A micro-grid source network load storage equipment construction arrangement conflict detection method based on three-dimensional vision and graph optimization algorithm

Abstract

The application provides a micro-grid source network load storage equipment construction arrangement conflict detection method based on three-dimensional vision and graph optimization algorithm, and belongs to the cross technical field of swarm intelligence and computer vision; three-dimensional point clouds of a construction site are collected and preprocessed to obtain registration point clouds, equipment and structural member data are synchronously extracted, and electrical safety distance and maintenance channel specification data are called. A three-dimensional semantic model is built and a bounding box is generated; the bounding box is taken as a node, and the safety distance is taken as an edge weight to build a physical constraint graph, and an improved graph convolution network is input to complete layout conflict detection. The channel width is obtained, and the operation and maintenance accessibility evaluation index is calculated in combination with the specification. A space constraint matrix is established relying on the safety distance, the operation and maintenance accessibility is taken as a hard constraint, and the layout conflict is reduced and the space utilization is improved as the goal, and the optimal equipment layout scheme is output by means of heuristic search iterative optimization. The scheme can provide intelligent decision support for micro-grid source network load storage equipment construction arrangement.

Description

Technical Field

[0001] This invention belongs to the field of swarm intelligence and computer vision, and in particular relates to a method for detecting conflicts in the construction layout of microgrid power generation, grid, load and storage equipment based on three-dimensional vision and graph optimization algorithms. Background Technology

[0002] With the continuous advancement of distributed energy and microgrid technologies, the spatial arrangement of various types of electrical equipment at power construction sites for integrated power generation, grid, load, and storage projects is becoming increasingly prominent. As core components of microgrid systems, the rationality of the spatial layout of energy storage cabinets, grid-connected cabinets, transformers, and distribution equipment directly affects construction safety, maintenance accessibility, and overall system efficiency. At the same time, the physical conflicts and insufficient maintenance channels caused by the dense arrangement of equipment within limited construction space are constantly intensifying, posing a serious challenge to the compliance of power construction regulations.

[0003] Currently, the main methods for detecting spatial conflicts in the construction layout of power equipment include the following: (1) Collision detection method based on traditional Building Information Modeling (BIM): This method mainly uses commercial software such as Navisworks and Solibri to perform hard and soft collision detection on building electromechanical pipelines, so that component conflicts in the design model can be exposed in the virtual environment. The operation process is relatively mature and easy to implement. However, this method usually relies on manually preset collision rules and screening of components by category, resulting in a lack of semantic-level recognition capability for microgrid-specific electrical equipment, and the inability to establish differentiated constraints based on the electrical characteristics of energy storage cabinets and grid-connected cabinets, which limits its application in the fine construction scenario of source-grid-load-storage.

[0004] (2) Substation site selection and capacity determination method based on heuristic search algorithm: Using genetic algorithm (GA) or particle swarm optimization (PSO) algorithm, the site and capacity of distribution transformers are planned macroscopically with the goal of minimizing grid losses and investment costs. This type of method has strong global search capability and can handle multi-objective trade-offs. However, since this method focuses on site selection at the geospatial level, it does not involve the three-dimensional fine layout of equipment in the room, and lacks a quantitative assessment of the accessibility of operation and maintenance channels, it is difficult to meet the fine-grained requirements of dense equipment layout inside the microgrid construction site.

[0005] (3) Spatial layout optimization methods based on general graph neural networks (GNN): Some research and engineering practices have attempted to model the spatial topology of thermal equipment or renewable energy facilities using graph convolutional networks (GCN), hoping to reduce the probability of layout conflicts through the propagation of node features. However, these methods rely on general graph structures and lack deep integration with power construction specifications. They fail to transform electrical safety distance standards and maintenance passage specifications into computable graph edge constraints, making it difficult for the optimization results to directly meet the mandatory requirements of the power industry.

[0006] In summary, existing power equipment construction and layout technologies have certain limitations in terms of 3D semantic modeling accuracy, physical constraint graph construction methods, and specification-driven optimization capabilities: traditional BIM collision detection lacks microgrid-specific semantics, heuristic site selection methods do not address fine-grained equipment-level layout, and general graph neural networks do not integrate power specification constraints. Summary of the Invention

