used to divide the distribution network area into Thiessen polygons according to the fault point to obtain the initial fault division area;

[0204] The island division module is used to divide the islands according to the initial fault division area;

[0205] The power restoration strategy solving module is used to solve the optimal distribution network power restoration strategy based on the islanding results by combining the graph attention mechanism with the bidirectional deep recurrent Q network;

[0206] The output module is used to execute the optimal distribution network power supply restoration strategy and realize distribution network fault power supply restoration based on the graph attention mechanism.

[0207] The beneficial effects of the present invention are as follows:

[0208] The partitioning mechanism of the present invention is based on Thiessen polygons and Delaunay triangulation, which naturally meets the "power-load spatial proximity" requirement, breaks through the limitations of traditional empirical partitioning, provides a physically feasible initial solution set for subsequent island optimization, avoids the generation of invalid solutions caused by random search in traditional island partitioning, and reduces the invalid solution generation rate.

[0209] The adjacency matrix constructed based on the island partitioning results of the Thiessen polygon boundary can accurately reflect the actual electrical connections between nodes, provide highly reliable topological correlation features for the subsequent weight calculation of the graph attention mechanism, improve the convergence efficiency of the reinforcement learning strategy, and only need to make action decisions for the tie switches at the boundaries of the Thiessen sub-region, reducing the action search space and the computational complexity of the global search strategy.

[0210] The present invention integrates the graph attention mechanism and the bidirectional deep recurrent Q network, overcoming the defect of traditional algorithms in insufficient modeling of topological correlation. The graph attention mechanism accurately captures the implicit coupling relationship of power interaction between islands by aggregating node electrical properties and neighborhood correlation characteristics through multiple attention heads; the bidirectional LSTM network synchronously captures the impact of previous operations and the potential for subsequent recovery, realizing the coordinated optimization of the timing of interconnection switch actions. The decision-making process considers the global optimality of state timing dependency, thereby improving the robustness of the strategy.

[0211] The present disclosure may be a system, method and / or computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for causing a processor to implement various aspects of the present disclosure.

[0212] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered by the scope of protection of the claims of the present invention.

Claims

1. A method for power supply restoration in distribution network failure based on graph attention mechanism, characterized by: The distribution network area is divided into Thiessen polygons according to the fault point to obtain the initial fault division area; Island division is performed based on the initial fault area; Based on the islanding results, the graph attention mechanism is combined with a bidirectional deep recurrent Q network to solve the optimal distribution network power restoration strategy, including: Extract the island adjacency matrix and island load characteristics based on the island division results, and construct the island graph data; According to the island graph data, the global state vector of each island is extracted; The global state vector of each island is used to make sequential decisions through a deep recurrent Q network based on a bidirectional long-short-term neural network to obtain the action instructions of each island. Execute the action instructions of each island and calculate the reward function to obtain the next state, immediate reward and termination flag; The target Q value is calculated for the new state, immediate reward and termination flag, and the TD error is minimized. The calculation is repeated until the termination condition is reached, and the optimal distribution network power restoration strategy is obtained. Execute the optimal distribution network power restoration strategy and realize distribution network fault power restoration based on graph attention mechanism.

2. A method for power supply restoration in a distribution network fault based on a graph attention mechanism according to claim 1, characterized in that: The method of dividing the distribution network area by Thiessen polygons according to the fault point to obtain the initial fault division area specifically includes: Set the fault point, line end, and tie switch node as the vertices of the Thiessen polygon; Input the vertices into the Delaunay algorithm to generate a triangulated network; Connect the centers of the circumscribed circles of the triangles in the generated triangulated network to form the Thiessen polygon boundary and obtain the Thiessen polygon sub-region; The Thiessen polygon sub-regions at both ends of the tie switch are merged to obtain the initial fault partition region.

3. The method for power supply restoration after a distribution network fault based on a graph attention mechanism according to claim 1 is characterized in that: The island division based on the initial fault division area specifically includes: The loads within each initial fault partition area form an initial island set; Based on the initial island set, an island partitioning model is constructed with the goal of maximizing power restoration. Solve the island partition model to obtain the island partition result.

