A Deep Learning-Based Intelligent Optimization Method for Power Design Parameters

By using deep learning-based graph neural networks and adaptive large neighborhood search technology, power design state diagrams and state tensors are generated. Parameter optimization is then performed using an improved Graphormer model and the ALNS algorithm, solving the problem of inaccurate parameter adjustment in power design and achieving efficient and feasible design parameter optimization.

CN123069903APending Publication Date: 2026-11-10PINGSHAN COUNTY JINCHENG POWER CONSTR CENT
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611336637.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-31
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing power design methods are unable to uniformly express different design objects and their topological influence relationships, resulting in inaccurate identification of the effect path of parameter adjustment on multi-condition operation response. Traditional deep learning methods are unable to determine the adjustment direction and correlation strength of design parameters in reverse based on engineering constraints, resulting in insufficient optimization efficiency and feasibility of results.

Method used

By employing deep learning-based graph neural networks and adaptive large neighborhood search technology, power design state diagrams and state tensors are generated. An improved Graphormer model is used for parameter correlation analysis and correction, and an improved ALNS algorithm is combined for parameter optimization, thereby achieving targeted correction and engineering verification of design parameters.

Benefits of technology

It improves the accuracy and engineering feasibility of power design parameter optimization, enhances the efficiency of parameter adjustment, and ensures the degree of matching between optimization results and multiple operating conditions and engineering constraints.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123069903A_ABST
    Figure CN123069903A_ABST
Patent Text Reader

Abstract

This invention discloses a deep learning-based intelligent optimization method for power design parameters, belonging to the field of intelligent design technology for power engineering. The method includes the following steps: receiving power design parameter optimization task information and generating a power design parameter optimization constraint set; reading original power design data and generating a power design state diagram based on electrical connection relationships; tensorizing node attributes and edge relationships and combining operating conditions to generate a power design state tensor; using an improved Graphormer model to perform multi-condition response propagation, design constraint residual reverse attribution, and parameter correction decoding to generate a parameter correction evidence set; constructing a candidate domain for power design parameters based on this, and generating a candidate power design parameter group using an improved ALNS algorithm; and generating the power design parameter optimization result through engineering verification, result feedback, and dual-objective screening. This invention can improve the engineering adaptability and feasibility of power design parameter optimization.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of intelligent design technology for power engineering, and in particular to an intelligent optimization method for power design parameters based on deep learning. Background Technology

[0002] With the expansion of power engineering design scale, the number of design objects, operating conditions, and engineering constraints is constantly increasing, making intelligent analysis and optimization of power design parameters a focus of widespread attention. Existing methods mainly rely on designers to perform manual calculations combining power flow calculations, short-circuit current calculations, equipment capacity verification, and protection coordination calculations, or to use heuristic algorithms to search within a preset range. However, these methods generally suffer from the following problems in practical applications: The equipment attributes, electrical connections, and operating conditions in the original power design data are interconnected. Existing methods cannot uniformly express different design objects and their topological influence relationships, resulting in inaccurate identification of the effect path of parameter adjustment on multi-condition operation response. Traditional deep learning methods focus on single parameter prediction or result evaluation, making it difficult to determine the adjustment direction, correction amount, and correlation strength of design parameters based on engineering constraint residuals. The generated candidate values ​​lack a correspondence with the actual constraints.

[0003] Traditional intelligent optimization algorithms adjust parameters using fixed neighborhoods or random perturbations, which cannot dynamically select associated parameters and search ranges based on constraint residuals. This easily generates a large number of candidate solutions that do not meet engineering constraints, and the engineering verification results are difficult to feed back to the search process, resulting in insufficient optimization efficiency and feasibility of results.

[0004] Therefore, how to provide a method for intelligent optimization of power design parameters based on deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an intelligent optimization method for power design parameters based on deep learning. This invention fully utilizes graph neural networks, deep learning, and adaptive large neighborhood search technology to achieve correlation analysis, directional correction, engineering verification, and optimization screening of power design parameters. It has the advantages of high optimization accuracy, strong engineering feasibility, and high parameter adjustment efficiency.

[0006] A method for intelligent optimization of power design parameters based on deep learning according to an embodiment of the present invention includes the following steps: Step 1: Receive power design parameter optimization task information, determine the design scheme to be optimized and engineering constraints, and generate a power design parameter optimization constraint set; Step 2: Read the original power design data according to the power design parameter optimization constraint set, perform correlation coding based on electrical connection relationships, and generate a power design state diagram; Step 3: Tensorize the node attributes and edge relationships in the power design state diagram, perform combinations according to the operating conditions, and generate the power design state tensor; Step 4: Input the power design state tensor into the improved Graphormer model, and generate a parameter correction evidence set through heterogeneous design object encoding, multi-condition response propagation under electrical influence path constraints, reverse attribution of design constraint residuals, and parameter correction decoding. Step 5: Determine candidate values ​​and adjust correlations according to the parameter correction evidence set to generate candidate domains for power design parameters; Step 6: Input the candidate domain of power design parameters into the improved ALNS algorithm, and generate a set of candidate power design parameters through parameter destruction driven by design constraint residuals, candidate domain constraint repair, adaptive selection of destruction-repair operator combination, and hierarchical acceptance with priority given to engineering constraints. Step 7: Perform engineering verification and result feedback on the candidate power design parameter set to generate a feasible power design parameter set; Step 8: Filter the feasible power design parameter set according to the parameter modification range and design performance, and generate power design parameter optimization results.

[0007] Optionally, step one specifically includes: Analyze the power design parameter optimization task information, locate the unique design scheme according to the task direction, and terminate the current process if no unique design scheme is obtained. The design scheme is checked according to the optimization scope defined by the task. The design content within the optimization scope is included in the adjustment scope, while the design content outside the optimization scope is kept in its original state, and a design scheme to be optimized is generated. Match the corresponding engineering constraints to the design scheme to be optimized, delete engineering constraints that are irrelevant to the design scheme to be optimized, and merge duplicate engineering constraints with the same target and restriction content. For engineering constraints that affect the same design content and have conflicting results, compare their applicable scope, retain the engineering constraints that cover the design scheme to be optimized and have a narrower scope, and terminate the current processing if a unique result cannot be determined. The engineering constraints that have been matched, merged, and conflict-handled are associated with the design scheme to be optimized, generating a set of power design parameter optimization constraints.

[0008] Optionally, step two specifically involves: According to the reading range limited by the power design parameter optimization constraint set, the design record corresponding to the design scheme to be optimized is read from the original power design data; Merge design records based on the object correspondence between them. Merge design records that correspond to the same object and have the same content into the same record location. For design records with inconsistent content, determine the merging result based on the corresponding content in the design scheme to be optimized. If the merging result cannot be determined, terminate the current process. Set a unique node for each group of merged design records, and write the design content carried by the design record into the corresponding node attributes; Based on the electrical connection relationship, the nodes that are related to each other are matched, each electrical connection relationship is encoded as the edge relationship between the corresponding nodes, and edge relationships with the same content between the same nodes are merged. The node attributes and edge relationships are combined according to the corresponding positions of the nodes to generate the power design state diagram.

[0009] Optionally, step three specifically includes: Read the power design state diagram, determine the node arrangement position according to the correspondence of each node in the electrical connection structure, and determine the writing position of each node attribute in the node arrangement position according to the design object category; Extract the node attributes carried by each node, convert the numerical content of the same attribute into a unified engineering unit, convert the category content into a category code, and combine them into a node attribute vector of the same length according to the corresponding writing position. The node attribute vectors are combined sequentially according to the node arrangement position, and the positions, connection directions and relationship contents of the interconnected nodes are written into the corresponding positions according to the edge relationship to form a node attribute matrix and an edge relationship matrix. Identify the node attributes corresponding to each operating condition in the power design state diagram. While keeping the node arrangement position and edge relationship matrix unchanged, replace the operating condition change content in the node attribute matrix with the attribute content corresponding to the current operating condition, and retain the attribute content that does not change with the operating condition. The node attribute matrix and edge relationship matrix corresponding to each operating condition are combined according to the correspondence of the operating conditions to generate the power design state tensor.

