A line differentiation lightning protection management method based on deep reinforcement learning

Abstract

The application discloses a line differentiation lightning protection management method based on deep reinforcement learning, comprising the following steps: step one: acquiring line topology, section parameters, lightning trip-out data and key load identification, and generating a state sequence; step two: obtaining exposure clusters based on lightning location information and meteorological information clustering, generating hyper-edges and calculating hyper-edge capacity parameters; step three: constructing a hypergraph and converting to obtain a converted graph with assigned capacity; step four: constructing a Gomory-Hu cut tree to obtain bisection partitioning and tree edge capacity; step five: generating macro actions and sorting; step six: determining target macro action identification and generating action parameters through a strategy network and a value network; step seven: issuing instructions to update data and calculating a reward value; and step eight: updating network parameters according to the reward value and outputting management target sections, action type identification and action parameters. The application improves the differentiation decision and iterative optimization capability of lightning protection management.

Description

Technical Field

[0001] This invention relates to the field of power system lightning protection technology, and in particular to a line-differentiated lightning protection management method based on deep reinforcement learning. Background Technology

[0002] Lightning protection management of transmission lines is an important part of power system operation and maintenance. Existing solutions are usually formulated based on line insulation coordination, grounding resistance, lightning protection device configuration and inspection results. Risk assessment is also conducted by combining lightning location records, meteorological observation information and tripping records. In engineering practice, zoning management and index scoring methods are commonly used. Sections are divided into a predetermined number of management units according to topography, thunderstorm days, ground flash density and historical tripping frequency. Then, under the constraints of budget and power outage window, measures such as lightning arrester installation, grounding resistance management and insulation configuration adjustment are selected to form annual and quarterly management plans.

[0003] To improve the targeting of governance solutions, some existing technologies introduce graph models to represent line topology, treating sections as nodes and adjacent connections as edges, and using edge weights to characterize the degree of weakness and the cost of cutting. Bottleneck locations are obtained through key section identification, minimum cut set analysis, and partitioning, which are used to guide resource allocation and maintenance scheduling. However, since the construction of edge weights largely depends on local information of adjacent sections, the correlation between common causes of lightning strikes and cascading trips caused by thunderstorm corridors and exposure zones is difficult to incorporate into the model. Higher-order correlations lack carriers in topological representation, making partitioning results sensitive to edge weight disturbances. Bottleneck identification and governance boundaries may change abruptly with spatiotemporal exposure, and governance objects and governance boundaries are not easily aligned with risk propagation characteristics.

[0004] At the decision optimization level, existing methods mostly use rule bases, heuristic search, and single-step optimization to generate governance lists. These methods are difficult to form closed-loop updates under conditions of rapid changes in lightning location and meteorological intensity. A few solutions introduce reinforcement learning for operation and maintenance decisions, but they often directly use segment-level actions as the action space. The action dimension expands sharply with the scale of the line, training convergence depends on a large number of interaction samples, and there is a lack of macro-action abstraction driven by topological bottleneck structure. It is difficult to incorporate key load pairs and tree-edge capacity into a unified sorting mechanism, resulting in insufficient interpretability of governance priorities and difficulty in maintaining a stable consistency between the selection of governance target segments and the issuance of action parameters.

[0005] Therefore, how to provide a line-differentiated lightning protection management method based on deep reinforcement learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a line-differentiated lightning protection management method based on deep reinforcement learning. This invention utilizes lightning location information and meteorological information to construct exposure clusters and generate hyperedges. It constructs a transformation graph based on hypergraph-assisted node transformation and generates a Gomory-Hu cut tree. Based on the cut tree, it generates macro actions by binary partitioning and sorting. It combines a policy network and a value network to determine the management target segment and action parameters and updates the network parameters according to the reward value. It has the advantages of risk partitioning that fits the exposure correlation, interpretable management priority, and fast decision-making closed-loop update speed.

[0007] A line-differentiated lightning protection management method based on deep reinforcement learning according to an embodiment of the present invention includes the following steps: Step 1: Obtain line topology, section parameters, lightning trip data, and critical load identifiers; determine section nodes, adjacent connection edges, and critical load pairs; and generate a state sequence. Step 2: Based on lightning location information and meteorological information, the segment nodes are spatiotemporally clustered to obtain exposed clusters, generate hyperedges, and calculate the hyperedge capacity parameters; Step 3: Construct a hypergraph from segment nodes, adjacent connecting edges, and super edges. Use auxiliary nodes to transform the graph and assign capacity based on segment parameters and super edge capacity parameters. Step 4: Construct a Gomory-Hu cut tree in the transformation graph to obtain the tree edges, tree edge capacity, and binary partition; Step 5: Generate macro actions from the binary partitioning. The macro action record includes the partition boundary segment and action type identifier. Generate macro actions in order of critical load pairs and tree edge capacity. Step 6: Construct the policy network and value network, determine the target macro-action identifier and generate action parameters based on the state sequence, tree edge capacity, and macro-action order; Step 7: Determine the target section for treatment based on the target macro action identifier, issue instructions according to the action type identifier and action parameters, update the section parameters, lightning trip data and status sequence, and calculate the return value; Step 8: Update the strategy network parameters and value network parameters based on the reward value, and output the governance target segment, action type identifier, and action parameters.

[0008] Optionally, step one specifically includes: The segment nodes and adjacent connecting edges are determined based on the line topology. The line topology includes segment identifiers and segment connection relationships. The section parameters associated with the section identifier are written into the section node. The section parameters include grounding resistance parameters, insulation configuration parameters, lightning withstand level parameters, and lightning protection device configuration parameters. Lightning trip data is associated with the section identifier and written to the section node. The lightning trip data includes the lightning timestamp, lightning location information, trip timestamp, and trip section identifier. Critical load nodes are identified and critical load pairs are generated based on critical load identifiers. The critical load identifier includes a critical load node identifier and a load weight value. A status sequence is generated based on section parameters, lightning trip data, and critical load identifiers. The status sequence records the section node status quantities and critical load identifiers by timestamp.