[0007] To address the above problems, this invention proposes a method for detecting conflicts in the construction layout of microgrid power generation, grid-load, and energy storage equipment based on 3D vision and graph optimization algorithms, comprising the following steps: S1 collects 3D point cloud data of the construction site, outputs registered point cloud data after preprocessing, extracts equipment parameter data and structural component data from BIM model data, and obtains electrical safety distance standard data and maintenance passage specification data. S2, input the registered point cloud data into the point cloud semantic segmentation sub-network and output the point cloud set with semantically labeled data. After the BIM semantic mapping layer is fused with the equipment parameter data, a three-dimensional semantic model is generated, and bounding box data is constructed. Physical constraint graph data is constructed with bounding box data as graph nodes and electrical safety distance standard data as edge weights. S3 inputs the physical constraint graph data into the improved graph convolutional network to output the conflict detection result data, extracts the maintenance channel width data through the accessibility assessment module, and calculates the operation and maintenance accessibility assessment index data based on the channel continuity and turning radius constraints; S4 constructs a spatial distance constraint matrix using electrical safety distance standard data. Within the feasible space defined by the constraint matrix, it uses operation and maintenance accessibility evaluation index data as hard constraints and aims to minimize the number of conflicts and maximize space utilization. It then iteratively optimizes the layout by using a heuristic search algorithm to output an optimized layout scheme.

[0008] Preferably, the equipment parameter data includes the equipment's three-dimensional bounding box dimensions, rated capacity, installation height, and wiring port locations; the structural component data includes the geometric boundaries and material properties of walls, columns, beams, and floor slabs, and establishes a mapping table between globally unique identifiers and geometric instances of components; the electrical safety distance standard data forms a rule entry library containing the minimum clearance between devices under different voltage levels and the minimum distance from live parts to grounded parts; the maintenance passage specification data includes quantitative thresholds for a minimum width of 1.5 meters for the main passage, a minimum width of 1.2 meters for the operating passage in front of the equipment, and a clearance height of not less than 2.2 meters.

[0009] Preferably, the point cloud semantic segmentation subnetwork consists of 4 cascaded SA modules. Each SA module performs three operations sequentially: First, it selects a fixed number of key points from the input point set as sub-sampling centers through farthest point sampling, and performs downsampling layer by layer; then, it collects neighborhood points within a specified search radius using a ball query algorithm, centered on each key point, to form a local point set; finally, it performs point-by-point feature mapping on each point in each local point set using a multilayer perceptron with shared weights, and then performs max pooling to aggregate the local feature vectors of the corresponding key points; 4 layers of SA The parameter configuration and dimensional changes of module A are as follows: Layer 1 downsamples the input point set to one-quarter and outputs 64-dimensional local features; Layer 2 downsamples to one-sixteenth and outputs 128-dimensional features; Layer 3 downsamples to one-sixty-fourth and outputs 256-dimensional features; Layer 4 downsamples to one-two-hundred-and-fifty-sixth and outputs a 512-dimensional global feature vector. Each point is then classified into one of seven categories—energy storage cabinet, grid-connected cabinet, transformer, busbar trunking, distribution box, structural component, or channel—through a 3-layer fully connected classification head, and the output is a point cloud set with semantically labeled data.

[0010] Preferably, the BIM semantic mapping layer rigidly registers the design coordinate system defined by the equipment parameter data in the BIM model data with the global coordinate system of the registered point cloud data, and extracts the geometric boundaries, material attributes, and globally unique identifiers of BIM components and transmits them to the point cloud clusters of the corresponding semantic label data to generate a unified three-dimensional semantic model. Furthermore, the semantic instances in the three-dimensional semantic model are distinguished into equipment instances and structural component instances, and axis-aligned bounding boxes are extracted from them respectively. The center coordinates, length, width, and height three-dimensional dimensions, and semantic category codes are recorded as bounding box data.

[0011] Preferably, the physical constraint graph data is constructed as follows: Each device instance and structural component instance is used as a graph node. The node attribute vector is initialized, where the first three dimensions are the bounding box center coordinates, the fourth dimension is the device category code, and the fifth dimension is the bounding box volume. Next, the initial topology graph is constructed using the spatial adjacency relationship between nodes as edges. The adjacency determination condition is that the projection overlap rate of the bounding boxes of two nodes in any coordinate axis direction exceeds 30%. The minimum safety distance threshold for the corresponding voltage level and equipment combination is queried based on the electrical safety distance standard data, and the threshold is assigned to the weight of the corresponding edge. A spatial constraint graph relaxation engine is introduced to perform global optimization of the initial topology graph. The optimization objective is to minimize the sum of squared residuals of all edges, where the residual is defined as the absolute value of the difference between the actual spatial distance and the edge weight threshold. The optimal configuration of node poses is solved iteratively by the LM algorithm, and the optimized physical constraint graph data is output.