[0010] Optionally, the improved Graphormer model includes an object encoding layer, a condition propagation layer, a residual attribution layer, and a corrected decoding layer, with an electrical path constraint mechanism embedded in the condition propagation layer; The object encoding layer reads the power design state tensor according to the operating conditions and extracts the numerical attributes and category codes corresponding to each design object. The numerical attributes and category codes are transformed to a unified feature dimension and fused using a corresponding mapping matrix. The bias between interconnected nodes is calculated based on the edge relationship to generate an object encoding tensor. The working condition propagation layer maps the object encoding tensor into a query vector, a key vector, and a value vector, and calculates the node relevance in each attention head respectively; Among them, the electrical path constraint mechanism determines the effective influence path based on the electrical connection direction, design object category and node attributes, and generates node interaction mask based on the effective influence path; Subsequently, the relational bias is added to the node correlation, normalization and weighted aggregation are performed within the scope of the node interaction mask, the node representation is updated along the direction of the design parameters acting on the running response, and the node correspondence between each running condition is maintained to generate a multi-condition response representation. The residual attribution layer maps the multi-condition response characterization to the response quantity corresponding to the engineering constraint, calculates the deviation of the response quantity from the range limited by the engineering constraint, and generates the design constraint residual. Starting from the node that generates the design constraint residual, the associated node representations are collected along the reverse direction of the effective influence path. The attribution weights of the design parameters on the design constraint residuals within the optimization range are calculated, and parameter attribution representations are generated. The correction decoding layer determines the adjustment direction and correction amount of each design parameter based on the parameter attribution characterization and design constraint residuals. It uses the attribution weight as the correlation strength between the design parameter and the associated design constraint, and combines the design constraint residuals, adjustment direction and correction amount according to the corresponding relationship to generate a parameter correction evidence set.

[0011] Optionally, the electrical path constraint mechanism specifically includes: Read the edge relationships, design object categories, and node attributes corresponding to the object encoding tensor, determine the node carrying the design parameters as the path start point, and determine the node carrying the running response as the path end point; Based on the electrical connection direction in the edge relationship, the path extends layer by layer from the starting point to the adjacent node, forming a directed node sequence from the starting point of the path through the associated design object to the end point of the path; According to the design object category, check the order of action of objects in the directed node sequence to confirm that the design parameters can be applied to the running response through the corresponding design objects, and determine the directed node sequence with continuous connection direction and consistent action order as the effective influence path. Effective influence paths with the same starting and ending points are organized accordingly, and the node arrangement and influence direction of each path are preserved to form a set of effective influence paths. Node interaction masks are generated according to the node arrangement positions in the object encoding tensor. Node pairs that are in the same effective influence path and conform to the influence direction are set to the allowed interaction state, the remaining node pairs are set to the masked state, and the node itself is set to the allowed interaction state. The node interaction mask is matched with the object encoding tensor corresponding to each operating condition, so that the operating condition propagation layer only calculates the correlation and aggregates node information between nodes corresponding to the allowed interaction states.

[0012] Optionally, step five specifically includes: Read the parameter correction evidence set and extract the current value, allowable value range, value type, value rule, adjustment direction, correction amount, design constraint residual, associated design constraint and attribution weight for each design parameter; For design parameters that use step-by-step values, the allowable adjustment step size is gradually increased or decreased from the current value according to the adjustment direction to generate candidate values ​​that do not exceed the correction amount and are within the allowable value range. When the adjustment direction is to maintain or the design constraint residual is 0, only the current value is retained. For design parameters that adopt specification values, the adjustment direction is moved step by step from the current specification position to the adjacent specification position. The standard specifications that are within the adjustment level corresponding to the correction amount and are within the allowable value range are determined as candidate values. The current processing is terminated when the current value is not in the standard specification sequence. Write the current value of each design parameter into the corresponding candidate value, merge duplicate candidate values, and arrange them according to the step value order or standard specification order; The corresponding design parameters are aggregated using the associated design constraints as an index, and the adjustment direction, correction amount, design constraint residual, attribution weight and candidate value of each design parameter are written into the same adjustment association; The current value, candidate value, allowed value range, value type, value rule, and adjustment association of each design parameter are combined accordingly to generate the candidate domain of power design parameters.

[0013] Optionally, step six specifically includes: The candidate domain of power design parameters is input into the improved ALNS algorithm. The current values ​​of each design parameter in the design scheme to be optimized form the initial parameter group, and the design constraint residuals corresponding to the candidate values ​​are read according to the adjustment correlation. Based on the sorting results of the design constraint residuals and the design parameters of the associated coverage, the destruction order and destruction range are determined. The residual priority destruction and associated parameter joint destruction are called, the current value binding of the corresponding parameter position is released, and the parameter group to be repaired is generated. According to the parameter position in the parameter group to be repaired, the adjustment direction repair and the associated coordination repair are called. The legal values ​​are selected from the corresponding candidate values ​​and the associated design parameters acting on the same engineering constraint are coordinated and written to generate the repair parameter group. Combining the failure and repair methods for the same design constraint into operator combinations, setting the same initial selection weights for each operator combination, and recording the design constraint corresponding to each operator combination; The combination of operators called in the current iteration is determined according to the selection weight, so that the destruction operation and the repair operation are processed continuously around the same design constraint, avoiding the separation of adjustment target caused by selecting destruction mode and repair mode separately; Based on the adjustment amount of the candidate value relative to the current value and the corresponding attribution weight, estimate the remaining design constraint residuals of the repair parameter group, count the number of engineering constraints with remaining design constraint residuals greater than 0, and calculate the total number of parameter modifications for the repair parameter group. First, compare the number of engineering constraint violations, then compare the total amount of remaining design constraint residuals. If the above results are consistent, compare the total amount of parameter modifications. Determine the acceptance result of the repair parameter group according to the comparison level, and update the accepted repair parameter group to the current parameter group. Based on the acceptance results and the improvement amount of design constraint residuals, update the selection weights of the operator combination and the destruction range of the next iteration. When the termination condition of the current search round is reached, combine the current parameter group and the parameter group that meets the hierarchical acceptance condition during the iteration to generate a candidate power design parameter group.

[0014] Optionally, step seven specifically includes: Substitute the candidate power design parameter groups into the design scheme to be optimized in turn, keeping the design content and electrical connection relationship not included in the optimization scope unchanged; The engineering verification items and verification order corresponding to each candidate power design parameter group are determined according to the power design parameter optimization constraint set. Power flow calculations are performed on operating status constraints, short-circuit current calculations and equipment capacity verification are performed on fault capacity constraints, and protection action coordination calculations are performed on protection coordination constraints. The calculation results of each project are compared with the corresponding project constraints. The comparison results are then collected according to the same candidate power design parameter group to generate the project verification results. Record the candidate power design parameter sets that satisfy all engineering constraints; Update the corresponding design constraint residuals based on the degree of deviation from the unmet engineering constraints, and feed the update results back to the current search state of the improved ALNS algorithm according to the adjustment correlation. Based on the updated design constraint residuals, the selection weights and violation ranges of the operator combination are adjusted, and candidate power design parameter sets are generated and engineering verification is performed. When the result feedback termination condition is met, the candidate power design parameter groups that have passed the engineering verification are collected to generate a feasible power design parameter set.

[0015] Optionally, step eight specifically includes: Read the set of feasible power design parameters, match each set of feasible power design parameters with the design scheme to be optimized, and determine the design parameters that have changed and the absolute amount of their value changes. For the design parameters of step values, calculate the ratio of the actual number of adjustment steps to the total number of steps within the allowable range; for the design parameters of specification values, calculate the ratio of the number of specification levels crossed to the total number of allowable specification levels. Sum the ratios corresponding to the same feasible power design parameter group to generate the parameter modification range; Read the engineering verification results corresponding to each feasible power design parameter group, and calculate the normalized engineering margin according to the constraint form of the engineering constraints. For the upper limit constraint, calculate the difference between the upper limit value and the engineering calculation result, and convert it according to the absolute value of the upper limit value; For the lower limit constraint, calculate the difference between the engineering calculation result and the lower limit value, and convert it according to the absolute value of the lower limit value; For interval constraints, take the smaller distance between the engineering calculation result and the upper and lower limits, and convert it according to the allowable interval span; The design performance is generated by taking the minimum value of the normalized engineering margin of each item in the same feasible power design parameter group; Compare the parameter modification range and design performance group by group, and retain the feasible power design parameter group that has no other feasible power design parameter group, whose parameter modification range is no greater than that of the parameter group, whose design performance is no less than that of the parameter group, and whose at least one aspect is strictly better than that of the parameter group; The remaining feasible power design parameter groups are arranged in ascending order according to the parameter modification range, and in descending order according to the design performance when the parameter modification range is the same. The feasible power design parameter groups that have been arranged and the corresponding engineering verification results are combined to generate power design parameter optimization results.