[0009] Optionally, step two specifically includes: Lightning timestamps and lightning location coordinates are extracted based on lightning location information, and meteorological timestamps and meteorological intensity values ​​are extracted based on meteorological information. Time alignment is performed between lightning timestamps and meteorological timestamps. Based on the section parameters, the location coordinates of the section nodes are determined, the spatial distance between the location coordinates of the section nodes and the lightning location coordinates is calculated, the lightning location records that meet the distance conditions are associated with the section nodes, and the lightning location records are aggregated according to the time window to obtain the section lightning exposure. The meteorological intensity value is aggregated according to the time window to obtain the meteorological exposure of the section, and the lightning exposure of the section and the meteorological exposure of the section are combined to form the section exposure feature vector; Spatiotemporal clustering is performed based on the segment exposure feature vector, segment node location coordinates, and lightning timestamp to obtain exposure clusters. The exposure cluster records the cluster identifier and the segment identifier of the cluster member. Hyperedges are generated based on exposed clusters. The hyperedges connect the cluster member segment identifiers to the corresponding segment nodes. The hyperedge capacity parameter is calculated by weighted aggregation based on the exposed feature vectors of the cluster member segments.

[0010] Optionally, step three specifically includes: A basic graph is constructed based on segment nodes and adjacent connecting edges. The basic graph has segment nodes as vertices and adjacent connecting edges as edges, and records the segment identifiers of the starting point and ending point of the edges. The hyperedge is incorporated into the base graph to generate a hypergraph, and the hypergraph records the hyperedge identifier and the cluster member segment identifier. For each superedge, generate an auxiliary node. The identifier of the auxiliary node is taken from the identifier of the superedge. Establish auxiliary connection edges between the auxiliary node and the corresponding segment node of the cluster member segment identifier. Generate a transformation graph. The transformation graph includes segment nodes, auxiliary nodes, adjacent connection edges, and auxiliary connection edges. The capacity of adjacent connecting edges is calculated based on the segment parameters. The capacity of adjacent connecting edges is obtained by combining the grounding resistance parameters, insulation configuration parameters, lightning withstand level parameters, and lightning protection device configuration parameters corresponding to the segment identifier at the edge start point with the grounding resistance parameters, insulation configuration parameters, lightning withstand level parameters, and lightning protection device configuration parameters corresponding to the segment identifier at the edge end point, according to weights. The capacity of auxiliary connection edges is calculated based on the superedge capacity parameter. The capacity of auxiliary connection edges is calculated according to the superedge capacity parameter and the number of cluster member segment identifiers and written into the transformation graph.

[0011] Optionally, step four specifically involves: The transformation graph node set is determined based on the transformation graph. The transformation graph node set includes segment nodes and auxiliary nodes. A reference node is selected, and a node sequence table is established based on the transformation graph node set. A parent node mapping table is established based on the node sequence number table. The parent node mapping table records the parent node identifier for each non-base node. The parent node identifier is initialized to the base node identifier. Traverse each non-base node according to the node number table, and perform maximum flow minimum cut calculation on the transformation graph for the current node identifier and the parent node identifier recorded in the parent node mapping table to obtain the cut value and the binary partition. The binary partition records the current node partition node identifier set and the parent node partition node identifier set. Generate a tree edge by combining the current node identifier and the parent node identifier, write the cut value into the tree edge capacity, and associate the binary partition with the tree edge to form a tree edge record. Based on the binary partitioning update of the parent node mapping table, for nodes whose parent node identifier is equal to the current parent node identifier and whose node identifier falls into the set of node identifiers of the current node partition, the parent node identifier is updated to the current node identifier. When the parent node identifier corresponding to the current parent node identifier in the parent node mapping table falls into the current node partition node identifier set, swap the current node identifier and the current parent node identifier, swap the parent node identifier record in the parent node mapping table, and swap the tree edge endpoint identifier, tree edge capacity, and binary partition in the tree edge record. After traversing the node index table, the Gomory-Hu cut tree is obtained by summarizing the tree edge records. The Gomory-Hu cut tree outputs the tree edge set, the tree edge capacity set, and the binary partition set.

[0012] Optionally, step five specifically includes: Generate a macro action set based on the tree edge set, tree edge capacity set, and binary partition set. Assign a macro action identifier to each tree edge, and store the macro action identifier in association with the tree edge endpoint identifier. Based on the transformation graph, the cross-partition edge set is determined. The cross-partition edge set contains adjacent connection edges and auxiliary connection edges whose endpoint node identifiers fall into the current node partition node identifier set and the parent node partition node identifier set, respectively. The segment identifiers corresponding to the segment nodes in the cross-partition edge set are aggregated to generate the partition boundary segment and written into the macro action. Based on the reading of section parameters from the partition boundary section, the action type identifier is determined and written into the macro action by comparing the grounding resistance parameter, insulation configuration parameter, lightning withstand level parameter, lightning protection device configuration parameter and action type threshold parameter. The partition contribution value is calculated based on critical load pairs and binary partitioning. The partition contribution value is the sum of the load weight values ​​of the source critical load node identifiers falling into the current node partition node identifier set and the destination critical load node identifiers falling into the parent node partition node identifier set in the critical load pair, and the sum of the load weight values ​​of the source critical load node identifiers falling into the parent node partition node identifier set and the destination critical load node identifiers falling into the current node partition node identifier set in the critical load pair. Macro actions are sorted based on partition contribution value and tree edge capacity. Macro actions are sorted in descending order of partition contribution value, and if partition contribution values ​​are the same, they are sorted in ascending order of tree edge capacity.

[0013] Optionally, step six specifically includes: Construct a policy network and a value network. The policy network includes a state encoding module, an action evaluation module, and a parameter generation module. The value network includes a state encoding module and a value estimation module. State records are extracted based on the state sequence and timestamp, and the state records are sent to the state encoding module of the policy network to generate state representation vectors. The macro action identifier sequence is determined based on the macro action sorting, and the macro action score of each macro action identifier in the macro action identifier sequence is calculated in the action evaluation module of the policy network in combination with the tree edge capacity set. The macro action identifier corresponding to the highest macro action score is determined as the target macro action identifier based on the macro action score. Action parameters are generated in the parameter generation module of the policy network based on the state representation vector and the target macro-action identifier. The action parameters are stored in association with the target macro-action identifier. The state record is sent to the state encoding module of the value network to generate a value representation vector. Based on the value representation vector, the value estimation module calculates the value estimate, and the value estimate is stored in association with the timestamp.