4. A method for power supply restoration after a distribution network fault based on a graph attention mechanism according to claim 1, characterized in that: The extraction of the island adjacency matrix and the island load characteristics according to the island division results and the construction of the island graph data specifically include: Extract each isolated node set and edge set; Construct an adjacency matrix based on the node set and edge set; Extract the initial node features in each island; Based on the adjacency matrix and initial node features, the graph structure data of each island is constructed.

5. The method for power supply restoration after a distribution network fault based on a graph attention mechanism according to claim 1 is characterized in that: The extracting of the global state vector of each island according to the island graph data specifically includes: Calculate attention weights based on island graph data; According to the normalized attention weight of neighbor node j on node i, combined with the linear transformation of the neighbor features, a nonlinear activation function is input, and the aggregated features of all attention heads are repeatedly calculated to obtain the aggregated result of the neighbor features of the i-th node; Based on the aggregation results of the neighbor features of the i-th node, the aggregation results of the neighbor features of all nodes in each island are globally pooled to obtain the global state vector of each island.

6. A method for power supply restoration after a distribution network fault based on a graph attention mechanism according to claim 5, characterized in that: The calculation of attention weights based on the island graph data specifically includes: When there is a power line between node i and node j in the adjacency matrix, the corresponding node is determined to be a neighbor node of node i. The features of node i and node j in the initial node features are mapped to a high-dimensional space through a linear transformation matrix, and then the two mapping results are spliced; Transpose the attention parameter vector and do a dot product with the concatenated result to get a scalar value; Input the scalar value into the activation function to get the raw attention score of neighbor node j to node i; According to the original attention score of neighbor node j to node i, the normalized attention weight of neighbor node j to node i is solved.

7. The method for power supply restoration after a distribution network fault based on a graph attention mechanism according to claim 6 is characterized in that: The step of calculating the normalized attention weight of neighbor node j to node i based on the original attention score of neighbor node j to node i specifically includes: The attention scores of all neighboring nodes directly connected to node i through power lines are summed by exponentially mapping the scores to the positive domain through the exponential function, and the summed result is set as the denominator; The exponential result of the original attention score of neighbor node j to node i is set as the numerator to obtain the normalized attention weight of neighbor node j to node i.

8. The method for power supply restoration after a distribution network fault based on a graph attention mechanism according to claim 1 is characterized in that: The global state vector of each island is subjected to a time sequence decision by a deep recurrent Q network based on a bidirectional long-term short-term neural network to obtain the action instructions of each island, specifically including: A bidirectional long-term and short-term neural network is independently constructed for each island, and the global state vector of each island is input into the corresponding bidirectional long-term and short-term neural network to obtain the time series state vector of each island; Calculate the Q value based on the time series state vector of each island to obtain the expected cumulative reward value of each action; According to the expected cumulative reward value of each action, the action is selected through the ε-greedy strategy to obtain the action instructions of each island.

9. A distribution network fault power supply restoration system based on a graph attention mechanism, running a distribution network fault power supply restoration method based on a graph attention mechanism according to any one of claims 1-8, characterized in that: The area division module is used to divide the distribution network area into Thiessen polygons according to the fault point to obtain the initial fault division area; The island division module is used to divide the islands according to the initial fault division area; The power restoration strategy solving module is used to solve the optimal distribution network power restoration strategy based on the islanding results by combining the graph attention mechanism with the bidirectional deep recurrent Q network. The module includes: According to the island division results, the island adjacency matrix and island load characteristics are extracted to construct the island graph data; According to the island graph data, the global state vector of each island is extracted; The global state vector of each island is used to make sequential decisions through a deep recurrent Q network based on a bidirectional long-short-term neural network to obtain the action instructions of each island. Execute the action instructions of each island and calculate the reward function to obtain the next state, immediate reward and termination flag; The target Q value is calculated for the new state, immediate reward and termination flag, and the TD error is minimized. The calculation is repeated until the termination condition is reached, and the optimal distribution network power restoration strategy is obtained. The output module is used to execute the optimal distribution network power supply restoration strategy and realize distribution network fault power supply restoration based on the graph attention mechanism.

Images