[0016] The beneficial effects of this invention are: This invention addresses the problems of diverse design object categories, complex electrical connection relationships, and significant differences in operating conditions in original power design data by cascading and integrating the improved Graphormer model and the improved ALNS algorithm. It constructs a power design state diagram based on electrical connection relationships and performs tensor quantization and multi-condition combination on node attributes and edge relationships to achieve a unified expression of design object attributes, electrical topology relationships, and operating condition changes. In the parameter correction analysis stage, this invention integrates numerical attributes, category codes, and edge relationships through heterogeneous design object encoding and multi-condition response propagation. It uses the electrical path constraint mechanism to limit the node interaction range, so that the node representation is updated along the effective influence path of the design parameters acting on the operating response. Based on the design constraint residual of the operating response relative to the engineering constraint limit range, the adjustment direction, correction amount, and attribution weight of each design parameter are determined in reverse, forming a parameter correction evidence set and a candidate domain of power design parameters. In the parameter search and result determination stage, this invention uses design constraint residuals to drive parameter destruction and candidate domain constraint repair. It combines destruction and repair methods for the same engineering constraint into operator combinations, and adaptively updates the selection weights and destruction ranges based on the acceptance results and residual improvement. It implements hierarchical acceptance through engineering constraint satisfaction results, design constraint residuals, and parameter modification ranges. At the same time, it feeds back the engineering verification results formed by power flow calculation, short-circuit current calculation, equipment capacity verification, and protection action coordination calculation to the search process. Finally, it selects feasible power design parameter groups according to parameter modification ranges and design performance, thereby improving the matching degree between power design parameter optimization results and multiple operating conditions and engineering constraints. Attached Figure Description

[0017] 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 an overall flowchart of a deep learning-based intelligent optimization method for power design parameters proposed in this invention. Figure 2 This is a schematic diagram of the dual-objective screening process for the feasible power design parameter set proposed in this invention; Figure 3 This is a diagram showing the dual-objective dominance relationship of the feasible power design parameter set proposed in this invention. Detailed Implementation

[0018] 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.

[0019] refer to Figures 1-3 A method for intelligent optimization of power design parameters based on deep learning includes the following steps: Step 1: Receive power design parameter optimization task information, determine the design scheme to be optimized and engineering constraints, and generate a power design parameter optimization constraint set; Step 2: Read the original power design data according to the power design parameter optimization constraint set, perform correlation coding based on electrical connection relationships, and generate a power design state diagram; Step 3: Tensorize the node attributes and edge relationships in the power design state diagram, perform combinations according to the operating conditions, and generate the power design state tensor; Step 4: Input the power design state tensor into the improved Graphormer model, and generate a parameter correction evidence set through heterogeneous design object encoding, multi-condition response propagation under electrical influence path constraints, reverse attribution of design constraint residuals, and parameter correction decoding. Step 5: Determine candidate values ​​and adjust correlations according to the parameter correction evidence set to generate candidate domains for power design parameters; Step 6: Input the candidate domain of power design parameters into the improved ALNS algorithm, and generate a set of candidate power design parameters through parameter destruction driven by design constraint residuals, candidate domain constraint repair, adaptive selection of destruction-repair operator combination, and hierarchical acceptance with priority given to engineering constraints. Step 7: Perform engineering verification and result feedback on the candidate power design parameter set to generate a feasible power design parameter set; Step 8: Filter the feasible power design parameter set according to the parameter modification range and design performance, and generate power design parameter optimization results.

[0020] In this embodiment, step one specifically includes: Receive power design parameter optimization task information. The task information includes project identifier, design version, optimization target scope and engineering constraint source. Perform joint matching on existing design schemes according to project identifier and design version. When a valid design scheme is found, it is identified as the unique design scheme. If the matching result is empty or multiple valid design schemes exist at the same time, record the termination reason and stop the current processing. Based on the scope of optimization, the unique design scheme was verified. The optimization content included the cross-section of outgoing conductors, transformer capacity, rated breaking current of corresponding circuit breakers and protection setting values. The above design content was included in the adjustment scope. The bus connection method, equipment installation location and design content not specified by the task were kept in their original state, thus forming the design scheme to be optimized. Match engineering constraints from the engineering design requirements and technical conditions applicable to the design scheme to be optimized. The voltage deviation shall not exceed 7% of the rated voltage, the transformer load rate shall not exceed 0.85, the rated breaking current of the circuit breaker shall not be lower than the maximum short-circuit current, and the time difference between the upper and lower protection actions shall not be lower than 0.30s as the corresponding design content restrictions. Engineering constraints that are not applicable to the scope of this adjustment will not be written into the processing results. Engineering constraints with the same target, constraint type, and constraint value are merged into one record. Engineering constraints with different constraint values ​​for the same design content are compared in terms of their applicable scope. When the general design requirement limits the transformer load rate to no more than 0.90, while the current engineering design requirement limits it to no more than 0.85, the 0.85 constraint that covers the current engineering and has a narrower limitation range is retained. When conflicting constraints have the same applicable scope and a unique limitation result cannot be determined, the current processing is terminated. The engineering constraints that have been matched, merged, and conflict-handled are associated with the design scheme to be optimized according to the correspondence of the design objects, thereby generating a set of power design parameter optimization constraints; By optimizing the constraint set of power design parameters, the scope of reading the original power design data is limited, and a unified benchmark is provided for subsequent engineering constraint calculations and result verification.

[0021] In this embodiment, step two specifically involves: Based on the power design parameter optimization constraint set, the corresponding objects of the design scheme to be optimized, the adjustment range, and the engineering constraints, the equipment design records, circuit design records, electrical connection records, and operating condition records are read from the original power design data; the design object identifier is used as the main matching basis, and the circuit affiliation and connection endpoints are used as the auxiliary matching basis, and only the design records that have a direct object correspondence with the design scheme to be optimized are retained. The read design records are merged according to the design object identifier. Records with the same object identifier, design object category and design content are merged into the same record location. When the same design object has different design content, the merging result is determined by the corresponding content that is in a valid state in the design scheme to be optimized, and the corresponding operating condition content is retained. The current processing is terminated when there is no corresponding content or multiple valid contents in the design scheme to be optimized. A unique node is set for each group of merged design records. The nodes are distinguished according to the design object categories corresponding to busbars, transformers, outgoing lines, circuit breakers and protection devices. The design object category, rated parameters, current design parameters and operating conditions are written into the node attributes. When a field is missing, a missing status mark is written. The attributes of other nodes are not used to replace it. Match the corresponding nodes based on the connection endpoints in the electrical connection record, determine the edge direction according to the connection endpoint order and the port direction of the design object, and encode the connection method and connection object category as edge relationship content; when there are duplicate edge relationships with the same content between nodes, merge them into one edge relationship, and when there are different connection contents between nodes, retain the corresponding edge relationships respectively. The node attributes are combined according to the generation order of the unique nodes, and the edge relationships are written between the corresponding nodes to generate the power design state diagram. By maintaining the corresponding structure between the design object attributes and electrical connection relationships through the power design state diagram, a unified graph structure input is provided for the tensor quantization processing of node attributes and edge relationships.