[0014] Optionally, step seven specifically includes: Macro actions are located based on target macro action identifiers. The partition boundary segments and action type identifiers of macro action records are read. The governance priority value is calculated for the segment identifiers in the partition boundary segments according to the action parameters. The governance target segments are selected according to the governance priority value. The governance priority value is synthesized by weighting the grounding resistance parameters, insulation configuration parameters, lightning withstand level parameters, lightning protection device configuration parameters, and trip count values ​​in the lightning trip data. Governance instructions are generated and issued based on the target area, action type identifier, and action parameters. The governance instructions record the area identifier, action type identifier, action parameters, and timestamp. Update segment parameters based on governance instructions, select segment parameter items according to action type identifier, and assign and update segment parameter items according to action parameters; Collect lightning trip data within the time window corresponding to the timestamp, associate and write the lightning trip data according to the section identifier, and generate status records based on section parameters, lightning trip data, and key load identifiers and write them into the status sequence. The return value is calculated based on the lightning trip data. The return value is obtained by combining the trip count value of the corresponding time window of the governance target section with the load weight value according to the weight and storing them together.

[0015] Optionally, step eight specifically includes: Read the return value based on the timestamp, read the value estimate corresponding to the timestamp in the value estimation sequence, and calculate the advantage value based on the return value and the value estimate; Based on the state records corresponding to timestamps in the state sequence, the order of macro actions, and the set of tree edge capacity, the selection probability value of the target macro action identifier is calculated on the policy network. The gradient of the policy network is determined and the policy network parameters are updated based on the advantage value and the selection probability value. Based on the state records corresponding to timestamps in the state sequence, the value estimate is calculated on the value network, the value error is calculated based on the return value and the value estimate, and the value network gradient is determined and the value network parameters are updated based on the value error. Write the policy network parameters and value network parameters, and output the governance target segment, action type identifier, and action parameters.

[0016] The beneficial effects of this invention are: This invention introduces exposure clusters driven by lightning location information and meteorological information into the line topology representation. Hyperedges are generated through exposure clusters and their capacity parameters are calculated. Auxiliary node transformations are performed on the hypergraph to obtain a transformation graph. Then, a Gomory-Hu cut tree is constructed on the transformation graph. The tree edge capacity and binary partitioning output by the cut tree are used to characterize the bottleneck boundaries of risk coupling between segments. Thus, the determination of the partition boundary segment no longer depends solely on the local weights of adjacent connecting edges. It can explicitly reflect the exposure correlation and inter-segment chain risks at the topology level, improving the accuracy and stability of the location of the target segment for differentiated governance.

[0017] This invention generates macro-actions based on binary partitioning and combines critical load pairs with tree-edge capacity to form a macro-action ranking. Under the framework of policy network and value network, it determines the target macro-action identifier and generates action parameters. Then, it calculates the governance priority value for the partition boundary segment and selects the governance target segment to issue instructions. At the same time, it calculates the reward value based on lightning trip data and updates the network parameters, realizing closed-loop governance optimization in response to changes in lightning risk. By abstracting macro-actions, it reduces the complexity of direct decision-making at the segment level, makes the governance priority ranking and action parameter generation have consistent structural constraints, improves the decision convergence speed in large-scale line scenarios, and enhances the interpretability and executability of the governance scheme. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a line-differentiated lightning protection management method based on deep reinforcement learning proposed in this invention; Figure 2 This is a schematic diagram of the hypergraph construction for a line-differentiated lightning protection management method based on deep reinforcement learning proposed in this invention. Figure 3 This is a schematic diagram of the Gomory-Hu tree construction for a line-differentiated lightning protection management method based on deep reinforcement learning proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figures 1-3 A line-differentiated lightning protection management method based on deep reinforcement learning includes the following steps: Step 1: Obtain line topology, section parameters, lightning trip data, and critical load identifiers; determine section nodes, adjacent connection edges, and critical load pairs; and generate a state sequence. Step 2: Based on lightning location information and meteorological information, the segment nodes are spatiotemporally clustered to obtain exposed clusters, generate hyperedges, and calculate the hyperedge capacity parameters; Step 3: Construct a hypergraph from segment nodes, adjacent connecting edges, and super edges. Use auxiliary nodes to transform the graph and assign capacity based on segment parameters and super edge capacity parameters. Step 4: Construct a Gomory-Hu cut tree in the transformation graph to obtain the tree edges, tree edge capacity, and binary partition; Step 5: Generate macro actions from the binary partitioning. The macro action record includes the partition boundary segment and action type identifier. Generate macro actions in order of critical load pairs and tree edge capacity. Step 6: Construct the policy network and value network, determine the target macro-action identifier and generate action parameters based on the state sequence, tree edge capacity, and macro-action order; Step 7: Determine the target section for treatment based on the target macro action identifier, issue instructions according to the action type identifier and action parameters, update the section parameters, lightning trip data and status sequence, and calculate the return value; Step 8: Update the strategy network parameters and value network parameters based on the reward value, and output the governance target segment, action type identifier, and action parameters.

[0021] In this embodiment, step one specifically includes: Based on the analysis of the segment identifier and segment connection relationship in the line topology, a segment node index table is established according to the segment identifier, and a set of adjacent connection edges is generated according to the segment connection relationship. The adjacent connection edges record the segment identifier of the edge start point and the segment identifier of the edge end point. The segment parameters are mapped to the segment node index table according to the segment identifier. A segment parameter vector is generated for each segment node in the segment node index table. The components of the segment parameter vector correspond to the grounding resistance parameter, insulation configuration parameter, lightning withstand level parameter, and lightning protection device configuration parameter in sequence. The lightning trip data is mapped to the section node index table according to the section identifier. The lightning timestamp and trip timestamp are sorted by time. The number of lightning strikes and trips are accumulated for each section node according to the time window parameter. The trip interval parameter is calculated based on the most recent trip timestamp to form a lightning trip feature vector and write it into the section node index table. Based on the critical load identifier, the critical load nodes are located and a set of critical load nodes is generated. The load weight values ​​are written to the corresponding critical load nodes. The critical load node sets are paired up according to the node identifiers to generate a set of critical load pairs. The critical load pairs record the source critical load node identifier and the destination critical load node identifier. The set of critical load pairs is used for the weight calculation of subsequent macro action sorting. A state sequence is generated based on the section parameter vector, lightning trip feature vector, and critical load node identifier. The state sequence is organized by timestamp, with each timestamp corresponding to a state record. The state record contains the section parameter vector, lightning trip feature vector, critical load node marker, and critical load pair identifier.