[0022] In this embodiment, step three specifically includes: Read the power design state diagram, take the power supply side node as the starting position, and determine the arrangement position of each node in the electrical connection structure along the edge direction; the nodes in the same connection level are stably arranged according to the design object category and design object identifier, and the numerical attribute, category code and missing status mark are written in the node attribute vector according to the design object category. Extract the node attributes carried by each node, and convert the voltage to kilovolts, the current to amperes, the capacity to kilovolt-amperes, the conductor cross-section to square millimeters, and the protection action time to seconds. The design object category adopts one-hot coding, and the coding length is equal to the total number of design object categories in the power design state diagram. Write 1 in the corresponding category position and 0 in the other positions. Combine the attributes of each node according to the determined writing position to form a node attribute vector of the same length; when a node is missing a corresponding numerical attribute, write 0 in the numerical position and write 1 in the corresponding missing status mark position; when the attribute is valid, write 0 in the missing status mark and do not use other node attributes to supplement the missing content. Write the node attribute vectors sequentially according to the node arrangement position to form a node attribute matrix; determine the row and column indices of the edge relationship matrix based on the edge relationship, and set a fixed-length relationship encoding vector at each row and column position. The encoding length is equal to the total number of encoding bits corresponding to the edge direction, connection method, and connection object category; write 1 according to the encoding position corresponding to each edge relationship. When multiple edge relationships exist for the same node pair, write 1 at the corresponding position of the same relationship encoding vector. Write all 0 vectors at the position where no edge relationship is established to form an edge relationship matrix. Read the node attributes corresponding to each operating condition in sequence according to the operating condition identifier. Replace the operating condition change content in the node attribute matrix with the load, current, voltage and protection action time under the current operating condition. Keep the rated parameters, current design parameters, design object category and edge relationship matrix unchanged. Combine the node attribute matrix corresponding to each operating condition with the same edge relationship matrix in the order of operating condition identifier to generate the power design state tensor. By maintaining the consistency of node locations and electrical connections under different operating conditions through the power design state tensor, a unified input is provided for improving the multi-condition response propagation of the Graphormer model.

[0023] In this embodiment, the improved Graphormer model includes an object encoding layer, a condition propagation layer, a residual attribution layer, and a corrected decoding layer. The condition propagation layer embeds an electrical path constraint mechanism, specifically: The power design state tensor is input into the improved Graphormer model. The input includes the node attribute matrix and the shared edge relationship matrix corresponding to each operating condition. Before input, the node arrangement position, node attribute length and the row and column positions of the edge relationship matrix under different operating conditions are checked. If any position is inconsistent, the current model processing is terminated. The object encoding layer reads the power design state tensor according to the operating conditions and extracts the numerical attributes, category codes and missing state markers corresponding to each design object. The numerical attributes are normalized by minimum-maximum according to the upper and lower limits of the attributes of the same type of design object in the model training data. The result below the lower limit is set to 0, the result above the upper limit is set to 1, and the normalization result is set to 0 when the upper and lower limits are the same. Missing values ​​are kept at 0 and distinguished by missing state markers. Numerical attributes are linearly mapped to 64-dimensional features. Category encoding and missing state labels are combined and linearly mapped to 64-dimensional features. These are then concatenated in the order of numerical features first, followed by category and missing state features, to form 128-dimensional features. After 128-dimensional fusion mapping and activation by Gaussian error linear units, an initial 128-dimensional node representation is generated. Edge relationships are encoded and mapped to relationship biases corresponding to 8 attention heads. When the same node pair has multiple edge relationships, the corresponding relationship biases are superimposed to generate an object encoding tensor. The operating condition propagation layer sets up four serially connected and parameter-independent multi-head attention propagation units, with different operating conditions sharing the same set of model parameters; each propagation unit maps the 128-dimensional node representation to query vector, key vector and value vector respectively, and divides it into eight 16-dimensional attention heads. The dot product of the query vector and the key vector is calculated and divided by 4 to obtain the node relevance in each attention head. The electrical path constraint mechanism determines the effective influence path based on the electrical connection direction, design object category, and node attributes, and generates node interaction masks based on the effective influence path. The mask value of the node pair that allows interaction is 0, the mask value of the node pair that is shielded is negative infinity, and the mask value of the node itself is 0. The node interaction mask and relationship bias are added to the node correlation, so that the node information only propagates along the effective influence path of the design parameters acting on the running response. Softmax normalization is applied to the node relevance. The value vector is then weighted and aggregated using the normalization weights. The aggregation results of the eight attention heads are concatenated into a 128-dimensional feature. After output mapping, the feature is added to the input of the propagation unit and subjected to residual addition and layer normalization. Then, the following steps are performed sequentially: 128-dimensional to 256-dimensional linear mapping, Gaussian error linear unit activation, 256-dimensional to 128-dimensional linear mapping, residual addition, and layer normalization. After completing four propagation units, the node representations are combined according to the operating condition sequence and node arrangement to generate multi-condition response representations. The residual attribution layer sets corresponding response mapping branches according to voltage deviation, maximum short-circuit current, load rate and protection action time. It reads the node representations of the corresponding operating conditions and design objects from the multi-condition response representation and outputs the response quantity with corresponding engineering units. A numerical stability term of 1×10^-6 is uniformly added to each normalized denominator. The excess ratio of the response quantity relative to the engineering constraint limit range is calculated according to the upper limit constraint, lower limit constraint and interval constraint respectively. When it does not exceed the limit range, the deviation is set to 0, and the design constraint residual is generated. Starting from the node that generates the design constraint residual, read the node representations corresponding to the design parameters within the optimization range in the reverse direction of the effective influence path; map the residual node representations and design parameter node representations into 64-dimensional query vectors and key vectors respectively, calculate the dot product and divide by 8, mask the design parameter nodes that are not on the corresponding effective influence path, perform Softmax normalization on the retained results, and generate the attribution weights of each design parameter under the same engineering constraint; The design parameter node representation, residual node representation, design constraint residual, and attribution weight are concatenated and a 128-dimensional parameter attribution representation is generated through linear mapping and Gaussian error linear unit activation. The correction decoding layer inputs the parameter attribution representation into the adjustment direction branch and the correction amount branch respectively. The adjustment direction branch outputs the probabilities of three directions: decrease, maintain, and increase. The correction amount branch outputs the correction ratio from 0 to 1 through the Sigmoid function. The step value parameter converts the correction ratio into an integer multiple of the allowable adjustment step size, and the specification value parameter converts the correction ratio into the standard specification adjustment level. When the adjustment direction is to maintain or the design constraint residual is 0, the correction amount is set to 0. According to the correspondence between the design parameters and the associated design constraints, the current value of the design parameter, the allowable value range, the value type, the value rule, the design constraint residual, the adjustment direction, the correction amount, and the attribution weight are combined. The attribution weight is used as the association strength to generate a parameter correction evidence set. Both the original Graphormer and the improved Graphormer models retain node representation, query-key-value mapping, multi-head self-attention, structural relation bias, feedforward mapping, residual connections, and layer normalization. The improved Graphormer model adds a numerical attribute and category encoding split mapping at the original node input encoding position, adds an electrical path constraint mechanism and electrical relation bias at the original attention correlation calculation position, and adds a residual attribution layer and a correction decoding layer at the original graph representation output position. This transforms the general graph structure representation process of the original model into a process of multi-condition response propagation, design constraint residual reverse attribution, and design parameter correction evidence generation. By improving the Graphormer model, the power design state tensor is transformed into a parameter correction evidence set with clear engineering constraints and parameter correlations, providing a computational basis for determining candidate values ​​and adjusting correlations for each design parameter.