[0022] In this embodiment, step two specifically involves: Based on lightning location information, lightning timestamps and lightning location coordinates are read. Based on meteorological information, meteorological timestamps and meteorological intensity values ​​are read. The lightning timestamps and meteorological timestamps are mapped to a unified time axis. The unified time axis uses a fixed sampling interval to generate a time grid. A correlation record is established between lightning location records and meteorological intensity values ​​that fall into the same time grid. Based on the segment parameters, the location coordinates of the segment nodes are read. For each lightning location record, the Euclidean distance between the location coordinates of the segment nodes and the lightning location coordinates is calculated. Lightning location records with Euclidean distances less than the distance threshold are written into the lightning location record list of the corresponding segment nodes. The lightning location record list is counted according to the time grid to generate the lightning count value of the time window. The maximum value of the current amplitude of the lightning location records within the time window is taken to obtain the maximum value of the current amplitude of the time window. The summation is used to obtain the lightning exposure of the segment. The meteorological intensity values ​​are aggregated by a sliding time window according to the time grid. For each segment node, the maximum value of the meteorological intensity value is taken within the time window to generate the segment meteorological exposure. The segment lightning exposure and the segment meteorological exposure are concatenated according to the preset component order to generate the segment exposure feature vector. The feature distance between segment nodes is calculated based on the segment exposure feature vector, the spatial distance between segment nodes is calculated based on the position coordinates of segment nodes, and the time difference of segment node exposure is calculated based on the time grid of a unified time axis. The feature distance, spatial distance, and time difference are compared with thresholds respectively. Segment nodes that meet the threshold constraints form a neighborhood set. Clustering is performed on the neighborhood set to obtain an exposure cluster. The exposure cluster records the member segment identifier set with a cluster identifier. Hyperedges are generated based on the member segment identifier set of the exposed cluster. The connection relationship of the hyperedge is defined by the member segment identifier set. For each hyperedge, the segment exposure feature vector of the member segment is read, and the weighted aggregation value is calculated according to the components of the member segment exposure feature vector as the hyperedge capacity parameter and stored in association with the hyperedge.

[0023] In this embodiment, step three specifically includes: A basic graph is constructed based on segment nodes and adjacent connecting edges. The basic graph uses segment nodes as the vertex set, and the vertices are indexed by segment identifiers. The basic graph uses adjacent connecting edges as the edge set, which is generated by the segment connection relationship of the line topology. The segment identifiers of the edge start point and the segment identifiers of the edge end point are written into the edge record according to the segment connection relationship. The hyperedge is incorporated into the base graph to generate the hypergraph. The hyperedge is generated by the exposed cluster, the hyperedge identifier is generated by the exposed cluster identifier, and the cluster member segment identifier is written into the hyperedge record by the exposed cluster member segment identifier. The hypergraph is composed of segment nodes, adjacent connecting edges, and hyperedges, and a hyperedge identifier index is established. For each superedge, an auxiliary node is generated. The auxiliary node identifier is taken from the superedge identifier and written into the auxiliary node set. For the cluster member segment identifier recorded by the superedge, the corresponding segment node is searched one by one. An auxiliary connection edge is established from the auxiliary node to the segment node. The auxiliary connection edge records the auxiliary node identifier and the segment identifier. The transformation graph is composed of the segment node set, the auxiliary node set, the adjacent connection edge set, and the auxiliary connection edge set. The capacity of adjacent connecting edges is calculated based on the segment parameters. The grounding resistance parameters, insulation configuration parameters, lightning withstand level parameters, and lightning protection device configuration parameters of the starting segment are identified from the starting segment of the edge, and the grounding resistance parameters, insulation configuration parameters, lightning withstand level parameters, and lightning protection device configuration parameters of the ending segment are identified from the ending segment of the edge. The parameters of the starting segment and the ending segment are normalized. The normalized parameters are weighted and summed according to the preset weights to obtain the capacity of adjacent connecting edges and written into the adjacent connecting edge record. Calculate the capacity of the auxiliary connection edge based on the superedge capacity parameter, read the superedge capacity parameter and the number of cluster member segment identifiers, calculate the quotient of the superedge capacity parameter and the number of cluster member segment identifiers to obtain the single-edge allocated capacity, and write the single-edge allocated capacity into the auxiliary connection edge record corresponding to the superedge to complete the assignment of the transformed graph edge capacity.

[0024] In this embodiment, step four specifically includes: The transformation graph node set is read based on the transformation graph. The transformation graph node set is obtained by merging the segment node set and the auxiliary node set. The reference node identifier is selected and written into the reference node variable. The node identifiers are sorted based on the transformation graph node set to generate a node sequence number table. The node sequence number table records the correspondence between the sequence number and the node identifier. A parent node mapping table is established based on the node sequence number table. The parent node mapping table records the parent node identifier with the node identifier as the key, and assigns the base node identifier as the initial parent node identifier to each non-base node identifier. Traverse the non-baseline node identifiers in ascending order of the node index table, read the parent node identifier from the corresponding parent node mapping table, and perform maximum flow minimum cut calculation on the transformation graph with the current node identifier as the source and the parent node identifier as the sink. The maximum flow minimum cut calculation includes initializing edge flow to zero, calculating residual capacity based on edge capacity and edge flow, constructing a residual network based on the residual capacity, subtracting edge flow from edge capacity to form the forward residual capacity, and taking edge flow from residual capacity to form the reverse residual capacity. The residual network contains nodes with residual capacity greater than zero. For each edge, search for an augmenting path from the source to the sink along the residual network and update the edge flow according to the minimum residual capacity of the augmenting path. Repeat the augmenting update until there is no augmenting path from the source to the sink in the residual network. Based on the set of node identifiers reachable from the source in the residual network, determine the current node partition node identifier set. Remove the current node partition node identifier set from the transformed graph node set to obtain the parent node partition node identifier set. The cut value is obtained by summing the directed edge capacities from the current node partition node identifier set to the parent node partition node identifier set and using it as the tree edge capacity. Combine the current node identifier and the parent node identifier to form a tree edge endpoint identifier pair. Take the cut value for the tree edge capacity. Take the current node partition node identifier set and the parent node partition node identifier set for the binary partition. Write the tree edge endpoint identifier pair, tree edge capacity, and binary partition as associations into the tree edge record set. Based on the binary partitioning update of the parent node mapping table, traverse the node identifiers in the parent node mapping table where the parent node identifier is equal to the current parent node identifier. For node identifiers that satisfy the condition that the node identifier falls into the set of node identifiers of the current node partition, update the parent node identifier from the current parent node identifier to the current node identifier. Determine whether the parent node identifier corresponding to the current parent node identifier in the parent node mapping table falls into the current node partition node identifier set. If the condition is met, swap the current node identifier and the current parent node identifier. In the parent node mapping table, swap the parent node identifier corresponding to the current node identifier with the parent node identifier corresponding to the current parent node identifier. Locate the tree edge record containing the current node identifier in the tree edge record set and replace the current node identifier in the tree edge endpoint identifier pair with the current parent node identifier. At the same time, the tree edge capacity and the binary partition are swapped along with the tree edge endpoint identifier pair. After traversing the node index table, the tree edge endpoint identifier pairs in the tree edge record set are used as the tree edge set, the tree edge capacity in the tree edge record set is used as the tree edge capacity set, and the binary partition in the tree edge record set is used as the binary partition set. Based on the tree edge set, construct the Gomory-Hu cut tree and output the tree edge set, tree edge capacity set, and binary partition set.