[0024] In this embodiment, the electrical path constraint mechanism is specifically as follows: The electrical path constraint mechanism reads the edge relationships, design object categories, and node attributes corresponding to the object encoding tensor. It determines the nodes that carry the design parameters based on the current design parameters in the node attributes and determines the nodes that carry the operating response based on the operating condition content in the node attributes. When an edge relationship lacks a connection endpoint, connection direction, or connection object category, the corresponding edge relationship is marked as invalid and does not participate in path expansion. Each node carrying design parameters is used as the starting point of the path. Adjacent nodes are visited according to the order of the connection endpoints in the edge relationship and the direction of the design object port. During the path expansion process, the visited nodes are recorded. When the same path reaches a visited node again, the current branch stops. When the node carrying the running response is reached, the current directed node sequence is written into the path to be checked. The current branch ends when there are no unvisited valid adjacent nodes for the current node. Check the order of action of objects in the path to be checked according to the design object category, and check the connection direction between adjacent nodes; if the connection direction is interrupted, the order of action of objects is inconsistent, or the running response carried by the end point of the path has no relationship with the design parameters corresponding to the start point of the path, delete the current path to be checked. A sequence of directed nodes that are continuous in direction, have the same order of action, and can reach the end point of the path from the starting point is identified as an effective influence path. Effective influence paths with the same starting point and ending point are organized accordingly, and the node arrangement relationship and influence direction of each path are retained to form a set of effective influence paths. Node interaction masks are generated according to the node arrangement positions in the object encoding tensor. The row positions of the node interaction mask correspond to the query nodes, and the column positions correspond to the key nodes. Node pairs that are in the same effective influence path and conform to the influence direction are written to 0, the node itself is written to 0, and the remaining node pairs are written to negative infinity. When the same node pair appears in multiple effective influence paths at the same time, the interaction state is allowed, and the interaction weight is not added repeatedly. The node interaction mask is matched with the object encoding tensor corresponding to each operating condition according to the node arrangement position. Different operating conditions share the same node interaction mask. The operating condition propagation layer adds the node interaction mask to the node relevance of each attention head, so that the normalized weight of the masked node pair is 0, and only aggregates node information based on the relationship bias and node relevance between nodes that are allowed to interact. The effective impact path set is provided to the residual attribution layer to aggregate the design parameter node representations along the path in reverse. The node interaction mask is provided to the operating condition propagation layer to limit the propagation range of multi-operating condition response. The electrical path constraint mechanism limits the node interactions in the object encoding tensor to an effective influence path that conforms to the electrical connection direction and the order of action of the design object, providing consistent path constraints for multi-condition response propagation and reverse attribution of design constraint residuals.

[0025] In this embodiment, step five specifically includes: Read the parameter correction evidence set and extract the current value, allowed value range, value type, value rule, adjustment direction, correction amount, design constraint residual, associated design constraint, and attribution weight for each design parameter. When the same design parameter corresponds to multiple parameter correction evidences, retain each adjustment basis according to the associated design constraint and do not merge the attribution weights corresponding to different associated design constraints. For step value parameters, read the allowable adjustment step size in the value selection rules, and gradually increase or decrease the allowable adjustment step size from the current value according to the adjustment direction to generate candidate values ​​that do not exceed the correction amount and are within the allowable value range; when the adjustment direction is to maintain or the design constraint residual is 0, only the current value is retained; For specification value parameters, the standard specification sequence and the standard specification adjustment level corresponding to the correction amount in the value rules are read. The adjustment is moved step by step from the current specification position to the adjacent specification position according to the adjustment direction. The standard specifications within the adjustment level and within the allowed value range are determined as candidate values. The current processing is terminated when the current value is not in the standard specification sequence. Write the current value of each design parameter into the corresponding candidate value, merge duplicate candidate values, and arrange them according to the step value order or standard specification order; The parameter correction evidence is grouped according to the associated design constraints. Adjustment relationships are established for design parameters that act on the same associated design constraint. The adjustment direction, correction amount, design constraint residual, attribution weight, and candidate value are recorded for each design parameter. When the same design parameter is associated with multiple design constraints, the corresponding adjustment relationships are retained separately. The current value, candidate value, allowed value range, value type, value rule, and adjustment association of each design parameter are combined accordingly to generate a candidate domain of power design parameters; By limiting the range of values ​​for each design parameter participating in the optimization search through the candidate domain of power design parameters, and maintaining the adjustment relationship between design parameters and associated design constraints, a search basis is provided for improving the execution parameter violation and candidate domain constraint repair of the ALNS algorithm.

[0026] In this embodiment, step six specifically includes: The candidate domain of power design parameters is input into the improved ALNS algorithm. The current value, candidate value, allowable value range, value type, value rule and adjustment association of each design parameter are read. The adjustment direction, correction amount, design constraint residual and attribution weight are read from the adjustment association. The initial parameter group is formed with the current value of each design parameter. For any parameter group, the adjustment completion ratio is calculated based on the actual adjustment amount of each design parameter relative to the current value. For step value parameters, the actual number of adjustment steps is divided by the number of steps corresponding to the correction amount. For specification value parameters, the actual number of adjustment levels is divided by the adjustment level corresponding to the correction amount. When the number of steps or adjustment levels corresponding to the correction amount is 0, the adjustment completion ratio is set to 0. When the obtained ratio is higher than 1, it is set to 1. The constraint improvement ratio is determined by the sum of the product of the adjustment completion ratio of each design parameter under the same associated design constraint and the attribution weight. The original design constraint residual is multiplied by the difference obtained by subtracting the constraint improvement ratio to generate the remaining design constraint residual. The number of engineering constraints with a remaining design constraint residual greater than 0 is counted. When calculating the total parameter modification amount, the step value parameter is calculated by dividing the actual number of adjustment steps by the total number of adjustable steps in the candidate domain, and the specification value parameter is calculated by dividing the actual number of adjustment levels by the total number of adjustable levels in the candidate domain. When the total number of adjustable steps or the total number of adjustable levels in the candidate domain is 0, the corresponding parameter modification amount is set to 0. The normalized results of each item are added together to obtain the total parameter modification amount. Set up single-parameter failure operators and associated parameter failure operators. The single-parameter failure operator selects design parameters from high to low according to the product of the remaining design constraint residuals and the attribution weights and removes their current values. The associated parameter failure operator selects the adjustment association with the largest remaining design constraint residuals and removes the current values ​​of all design parameters in the adjustment association to form the position of the parameter to be repaired. The system sets up directional repair operators and correlation repair operators. The directional repair operator selects values ​​from the corresponding candidate values ​​that meet the adjustment direction and value selection rules, calculates the number of engineering constraint violations, the total amount of remaining design constraint residuals, and the total amount of parameter modifications after writing, and selects the value with the smaller result according to the order of the three indicators. The correlation repair operator combines the candidate values ​​of each design parameter to be repaired in turn according to the adjustment correlation. After writing each parameter, it retains the combination with the better order of the three indicators until all parameters to be repaired are written, generating a candidate parameter group. If no available candidate values ​​are found, the value of the corresponding design parameter before the damage is restored. The destruction operator and the repair operator are paired and set with the same initial weight. The calling probability is determined according to the proportion of the combined weight to the sum of all combined weights. After each operator combination is called 10 times, the combined weight is updated. The combination that reduces the number of engineering constraint violations scores 2. The combination that has the same number of engineering constraint violations and reduces the total amount of remaining design constraint residuals scores 1. The combination that has the same first two items and reduces the total amount of parameter modification scores 1. The other combinations score 0. The combined weight is updated by the sum of 0.8 of the original combined weight and 0.2 of the average score of this round. If the update result is lower than 0.1, it is set to 0.1. For candidate parameter groups, a hierarchical acceptance process prioritizing engineering constraints is implemented. First, the number of engineering constraint violations is compared, then the total amount of remaining design constraint residuals is compared, and finally the total amount of parameter modifications is compared. Parameter groups with smaller corresponding indices are retained in turn. If all three indices are the same, the parameter group generated earlier is retained. The parameter groups that pass the hierarchical acceptance are used as the current parameter groups for the next round and written into the candidate file. Only one set of parameter groups with completely identical values ​​in the candidate file is retained. The search stops when no new parameter set is written to the candidate file for 30 consecutive iterations, and stops synchronously when the number of iterations reaches 200. The candidate files are arranged in the order of the number of engineering constraint violations, the total amount of remaining design constraint residuals, and the total amount of parameter modifications to generate candidate power design parameter sets. Both the original ALNS algorithm and the improved ALNS algorithm retain the search framework of initial parameter set construction, parameter destruction, parameter repair, operator adaptive selection, candidate parameter set acceptance, and iteration termination. The improved ALNS algorithm adds design constraint residuals and attribution weights at the position where the original algorithm selects the destruction object, adds power design parameter candidate domains and adjustment direction constraints at the position where the original algorithm performs repair from the complete value space, changes the position where the original algorithm evaluates destruction operators and repair operators separately to evaluate the combination of destruction and repair operators, and sets a hierarchical acceptance order for the number of engineering constraint violations, the total amount of remaining design constraint residuals, and the total amount of parameter modifications at the position where the original algorithm uniformly accepts candidate parameter sets, so that the general neighborhood search is transformed into a power design parameter search with engineering constraints as the priority. By improving the ALNS algorithm, a set of candidate power design parameters that meet the value selection rules and maintain adjustment correlation is generated within the candidate domain of power design parameters, providing input parameters to be verified for subsequent engineering verification and result feedback.

[0027] In this embodiment, step seven specifically includes: Read the candidate power design parameter group, substitute each group into the design scheme to be optimized in the order of arrangement, write the candidate parameter value into the corresponding design object, and keep the design content, electrical connection relationship and operating condition not included in the optimization scope in their original state to form the power design scheme to be verified. Based on the power design parameter optimization constraint set, power flow calculation, short-circuit current calculation, equipment capacity verification and protection action coordination calculation are performed to obtain the node voltage, branch current, transformer load rate, maximum short-circuit current and upper and lower level protection action time under each operating condition. Calculate the deviation of the node voltage from the rated voltage, check whether it exceeds 7% of the rated voltage, check whether the transformer load rate exceeds 0.85, check whether the rated breaking current of the circuit breaker is lower than the maximum short-circuit current, and check whether the time difference between the upper and lower level protection actions is less than 0.30s, and form the engineering verification results. Calculate the actual design constraint residuals according to the upper limit, lower limit or allowable range of each engineering constraint. When the corresponding limit is met, the residual is set to 0. When the limit is exceeded, the corresponding excess ratio is retained. The actual design constraint residuals are associated with the corresponding candidate power design parameter group. When all actual design constraint residuals are 0, the corresponding candidate power design parameter group is written into the temporary result of the feasible parameter group; when there are actual design constraint residuals, the actual design constraint residuals are replaced with the original design constraint residuals and written into the corresponding adjustment association in the power design parameter candidate domain according to the evidence set of the associated design constraint matching parameters. Keeping the candidate values ​​and value rules in the candidate domain of power design parameters unchanged, the improved ALNS algorithm is called to re-execute parameter destruction, candidate domain constraint repair and hierarchical acceptance, generate an updated candidate power design parameter group, and re-execute engineering verification; The same parameter group is not repeatedly verified. The parameter group that satisfies all engineering constraints is written into the temporary result of the feasible parameter group. The result feedback ends when all candidate power design parameter groups have been verified and no new parameter groups to be verified have been generated. Duplicate parameter groups in the temporary results of feasible parameter groups are merged, and the optimized values, current values ​​of corresponding design parameters, allowable value ranges, value types, value rules, and engineering verification results of each parameter group are combined according to the correspondence of design parameters to generate a feasible power design parameter set. By verifying engineering results and providing feedback, the candidate power design parameter set is transformed into a feasible power design parameter set that meets the requirements of electrical calculation and equipment coordination, providing input for screening based on the parameter modification range and design performance.

[0028] In this embodiment, step eight specifically includes: Read the feasible power design parameter set, and extract the optimized values, current values ​​of the corresponding design parameters, allowable value ranges, value types, value rules, and engineering verification results from each feasible power design parameter group; For step-value parameters, calculate the ratio of the actual number of adjustment steps to the total number of steps within the allowable value range; for specification-value parameters, calculate the ratio of the actual number of adjustment levels to the total number of adjustment levels within the allowable value range; add the calculation results of each design parameter in the same feasible power design parameter group to generate the parameter modification range; Calculate the normalized engineering margin of each associated design constraint based on the engineering verification results; for upper limit constraints, calculate the difference between the upper limit of the constraint and the engineering calculation result, and divide it by the absolute value of the upper limit of the constraint; for lower limit constraints, calculate the difference between the engineering calculation result and the lower limit of the constraint, and divide it by the absolute value of the lower limit of the constraint; for interval constraints, calculate the difference between the engineering calculation result and the lower limit of the interval and the difference between the upper limit of the interval and the engineering calculation result, respectively, and take the smaller of the two differences and divide it by the allowable interval span. Compare the normalized engineering margins corresponding to the same feasible power design parameter group, and determine the minimum normalized engineering margin as the design performance of the parameter group, so that the design performance characterizes the constraint with the lowest margin among all engineering constraints. The parameter modification range and design performance of each feasible power design parameter group are compared; when the parameter modification range of one parameter group is no greater than that of another parameter group, the design performance is no less than that of another parameter group, and at least one of them is strictly better than the other parameter group, the former parameter group is determined to dominate the latter parameter group, and the dominated parameter group is deleted. like Figure 3 The diagram shows the dual-objective dominance relationship of feasible power design parameter groups. With the parameter modification range as the horizontal axis and the design performance as the vertical axis, each feasible power design parameter group is mapped to a corresponding coordinate point. The horizontal axis value increases from left to right, and the vertical axis value increases from bottom to top. The parameter group located in the upper left has a smaller parameter modification range and higher design performance. Solid dots represent feasible power design parameter groups that are not dominated by other parameter groups, while hollow dots represent feasible power design parameter groups that are dominated by other parameter groups. Connect the solid dots in ascending order of parameter modification range to form the distribution boundary of the non-dominated parameter group. This boundary is used to represent the trade-off between parameter modification range and design performance. This distribution boundary does not participate in the recalculation of parameter modification range and design performance. The feasible power design parameter groups that are not dominated by other parameter groups are retained and arranged in ascending order of parameter modification range. When the parameter modification ranges are the same, they are arranged in descending order of design performance. The optimized values, parameter modification ranges, design performance, and corresponding engineering verification results are combined to generate the power design parameter optimization results. By jointly screening the parameter modification range and design performance, parameter groups with small parameter adjustment requirements and high engineering margins are retained in the feasible power design parameter set, providing an output basis for parameter modification of power design schemes.

[0029] Example 1: To verify the feasibility of this invention in practice, it was applied to optimize the design parameters of a 10kV power distribution project in an industrial park. The design to be optimized includes transformers, outgoing circuits, circuit breakers, and protection devices. Under normal operation, peak load, and maintenance / transfer conditions, the cross-section of the outgoing conductors, transformer capacity, rated breaking current of the circuit breaker, and protection setting values ​​were adjusted. The original design had problems such as a high transformer load rate, voltage deviation at the end approaching the limit, insufficient breaking capacity of the circuit breaker, and insufficient time difference between the action of upstream and downstream protection devices. Manual methods required repeated parameter adjustments and electrical calculations.

[0030] In the implementation process, firstly, a set of power design parameter optimization constraints is generated based on the optimization task information, the design scheme to be optimized, and engineering constraints. Then, equipment design records, circuit design records, electrical connection records, and operating condition records are read to generate a power design state diagram and a power design state tensor. Subsequently, an improved Graphormer model is used to perform heterogeneous design object encoding, multi-condition response propagation under electrical path constraints, reverse attribution of design constraint residuals, and parameter correction decoding to generate a set of parameter correction evidence, and based on this, candidate values ​​and adjustment correlations are determined. Then, an improved ALNS algorithm is used to complete parameter destruction, candidate domain constraint repair, adaptive selection of destruction-repair operator combinations, and hierarchical acceptance with priority given to engineering constraints, generating a set of candidate power design parameters. Finally, a feasible set of power design parameters is formed through power flow calculation, short-circuit current calculation, equipment capacity verification, and protection action coordination calculation, and power design parameter optimization results are generated according to the parameter modification range and design performance.

[0031] To verify the processing effect of this invention, three sets of comparison methods were set up: a manual rule iteration method, a standard Graphermer combined with ALNS method, and the method of this invention. The manual rule iteration method adjusts parameters item by item based on the experience of engineers; the standard Graphermer combined with ALNS method uses a general graph representation and a random destruction-repair strategy; the method of this invention uses an improved Graphermer model with embedded electrical path constraint mechanisms and an improved ALNS algorithm driven by design constraint residuals. The test objects included 30 sets of power distribution engineering design data, with 3 optimization tasks set for each set, resulting in 90 sets of effective design tasks. All methods used the same original power design data, operating conditions, and engineering constraints. The statistical results are shown in Table 1.