[0025] In this embodiment, step five specifically includes: Generate a macro action set based on the tree edge set, tree edge capacity set, and binary partition set. Traverse each tree edge endpoint identifier pair in the tree edge set and generate a macro action identifier according to the traversal sequence number. The macro action records the tree edge endpoint identifier pair, tree edge capacity, and binary partition identifier with the macro action identifier as the primary key. Based on the transformation graph, determine the cross-partition edge set. For each tree edge, read the current node partition node identifier set and the parent node partition node identifier set for the corresponding binary partition. Traverse the adjacent connection edge set and auxiliary connection edge set of the transformation graph. Filter the edge records where one end of the endpoint node identifier falls into the current node partition node identifier set and the other end falls into the parent node partition node identifier set and write them into the cross-partition edge set. Read the node type corresponding to the endpoint node identifier in the cross-partition edge set. If the node type is a segment node, read the segment identifier. Aggregate the segment identifiers and remove duplicates to obtain the partition boundary segment. Write the partition boundary segment into the macro action. Based on reading the segment parameters of the partition boundary segment, for each segment identifier in the partition boundary segment, the grounding resistance parameter, insulation configuration parameter, lightning withstand level parameter, and lightning protection device configuration parameter are read, and the action type threshold parameter is compared with the action type threshold parameter to generate the segment action type identifier. The action type identifier of the macro action is determined according to the frequency statistics of the segment action type identifier and written into the macro action. The partition contribution value is calculated based on critical load pairs and binary partitions. The source critical load node identifier, destination critical load node identifier, and load weight value of each critical load pair in the critical load pair set are read. The relationship between the source critical load node identifier and the destination critical load node identifier falling into the node identifier set of the binary partition is determined. When the source critical load node identifier falls into the node identifier set of the current node partition and the destination critical load node identifier falls into the node identifier set of the parent node partition, the load weight value is accumulated. When the source critical load node identifier falls into the node identifier set of the parent node partition and the destination critical load node identifier falls into the node identifier set of the current node partition, the load weight value is accumulated. The accumulated results are summed to obtain the partition contribution value and stored in association with the macro action identifier. Macro-action sorting is generated based on partition contribution value and tree edge capacity. Macro-action sorting takes macro-action identifier as the sorting object. The partition contribution value and tree edge capacity corresponding to the macro-action identifier are read. The macro-action identifiers are sorted in descending order by partition contribution value. If the partition contribution values ​​are the same, the macro-action identifiers are sorted in ascending order by tree edge capacity. The sorting result is written into the macro-action sorting sequence and stored in association with the macro-action set.

[0026] In this embodiment, step six specifically includes: Construct a policy network and a value network. The policy network includes a state encoding module, an action evaluation module, and a parameter generation module. The value network includes a state encoding module and a value estimation module. The state encoding module receives state records and outputs a state representation vector. The action evaluation module receives the state representation vector, the macro-action identifier sequence, and the tree edge capacity set and outputs the macro-action score. The parameter generation module receives the state representation vector and the target macro-action identifier and outputs the action parameters. The value estimation module receives the value representation vector and outputs the value estimate. State records are extracted based on the state sequence and timestamp. The state record consists of the state quantity of the segment node and the identifier of the key load. The state records are sent to the state encoding module of the policy network in the order of timestamp. The state encoding module performs vectorization encoding on the state record to obtain the state representation vector. The state representation vector is stored in association with the timestamp. The macro action identifier sequence is determined based on the macro action sorting. The macro action identifier sequence is arranged in the order of macro action sorting. The macro action score of each macro action identifier in the action evaluation module of the policy network is calculated in combination with the tree edge capacity set. The macro action score is calculated by fusing the state representation vector with the embedding vector of the corresponding macro action identifier's tree edge capacity, partition boundary segment, and action type identifier. The macro action score is stored in association with the macro action identifier. The macro action identifier corresponding to the maximum macro action score is determined based on the macro action score and is used as the target macro action identifier. The target macro action identifier is selected by performing the maximum value on the macro action score and is written into the target macro action identifier variable. Action parameters are generated in the parameter generation module of the policy network based on the state representation vector and the target macro action identifier. The parameter generation module calculates the action parameters by fusing the embedding vectors of the state representation vector and the target macro action identifier. The action parameters are associated with the target macro action identifier and written into the action parameter vector storage table. The state record is sent to the state encoding module of the value network to generate a value representation vector. The value representation vector is obtained by encoding the state record by the state encoding module of the value network. Based on the value representation vector, the value estimation module calculates the value estimate. The value estimate is obtained by performing regression calculation on the value representation vector by the value estimation module. The value estimate is associated with the timestamp and written into the value estimation sequence.

[0027] In this embodiment, step seven specifically includes: Macro actions are located based on target macro action identifiers. Macro action records are retrieved from the macro action set according to the target macro action identifier. The partition boundary segment set and action type identifier in the macro action record are read. The segment identifiers in the partition boundary segment set are traversed. The grounding resistance parameters, insulation configuration parameters, lightning withstand level parameters, and lightning protection device configuration parameters corresponding to the segment identifiers are read. The trip count value corresponding to the segment identifiers in the lightning trip data is read. Normalization processing is performed on the grounding resistance parameters, insulation configuration parameters, lightning withstand level parameters, lightning protection device configuration parameters, and trip count values. The normalized values ​​are weighted and summed according to preset weights to obtain the governance priority value. The governance priority value is associated with the segment identifiers and stored. The segment identifiers are sorted in descending order according to the governance priority value, and the preset number of segment identifiers before the sorting is taken to form the governance target segment. Governance instructions are generated and issued based on the governance target segment, action type identifier, and action parameters. A governance instruction record is generated for each segment identifier in the governance target segment. The governance instruction record is written with the segment identifier, action type identifier, action parameters, and timestamp. The governance instructions are issued by writing them into the execution queue using the sending interface call method. The section parameters are updated based on the governance instructions. The governance instructions record in the execution queue is traversed. The corresponding section parameter item name in the section parameter item mapping table is determined according to the action type identifier. The section parameter item name is taken from one of the grounding resistance parameter, insulation configuration parameter, lightning withstand level parameter, and lightning protection device configuration parameter. The section parameter item is updated to the corresponding assignment result of the action parameter according to the action parameter. The updated section parameters are associated with the section identifier and stored. Collect lightning trip data within the time window corresponding to the timestamp, determine the start and end times of the time window by the timestamp and the time window length, write the lightning trip data within the time window into the lightning trip data record corresponding to the section identifier according to the section identifier, generate a status record based on the updated section parameters, lightning trip data, and critical load identifier, associate the status record with the section node status quantity and critical load identifier according to the timestamp, and write the status record into the status sequence; The return value is calculated based on the lightning trip data. For each segment identifier of the governance target segment, the trip count value of the time window is read, the load weight value in the key load identifier is read and mapped to the governance target segment according to the segment identifier, and the trip count value of the time window and the load weight value are weighted and synthesized according to the preset weight to obtain the return value. The return value is stored in association with the timestamp.

[0028] In this embodiment, step eight specifically includes: The return value is read based on the timestamp. The return value and the timestamp are associated with each other in the return value storage table. The value estimate corresponding to the timestamp in the value estimation sequence is read. The value estimate is associated with each timestamp in the value estimation sequence. The advantage value is calculated based on the return value and the value estimate. The advantage value is the return value minus the value estimate and is associated with the timestamp and stored. Based on the state records corresponding to timestamps in the state sequence, the macro-action order, and the tree edge capacity set, the selection probability value of the target macro-action identifier is calculated on the policy network. The state records and timestamps are associated with each other in the state sequence. The macro-action identifier sequence is recorded in the macro-action order. The tree edge capacity set is associated with the macro-action identifier and the macro-action set. The policy network generates a state representation vector in the state encoding module. The policy network calculates the macro-action score for the macro-action identifier sequence in the action evaluation module. The policy network normalizes the macro-action score and converts it into the selection probability distribution of the macro-action identifier sequence. The selection probability value corresponding to the target macro-action identifier is read. The policy network gradient is determined based on the dominance value and the selection probability value. The policy network gradient is calculated by multiplying the logarithmic gradient of the selection probability value by the dominance value and updating the policy network parameters. Based on the state records corresponding to the timestamps in the state sequence, the value estimate is calculated on the value network. The value network generates a value representation vector in the state encoding module. The value network performs regression calculation on the value representation vector in the value estimation module to obtain the value estimate. The value error is calculated based on the return value and the value estimate. The value error is the return value minus the value estimate. The value network gradient is determined based on the value error. The value network gradient is obtained by differentiating the parameters with respect to the squared loss of the value error and then updating the value network parameters. Write the policy network parameters and value network parameters. The policy network parameters are written to the policy network parameter storage area, and the value network parameters are written to the value network parameter storage area. The governance target segment, action type identifier, and action parameters are taken from the records in the governance instruction generation stage and are used as the method output results.

[0029] Example 1: To verify the feasibility of this invention in practice, it was applied to a differentiated lightning protection management scenario for an overhead transmission line in a certain region. The line consists of several sections, traversing hilly, plain, and water-rich areas. During the thunderstorm season, lightning activity is frequent, and lightning tripping exhibits clustered occurrences. Some sections experience repeated tripping within similar time windows. When maintenance personnel formulate management plans using experience-based zoning and fixed threshold lists, problems often arise such as mismatch between management boundaries and thunderstorm exposure zones, fragmented management priorities for sections within the same exposure zone, and the inability to quickly reflect the effects of management instructions in the next round of decision-making. This leads to dispersed management resource allocation, delayed plan adjustments, and insignificant reduction in the risk of power supply to critical loads.

[0030] In the application scenario, the following steps are first taken: First, the line topology, section parameters, lightning trip data, and critical load identifiers are collected to determine section nodes, adjacent connecting edges, and critical load pairs, generating a state sequence. The state sequence is organized by timestamp, showing the section node state variables and critical load pair identifiers. Then, spatiotemporal clustering of the section nodes is performed using lightning location information and meteorological information to obtain exposed clusters. These exposed clusters are aggregated using the segment identifiers of cluster members to form hyperedges, and the hyperedge capacity parameters are obtained by weighted aggregation based on the exposed feature vectors of the cluster members. A hypergraph is constructed based on the section nodes, adjacent connecting edges, and hyperedges. Auxiliary nodes are generated for the hyperedges, and auxiliary connecting edges are established to obtain a transformation graph. The edge capacity of the transformation graph is composed of the adjacent connecting edge capacity synthesized from the section parameters and the auxiliary connecting edge capacity allocated from the hyperedge capacity parameters. A Gomory-Hu cut tree is constructed on the transformation graph, outputting tree edges, tree edge capacities, and binary partitions. The binary partitions are used to generate macro-actions, recording the partition boundary segments and action type identifiers, and generating a macro-action sorting based on critical load pairs and tree edge capacities. Under the constraints of state sequence, tree edge capacity, and macro-action ordering, the policy network and value network determine the target macro-action identifier and generate action parameters. In the governance phase, the boundary segment of the partition is located based on the target macro-action identifier, the governance priority value is calculated according to the action parameters, the target segment of governance is selected, the governance instruction is issued to update the segment parameters, and lightning trip data within the time window is collected. The reward value is calculated based on the lightning trip data. The reward value is used to update the parameters of the policy network and the value network, so that the selection of governance actions and action parameters converge in the direction of reducing trip risk in continuous iteration.