[0032] Table 1. Comparison of Test Results for Different Power Design Parameter Optimization Methods

[0033] As shown in Table 1, the manual rule-based iterative method struggles to uniformly handle the interrelationships between different design parameters, achieving an 82.2% first-order optimization constraint satisfaction rate and an average optimization time of 39.4 min. The ordinary Graphormer combined with the ALNS method can utilize graph structure information to perform parameter searches, but lacks electrical influence path constraints and reverse attribution of design constraint residuals, resulting in an average parameter modification range of 0.176 and a protection coordination pass rate of 94.4%. The method of this invention achieves a 97.8% first-order optimization constraint satisfaction rate, reduces the average parameter modification range to 0.129, increases the minimum normalized engineering margin to 0.093, reduces the maximum voltage deviation to 4.7%, achieves a protection coordination pass rate of 98.9%, and shortens the average optimization time to 10.8 min.

[0034] Among the above indicators, the first-time optimization engineering constraint satisfaction rate reflects the proportion of the candidate power design parameter group generated for the first time that satisfies all engineering constraints; the average parameter modification range reflects the comprehensive adjustment degree of the optimized parameters relative to the current design parameters; the minimum normalized engineering margin reflects the safety margin of the most restrictive engineering constraints; the maximum voltage deviation reflects the maximum change of node voltage relative to the rated voltage; the protection coordination pass rate reflects the proportion of the time difference between upper and lower level protection actions that meets the constraints; and the average optimization time reflects the processing time required to complete a single set of optimization tasks.

[0035] The above results are achieved by the method of the present invention mainly because the improved Graphormer model uses the electrical connection direction and the order of action of design objects to limit the propagation of node information, and determines the attribution weight of design parameters to design constraint residuals in reverse along the effective influence path; the improved ALNS algorithm selects the destructive objects according to the design constraint residuals and attribution weights, performs targeted repair within the candidate domain of power design parameters, and accepts them in layers according to the number of engineering constraint violations, the total amount of remaining design constraint residuals, and the total amount of parameter modifications, thereby reducing invalid searches and unnecessary parameter adjustments.

[0036] As can be seen from the above implementation results, the present invention can complete the correlation analysis of power design parameters under multiple operating conditions, candidate value generation, parameter combination search, engineering verification and result screening. Under the conditions of meeting the requirements of voltage deviation, transformer load rate, equipment breaking capacity and protection coordination, it can obtain power design parameter optimization results with a small parameter modification range and high engineering margin.

[0037] 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 method for intelligent optimization of power design parameters based on deep learning, characterized in that, Includes the following steps: Step 1: Receive power design parameter optimization task information, determine the design scheme to be optimized and engineering constraints, and generate a power design parameter optimization constraint set; Step 2: Read the original power design data according to the power design parameter optimization constraint set, perform correlation coding based on electrical connection relationships, and generate a power design state diagram; Step 3: Tensorize the node attributes and edge relationships in the power design state diagram, perform combinations according to the operating conditions, and generate the power design state tensor; Step 4: Input the power design state tensor into the improved Graphormer model, and generate a parameter correction evidence set through heterogeneous design object encoding, multi-condition response propagation under electrical influence path constraints, reverse attribution of design constraint residuals, and parameter correction decoding. Step 5: Determine candidate values ​​and adjust correlations according to the parameter correction evidence set to generate candidate domains for power design parameters; Step 6: Input the candidate domain of power design parameters into the improved ALNS algorithm, and generate a set of candidate power design parameters through parameter destruction driven by design constraint residuals, candidate domain constraint repair, adaptive selection of destruction-repair operator combination, and hierarchical acceptance with priority given to engineering constraints. Step 7: Perform engineering verification and result feedback on the candidate power design parameter set to generate a feasible power design parameter set; Step 8: Filter the feasible power design parameter set according to the parameter modification range and design performance, and generate power design parameter optimization results.

2. The intelligent optimization method for power design parameters based on deep learning according to claim 1, characterized in that, Step one specifically involves: Analyze the power design parameter optimization task information, locate the unique design scheme according to the task direction, and terminate the current process if no unique design scheme is obtained. The design scheme is checked according to the optimization scope defined by the task. The design content within the optimization scope is included in the adjustment scope, while the design content outside the optimization scope is kept in its original state, and a design scheme to be optimized is generated. Match the corresponding engineering constraints to the design scheme to be optimized, delete engineering constraints that are irrelevant to the design scheme to be optimized, and merge duplicate engineering constraints with the same target and restriction content. For engineering constraints that affect the same design content and have conflicting results, compare their applicable scope, retain the engineering constraints that cover the design scheme to be optimized and have a narrower scope, and terminate the current processing if a unique result cannot be determined. The engineering constraints that have been matched, merged, and conflict-handled are associated with the design scheme to be optimized, generating a set of power design parameter optimization constraints.

3. The intelligent optimization method for power design parameters based on deep learning according to claim 1, characterized in that, Step two specifically involves: According to the reading range limited by the power design parameter optimization constraint set, the design record corresponding to the design scheme to be optimized is read from the original power design data; Merge design records based on the object correspondence between them. Merge design records that correspond to the same object and have the same content into the same record location. For design records with inconsistent content, determine the merging result based on the corresponding content in the design scheme to be optimized. If the merging result cannot be determined, terminate the current process. Set a unique node for each group of merged design records, and write the design content carried by the design record into the corresponding node attributes; Based on the electrical connection relationship, the nodes that are related to each other are matched, each electrical connection relationship is encoded as the edge relationship between the corresponding nodes, and edge relationships with the same content between the same nodes are merged. The node attributes and edge relationships are combined according to the corresponding positions of the nodes to generate the power design state diagram.

4. The intelligent optimization method for power design parameters based on deep learning according to claim 1, characterized in that, Step three specifically involves: Read the power design state diagram, determine the node arrangement position according to the correspondence of each node in the electrical connection structure, and determine the writing position of each node attribute in the node arrangement position according to the design object category; Extract the node attributes carried by each node, convert the numerical content of the same attribute into a unified engineering unit, convert the category content into a category code, and combine them into a node attribute vector of the same length according to the corresponding writing position. The node attribute vectors are combined sequentially according to the node arrangement position, and the positions, connection directions and relationship contents of the interconnected nodes are written into the corresponding positions according to the edge relationship to form a node attribute matrix and an edge relationship matrix. Identify the node attributes corresponding to each operating condition in the power design state diagram. While keeping the node arrangement position and edge relationship matrix unchanged, replace the operating condition change content in the node attribute matrix with the attribute content corresponding to the current operating condition, and retain the attribute content that does not change with the operating condition. The node attribute matrix and edge relationship matrix corresponding to each operating condition are combined according to the correspondence of the operating conditions to generate the power design state tensor.

5. The intelligent optimization method for power design parameters based on deep learning according to claim 1, characterized in that, The improved Graphormer model includes an object encoding layer, a condition propagation layer, a residual attribution layer, and a corrected decoding layer, with an electrical path constraint mechanism embedded in the condition propagation layer. The object encoding layer reads the power design state tensor according to the operating conditions and extracts the numerical attributes and category codes corresponding to each design object. The numerical attributes and category codes are transformed to a unified feature dimension and fused using a corresponding mapping matrix. The bias between interconnected nodes is calculated based on the edge relationship to generate an object encoding tensor. The working condition propagation layer maps the object encoding tensor into a query vector, a key vector, and a value vector, and calculates the node relevance in each attention head respectively; Among them, the electrical path constraint mechanism determines the effective influence path based on the electrical connection direction, design object category and node attributes, and generates node interaction mask based on the effective influence path; Subsequently, the relational bias is added to the node correlation, normalization and weighted aggregation are performed within the scope of the node interaction mask, the node representation is updated along the direction of the design parameters acting on the running response, and the node correspondence between each running condition is maintained to generate a multi-condition response representation. The residual attribution layer maps the multi-condition response characterization to the response quantity corresponding to the engineering constraint, calculates the deviation of the response quantity from the range limited by the engineering constraint, and generates the design constraint residual. Starting from the node that generates the design constraint residual, the associated node representations are collected along the reverse direction of the effective influence path. The attribution weights of the design parameters on the design constraint residuals within the optimization range are calculated, and parameter attribution representations are generated. The correction decoding layer determines the adjustment direction and correction amount of each design parameter based on the parameter attribution characterization and design constraint residuals. It uses the attribution weight as the correlation strength between the design parameter and the associated design constraint, and combines the design constraint residuals, adjustment direction and correction amount according to the corresponding relationship to generate a parameter correction evidence set.