[0031] To verify the beneficial effects, within the same line and the same thunderstorm season window, comparison scheme A was set up using experience partitioning + fixed threshold list, comparison scheme B was set up using only adjacent connection edges to construct the graph and perform minimum cut partitioning before deployment according to rules, and the scheme of this invention was set up using exposed cluster superedges + transformation graph + Gomory-Hu tree cutting + macro action sorting + closed-loop update of policy network and value network. Evaluation indicators include the number of section-level trips, the duration of power outages associated with critical load pairs, the calculation time of a single round of decision-making, the concentration of coverage of governance objects, and the number of policy convergence rounds. During the sampling period, approximately 26,000 lightning location events, approximately 8,000 effective meteorological time grids, and 286 lightning trip records were recorded. The critical load node identification covered 12 important load points, forming 66 pairs of critical loads. The results are shown in Table 1: Table 1 Comparison of Differentiated Lightning Protection Treatment Effects

[0032] The data shows that the proposed solution significantly outperforms both comparative solutions A and B in two core indicators: the number of trips and the duration of outages associated with critical load pairs. Furthermore, the solution demonstrates improved coverage concentration, indicating that macro-action sequencing, under the constraints of critical load pairs and tree-edge capacity, more effectively focuses governance resources on high-risk bottleneck locations within partition boundary segments. The reduced number of partition boundary flips reflects that after the generation of superedges from exposed clusters, risk correlation is expressed at the structural level, and the tree-cutting binary partitioning is more stable as it updates with time windows. While the single-round decision calculation time is slightly higher than the experience list, it remains within an acceptable range for engineering applications. The policy network and value network converge within a limited number of rounds, and the governance target segments and action parameters show a stable trend during iteration. This solves the problems of feedback lag, governance boundaries not conforming to exposed zones, and dispersed resource allocation inherent in traditional solutions.

[0033] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A line-differentiated lightning protection management method based on deep reinforcement learning, characterized in that, Includes the following steps: Step 1: Obtain line topology, section parameters, lightning trip data, and critical load identifiers; determine section nodes, adjacent connection edges, and critical load pairs; and generate a state sequence. Step 2: Based on lightning location information and meteorological information, the segment nodes are spatiotemporally clustered to obtain exposed clusters, generate hyperedges, and calculate the hyperedge capacity parameters; Step 3: Construct a hypergraph from segment nodes, adjacent connecting edges, and super edges. Use auxiliary nodes to transform the graph and assign capacity based on segment parameters and super edge capacity parameters. Step 4: Construct a Gomory-Hu cut tree in the transformation graph to obtain the tree edges, tree edge capacity, and binary partition; Step 5: Generate macro actions from the binary partitioning. The macro action record includes the partition boundary segment and action type identifier. Generate macro actions in order of critical load pairs and tree edge capacity. Step 6: Construct the policy network and value network, determine the target macro-action identifier and generate action parameters based on the state sequence, tree edge capacity, and macro-action order; Step 7: Determine the target section for treatment based on the target macro action identifier, issue instructions according to the action type identifier and action parameters, update the section parameters, lightning trip data and status sequence, and calculate the return value; Step 8: Update the strategy network parameters and value network parameters based on the reward value, and output the governance target segment, action type identifier, and action parameters.

2. The line-differentiated lightning protection management method based on deep reinforcement learning according to claim 1, characterized in that, Step one specifically involves: The segment nodes and adjacent connecting edges are determined based on the line topology. The line topology includes segment identifiers and segment connection relationships. The section parameters associated with the section identifier are written into the section node. The section parameters include grounding resistance parameters, insulation configuration parameters, lightning withstand level parameters, and lightning protection device configuration parameters. Lightning trip data is associated with the section identifier and written to the section node. The lightning trip data includes the lightning timestamp, lightning location information, trip timestamp, and trip section identifier. Critical load nodes are identified and critical load pairs are generated based on critical load identifiers. The critical load identifier includes a critical load node identifier and a load weight value. A status sequence is generated based on section parameters, lightning trip data, and critical load identifiers. The status sequence records the section node status quantities and critical load identifiers by timestamp.

3. The line-differentiated lightning protection management method based on deep reinforcement learning according to claim 1, characterized in that, Step two specifically involves: Lightning timestamps and lightning location coordinates are extracted based on lightning location information, and meteorological timestamps and meteorological intensity values ​​are extracted based on meteorological information. Time alignment is performed between lightning timestamps and meteorological timestamps. Based on the section parameters, the location coordinates of the section nodes are determined, the spatial distance between the location coordinates of the section nodes and the lightning location coordinates is calculated, the lightning location records that meet the distance conditions are associated with the section nodes, and the lightning location records are aggregated according to the time window to obtain the section lightning exposure. The meteorological intensity value is aggregated according to the time window to obtain the meteorological exposure of the section, and the lightning exposure of the section and the meteorological exposure of the section are combined to form the section exposure feature vector; Spatiotemporal clustering is performed based on the segment exposure feature vector, segment node location coordinates, and lightning timestamp to obtain exposure clusters. The exposure cluster records the cluster identifier and the segment identifier of the cluster member. Hyperedges are generated based on exposed clusters. The hyperedges connect the cluster member segment identifiers to the corresponding segment nodes. The hyperedge capacity parameter is calculated by weighted aggregation based on the exposed feature vectors of the cluster member segments.

4. The line-differentiated lightning protection management method based on deep reinforcement learning according to claim 1, characterized in that, Step three specifically involves: A basic graph is constructed based on segment nodes and adjacent connecting edges. The basic graph has segment nodes as vertices and adjacent connecting edges as edges, and records the segment identifiers of the starting point and ending point of the edges. The hyperedge is incorporated into the base graph to generate a hypergraph, and the hypergraph records the hyperedge identifier and the cluster member segment identifier. For each superedge, generate an auxiliary node. The identifier of the auxiliary node is taken from the identifier of the superedge. Establish auxiliary connection edges between the auxiliary node and the corresponding segment node of the cluster member segment identifier. Generate a transformation graph. The transformation graph includes segment nodes, auxiliary nodes, adjacent connection edges, and auxiliary connection edges. The capacity of adjacent connecting edges is calculated based on the segment parameters. The capacity of adjacent connecting edges is obtained by combining the grounding resistance parameters, insulation configuration parameters, lightning withstand level parameters, and lightning protection device configuration parameters corresponding to the segment identifier at the edge start point with the grounding resistance parameters, insulation configuration parameters, lightning withstand level parameters, and lightning protection device configuration parameters corresponding to the segment identifier at the edge end point, according to weights. The capacity of auxiliary connection edges is calculated based on the superedge capacity parameter. The capacity of auxiliary connection edges is calculated according to the superedge capacity parameter and the number of cluster member segment identifiers and written into the transformation graph.