6. The intelligent optimization method for power design parameters based on deep learning according to claim 5, characterized in that, The electrical path constraint mechanism is specifically as follows: Read the edge relationships, design object categories, and node attributes corresponding to the object encoding tensor, determine the node carrying the design parameters as the path start point, and determine the node carrying the running response as the path end point; Based on the electrical connection direction in the edge relationship, the path extends layer by layer from the starting point to the adjacent node, forming a directed node sequence from the starting point of the path through the associated design object to the end point of the path; According to the design object category, check the order of action of objects in the directed node sequence to confirm that the design parameters can be applied to the running response through the corresponding design objects, and determine the directed node sequence with continuous connection direction and consistent action order as the effective influence path. Effective influence paths with the same starting and ending points are organized accordingly, and the node arrangement and influence direction of each path are preserved to form a set of effective influence paths. Node interaction masks are generated according to the node arrangement positions in the object encoding tensor. Node pairs that are in the same effective influence path and conform to the influence direction are set to the allowed interaction state, the remaining node pairs are set to the masked state, and the node itself is set to the allowed interaction state. The node interaction mask is matched with the object encoding tensor corresponding to each operating condition, so that the operating condition propagation layer only calculates the correlation and aggregates node information between nodes corresponding to the allowed interaction states.

7. The intelligent optimization method for power design parameters based on deep learning according to claim 1, characterized in that, Step five specifically involves: Read the parameter correction evidence set and extract the current value, allowable value range, value type, value rule, adjustment direction, correction amount, design constraint residual, associated design constraint and attribution weight for each design parameter; For design parameters that use step-by-step values, the allowable adjustment step size is gradually increased or decreased from the current value according to the adjustment direction to generate candidate values ​​that do not exceed the correction amount and are within the allowable value range. When the adjustment direction is to maintain or the design constraint residual is 0, only the current value is retained. For design parameters that adopt specification values, the adjustment direction is moved step by step from the current specification position to the adjacent specification position. The standard specifications that are within the adjustment level corresponding to the correction amount and are within the allowable value range are determined as candidate values. The current processing is terminated when the current value is not in the standard specification sequence. Write the current value of each design parameter into the corresponding candidate value, merge duplicate candidate values, and arrange them according to the step value order or standard specification order; The corresponding design parameters are aggregated using the associated design constraints as an index, and the adjustment direction, correction amount, design constraint residual, attribution weight and candidate value of each design parameter are written into the same adjustment association; The current value, candidate value, allowed value range, value type, value rule, and adjustment association of each design parameter are combined accordingly to generate the candidate domain of power design parameters.

8. The intelligent optimization method for power design parameters based on deep learning according to claim 1, characterized in that, Step six specifically involves: The candidate domain of power design parameters is input into the improved ALNS algorithm. The current values ​​of each design parameter in the design scheme to be optimized form the initial parameter group, and the design constraint residuals corresponding to the candidate values ​​are read according to the adjustment correlation. Based on the sorting results of the design constraint residuals and the design parameters of the associated coverage, the destruction order and destruction range are determined. The residual priority destruction and associated parameter joint destruction are called, the current value binding of the corresponding parameter position is released, and the parameter group to be repaired is generated. According to the parameter position in the parameter group to be repaired, the adjustment direction repair and the associated coordination repair are called. The legal values ​​are selected from the corresponding candidate values ​​and the associated design parameters acting on the same engineering constraint are coordinated and written to generate the repair parameter group. Combining the failure and repair methods for the same design constraint into operator combinations, setting the same initial selection weights for each operator combination, and recording the design constraint corresponding to each operator combination; The combination of operators called in the current iteration is determined according to the selection weight, so that the destruction operation and the repair operation are processed continuously around the same design constraint, avoiding the separation of adjustment target caused by selecting destruction mode and repair mode separately; Based on the adjustment amount of the candidate value relative to the current value and the corresponding attribution weight, estimate the remaining design constraint residuals of the repair parameter group, count the number of engineering constraints with remaining design constraint residuals greater than 0, and calculate the total number of parameter modifications for the repair parameter group. First, compare the number of engineering constraint violations, then compare the total amount of remaining design constraint residuals. If the above results are consistent, compare the total amount of parameter modifications. Determine the acceptance result of the repair parameter group according to the comparison level, and update the accepted repair parameter group to the current parameter group. Based on the acceptance results and the improvement amount of design constraint residuals, update the selection weights of the operator combination and the destruction range of the next iteration. When the termination condition of the current search round is reached, combine the current parameter group and the parameter group that meets the hierarchical acceptance condition during the iteration to generate a candidate power design parameter group.

9. The intelligent optimization method for power design parameters based on deep learning according to claim 1, characterized in that, Step seven specifically involves: Substitute the candidate power design parameter groups into the design scheme to be optimized in turn, keeping the design content and electrical connection relationship not included in the optimization scope unchanged; The engineering verification items and verification order corresponding to each candidate power design parameter group are determined according to the power design parameter optimization constraint set. Power flow calculations are performed on operating status constraints, short-circuit current calculations and equipment capacity verification are performed on fault capacity constraints, and protection action coordination calculations are performed on protection coordination constraints. The calculation results of each project are compared with the corresponding project constraints. The comparison results are then collected according to the same candidate power design parameter group to generate the project verification results. Record the candidate power design parameter sets that satisfy all engineering constraints; Update the corresponding design constraint residuals based on the degree of deviation from the unmet engineering constraints, and feed the update results back to the current search state of the improved ALNS algorithm according to the adjustment correlation. Based on the updated design constraint residuals, the selection weights and violation ranges of the operator combination are adjusted, and candidate power design parameter sets are generated and engineering verification is performed. When the result feedback termination condition is met, the candidate power design parameter groups that have passed the engineering verification are collected to generate a feasible power design parameter set.

10. The intelligent optimization method for power design parameters based on deep learning according to claim 1, characterized in that, Step eight specifically involves: Read the set of feasible power design parameters, match each set of feasible power design parameters with the design scheme to be optimized, and determine the design parameters that have changed and the absolute amount of their value changes. For the design parameters of step values, calculate the ratio of the actual number of adjustment steps to the total number of steps within the allowable range; for the design parameters of specification values, calculate the ratio of the number of specification levels crossed to the total number of allowable specification levels. Sum the ratios corresponding to the same feasible power design parameter group to generate the parameter modification range; Read the engineering verification results corresponding to each feasible power design parameter group, and calculate the normalized engineering margin according to the constraint form of the engineering constraints. For the upper limit constraint, calculate the difference between the upper limit value and the engineering calculation result, and convert it according to the absolute value of the upper limit value; For the lower limit constraint, calculate the difference between the engineering calculation result and the lower limit value, and convert it according to the absolute value of the lower limit value; For interval constraints, take the smaller distance between the engineering calculation result and the upper and lower limits, and convert it according to the allowable interval span; The design performance is generated by taking the minimum value of the normalized engineering margin of each item in the same feasible power design parameter group; Compare the parameter modification range and design performance group by group, and retain the feasible power design parameter group that has no other feasible power design parameter group, whose parameter modification range is no greater than that of the parameter group, whose design performance is no less than that of the parameter group, and whose at least one aspect is strictly better than that of the parameter group; The remaining feasible power design parameter groups are arranged in ascending order according to the parameter modification range, and in descending order according to the design performance when the parameter modification range is the same. The feasible power design parameter groups that have been arranged and the corresponding engineering verification results are combined to generate power design parameter optimization results.