5. The line-differentiated lightning protection management method based on deep reinforcement learning according to claim 1, characterized in that, Step four specifically involves: The transformation graph node set is determined based on the transformation graph. The transformation graph node set includes segment nodes and auxiliary nodes. A reference node is selected, and a node sequence table is established based on the transformation graph node set. A parent node mapping table is established based on the node sequence number table. The parent node mapping table records the parent node identifier for each non-base node. The parent node identifier is initialized to the base node identifier. Traverse each non-base node according to the node number table, and perform maximum flow minimum cut calculation on the transformation graph for the current node identifier and the parent node identifier recorded in the parent node mapping table to obtain the cut value and the binary partition. The binary partition records the current node partition node identifier set and the parent node partition node identifier set. Generate a tree edge by combining the current node identifier and the parent node identifier, write the cut value into the tree edge capacity, and associate the binary partition with the tree edge to form a tree edge record. Based on the binary partitioning update of the parent node mapping table, for nodes whose parent node identifier is equal to the current parent node identifier and whose node identifier falls into the set of node identifiers of the current node partition, the parent node identifier is updated to the current node identifier. When the parent node identifier corresponding to the current parent node identifier in the parent node mapping table falls into the current node partition node identifier set, swap the current node identifier and the current parent node identifier, swap the parent node identifier record in the parent node mapping table, and swap the tree edge endpoint identifier, tree edge capacity, and binary partition in the tree edge record. After traversing the node index table, the Gomory-Hu cut tree is obtained by summarizing the tree edge records. The Gomory-Hu cut tree outputs the tree edge set, the tree edge capacity set, and the binary partition set.

6. The line-differentiated lightning protection management method based on deep reinforcement learning according to claim 1, characterized in that, Step five specifically involves: Generate a macro action set based on the tree edge set, tree edge capacity set, and binary partition set. Assign a macro action identifier to each tree edge, and store the macro action identifier in association with the tree edge endpoint identifier. Based on the transformation graph, the cross-partition edge set is determined. The cross-partition edge set contains adjacent connection edges and auxiliary connection edges whose endpoint node identifiers fall into the current node partition node identifier set and the parent node partition node identifier set, respectively. The segment identifiers corresponding to the segment nodes in the cross-partition edge set are merged to generate the partition boundary segment and written into the macro action. Based on the reading of section parameters from the partition boundary section, the action type identifier is determined and written into the macro action by comparing the grounding resistance parameter, insulation configuration parameter, lightning withstand level parameter, lightning protection device configuration parameter and action type threshold parameter. The partition contribution value is calculated based on critical load pairs and binary partitioning. The partition contribution value is the sum of the load weight values ​​of the source critical load node identifiers falling into the current node partition node identifier set and the destination critical load node identifiers falling into the parent node partition node identifier set in the critical load pair, and the sum of the load weight values ​​of the source critical load node identifiers falling into the parent node partition node identifier set and the destination critical load node identifiers falling into the current node partition node identifier set in the critical load pair. Macro actions are sorted based on partition contribution value and tree edge capacity. Macro actions are sorted in descending order of partition contribution value, and if partition contribution values ​​are the same, they are sorted in ascending order of tree edge capacity.

7. The line-differentiated lightning protection management method based on deep reinforcement learning according to claim 1, characterized in that, Step six specifically involves: Construct a policy network and a value network. The policy network includes a state encoding module, an action evaluation module, and a parameter generation module. The value network includes a state encoding module and a value estimation module. State records are extracted based on the state sequence and timestamp, and the state records are sent to the state encoding module of the policy network to generate state representation vectors. The macro action identifier sequence is determined based on the macro action sorting, and the macro action score of each macro action identifier in the macro action identifier sequence is calculated in the action evaluation module of the policy network in combination with the tree edge capacity set. The macro action identifier corresponding to the highest macro action score is determined as the target macro action identifier based on the macro action score. Action parameters are generated in the parameter generation module of the policy network based on the state representation vector and the target macro-action identifier. The action parameters are stored in association with the target macro-action identifier. The state record is sent to the state encoding module of the value network to generate a value representation vector. Based on the value representation vector, the value estimation module calculates the value estimate, and the value estimate is stored in association with the timestamp.

8. The line-differentiated lightning protection management method based on deep reinforcement learning according to claim 1, characterized in that, Step seven specifically involves: Macro actions are located based on target macro action identifiers. The partition boundary segments and action type identifiers of macro action records are read. The governance priority value is calculated for the segment identifiers in the partition boundary segments according to the action parameters. The governance target segments are selected according to the governance priority value. The governance priority value is synthesized by weighting the grounding resistance parameters, insulation configuration parameters, lightning withstand level parameters, lightning protection device configuration parameters, and trip count values ​​in the lightning trip data. Governance instructions are generated and issued based on the target area, action type identifier, and action parameters. The governance instructions record the area identifier, action type identifier, action parameters, and timestamp. Update segment parameters based on governance instructions, select segment parameter items according to action type identifier, and assign and update segment parameter items according to action parameters; Collect lightning trip data within the time window corresponding to the timestamp, associate and write the lightning trip data according to the section identifier, and generate status records based on section parameters, lightning trip data, and key load identifiers and write them into the status sequence. The return value is calculated based on the lightning trip data. The return value is obtained by combining the trip count value of the corresponding time window of the governance target section with the load weight value according to the weight and storing them together.

9. The line-differentiated lightning protection management method based on deep reinforcement learning according to claim 1, characterized in that, Step eight specifically involves: Read the return value based on the timestamp, read the value estimate corresponding to the timestamp in the value estimation sequence, and calculate the advantage value based on the return value and the value estimate; Based on the state records corresponding to timestamps in the state sequence, the order of macro actions, and the set of tree edge capacity, the selection probability value of the target macro action identifier is calculated on the policy network. The gradient of the policy network is determined and the policy network parameters are updated based on the advantage value and the selection probability value. Based on the state records corresponding to timestamps in the state sequence, the value estimate is calculated on the value network, the value error is calculated based on the return value and the value estimate, and the value network gradient is determined and the value network parameters are updated based on the value error. Write the policy network parameters and value network parameters, and output the governance target segment, action type identifier, and action parameters.

Images