[0012] Preferably, the improved graph convolutional network consists of two graph attention layers, one global average pooling layer, and a fully connected conflict classification head. First, the node attribute vectors and edge weight matrices of the physical constraint graph data are input into the first graph attention layer. This layer employs an 8-head attention mechanism, where each head maps 16-dimensional input features to a 4-dimensional attention space. After concatenation, a 32-dimensional primary device node feature vector is output. The attention coefficients are calculated as follows:

[0013] in For nodes For nodes Attention weights For nodes Feature representation, First, a learnable attention parameter vector is generated. Second, the primary device node feature vector and edge weight matrix are input into the second graph attention layer, outputting a 64-dimensional advanced device node feature vector. The edge weights are attenuated by a gating mechanism, with larger weights resulting in lower message passing strength. Third, the advanced device node feature vector is globally averaged and concatenated with the spatial relative position encoding of node pairs. This concatenation is then input into a fully connected conflict classification head, which consists of two 128-dimensional FC layers and one Sigmoid output layer, outputting the spatial crossover probability between each pair of adjacent nodes. Finally, node pairs with a spatial crossover probability greater than 0.5 are marked as conflict states, generating conflict detection result data. Each conflict record includes the conflict node number, the actual spatial distance, and the threshold specified by the corresponding electrical safety distance standard.

[0014] Preferably, the specific data processing procedure of the accessibility assessment module is as follows: First, semantically labeled point cloud clusters representing channel categories are extracted from the generated 3D semantic model. Connected regions are then clustered into independent channel instances using Euclidean clustering. The central path skeleton for each channel instance is extracted, and the maintenance channel width data is output. The width is defined as the length of the 95% confidence interval of the point cloud distribution perpendicular to the central path direction. Second, based on the minimum width threshold of 1.5 meters and the minimum clearance height threshold of 2.2 meters in the maintenance channel specification data, compliance is assessed for each channel instance. Channels with widths below the thresholds or with insufficient height sections are marked as accessibility defect areas. Third, the A* algorithm shortest path distance from each device instance to the nearest compliant channel is calculated, and the weighted sum of this path distance and channel width is used as the device's individual accessibility score. Finally, the individual accessibility scores of all devices are averaged and normalized to the range of 0 to 1, outputting maintenance accessibility evaluation index data, where a value closer to 1 indicates better overall accessibility.

[0015] Preferably, the S4 process specifically includes: First, a spatial distance constraint matrix is ​​constructed based on electrical safety distance standard data. Second, an initial candidate layout scheme population is randomly generated within the feasible space defined by the spatial distance constraint matrix. Third, for each candidate individual, an improved graph convolutional network is called to obtain conflict detection result data, and a reachability assessment module is called to obtain maintenance channel width data and operation and maintenance reachability assessment index data. The fitness value is calculated with the number of conflicts and space utilization rate as optimization objectives and the operation and maintenance reachability assessment index data as hard constraints. The population is iteratively evolved through selection, crossover, mutation, and simulated annealing perturbation operations until the maximum number of iterations or convergence condition is reached. Finally, the individual with the best fitness is selected from the final population for decoding, and the optimized layout scheme data is output. The scheme includes the optimal three-dimensional coordinates of all mobile devices and the corresponding conflict detection result data and operation and maintenance reachability assessment index data.

[0016] Preferably, the spatial distance constraint matrix and initial candidate layout schemes are constructed as follows: A spatial distance constraint matrix is ​​constructed using electrical safety distance standard data and maintenance access specification data, where the matrix elements... Indicates device With equipment exist The minimum permissible displacement increment in the direction is obtained by subtracting the current actual distance from the standard electrical safety distance between the two devices and then dividing by 2. Similarly, constraint matrices are constructed in the Y and Z directions. and Secondly, an initial candidate layout scheme population is randomly generated within the feasible space defined by the constraint matrix. The population size is set to 50 individuals, and each individual is composed of the three-dimensional coordinate offset encoding of all mobile devices. The encoding dimension is M×3, where M is the number of mobile devices. Mobile devices refer to energy storage cabinets, grid-connected cabinets, transformers, busbar trunking, and distribution boxes. Finally, boundary verification is performed on each candidate individual to ensure that the equipment does not exceed the physical boundary of the construction site after being moved and does not intrude into the space occupied by structural components.

[0017] Preferably, the iterative optimization using a heuristic search algorithm specifically involves the following process: First, for each candidate individual, an improved graph convolutional network is called to obtain conflict detection results data. Then, the reachability assessment module is called to obtain maintenance channel width data and operation and maintenance reachability assessment index data. Finally, its fitness value is calculated. The fitness function is defined as:

[0018] in For fitness value, The number of collisions in the collision detection results data. This is the sum of the volume occupied by the equipment and the volume of the maintenance passage. The total volume of the construction site. and These are conflict weights and space utilization weights, respectively. add The fitness value is set to 1. In the fitness calculation, the operational accessibility assessment index is introduced as a hard constraint: if the operational accessibility assessment index of the current individual is lower than 0.6, the fitness value of that individual is directly set to zero to ensure its elimination in the selection process. Secondly, a tournament selection strategy is used to select superior individuals from the current population to enter the mating pool, and offspring individuals are generated through single-point crossover in the mating pool, with a crossover probability set to 0.8. Thirdly, a Gaussian mutation operation is performed on the offspring individuals, with a mutation probability set to 0.15. The mutation amplitude is adaptively reduced by the current iteration number. Simultaneously, simulated annealing perturbation is performed on the bottom 20% of individuals in fitness ranking, accepting inferior solutions with a certain probability to maintain population diversity. Then, the parent and offspring generations are merged, and elite retention is performed based on fitness values, retaining the top 50 individuals to form the next generation population. Finally, the above evolutionary process is repeated until the maximum number of iterations is reached or the optimal fitness change is less than 0.001 for 20 consecutive generations. The individual with the best fitness is selected from the final population for decoding, and the optimized layout scheme data is output.

[0019] Compared with the prior art, the present invention has the following beneficial effects: (1) A three-dimensional semantic modeling method for all elements of source-grid-load-storage construction site was constructed by integrating LiDAR point cloud and BIM model. The equipment parameter data, structural component data and electrical safety distance standard data were semantically aligned and uniformly encoded to realize the mapping of construction site from discrete scanning data to structured semantic model. The three-dimensional semantic model generated by this method can be directly used for quantitative calculation of spatial conflicts, avoiding information loss and accuracy deviation caused by traditional manual modeling or format conversion. (2) An improved microgrid equipment spatial layout optimization architecture coupled with a graph convolutional network and heuristic search was designed. While extracting high-order topological features from the physical constraint graph data, spatial iterative optimization was performed using operation and maintenance accessibility assessment index data as hard constraints. The optimized layout scheme data was dynamically distributed to the on-site construction terminal through a visualization interface. This architecture can automatically generate compliant equipment layout schemes while meeting electrical safety distance standards, reducing the workload of repeated manual trial layouts and scheme verification. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the following description is only one embodiment of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating the overall technical route of the present invention.

[0022] Figure 2 This is a flowchart of the data acquisition and preprocessing process of the present invention.

[0023] Figure 3 This is a schematic diagram illustrating the semantic model construction and constraint graph establishment structure of this invention.

[0024] Figure 4 This is a diagram of the conflict detection and reachability assessment control architecture of the present invention.

[0025] Figure 5 This is a layout diagram of the spatial iterative optimization system of the present invention.

[0026] Figure 6 This is a comparison chart of collision detection accuracy in embodiments of the present invention.

[0027] Figure 7 This is a distribution diagram of operation and maintenance accessibility evaluation indicators in an embodiment of the present invention.

[0028] Figure 8 This is a layout optimization convergence curve diagram in an embodiment of the present invention. Detailed Implementation

[0029] This invention proposes a conflict detection method for the construction layout of microgrid power generation, grid-load, and storage equipment based on 3D vision and graph optimization algorithms. The specific steps are as follows: Figure 1 As shown: S1. Three-dimensional data acquisition and preprocessing at the construction site: This step aims to acquire comprehensive geometric and semantic information of the microgrid construction site, providing a data foundation for subsequent 3D modeling. First, a 3D point cloud data set of the construction site is collected using a 3D laser scanner, outputting an N×4 dimensional original point cloud matrix containing X, Y, and Z coordinates and reflection intensity. Second, BIM model data from the design phase is imported, extracting equipment parameter data and structural component data, and storing them serially in IFC format. Third, electrical safety distance standards and maintenance access specifications are compiled according to power industry regulations, forming a searchable rule entry library. Finally, the 3D point cloud data of the construction site is normalized in coordinate system and scaled, outputting registered point cloud data.

[0030] S2. Construction of 3D semantic model and establishment of physical constraint diagram: This step aims to construct a full-element 3D semantic model of the construction site and establish unique physical constraints between equipment. The semantic fusion and graph optimization module consists of a point cloud semantic segmentation sub-network, a BIM semantic mapping layer, and a spatial constraint graph relaxation engine (Pose Graph Optimization, PGO). First, the registered point cloud data is input into the point cloud semantic segmentation sub-network, outputting a set of point cloud data with semantic labels. Each point is labeled as one of ten categories: energy storage cabinet, grid-connected cabinet, transformer, busbar, distribution box, wall, column, beam, floor slab, or passageway. Second, the equipment parameter data and semantic label data in the BIM model data are aligned and their attributes are fused through the BIM semantic mapping layer to generate a unified 3D semantic model. Third, physical constraint graph data is constructed using equipment and structural components as nodes and spatial adjacency relationships and electrical safety distance standard data as edge weights, utilizing a spatial constraint graph relaxation engine. Each node stores 3D bounding box coordinates and equipment category code, and each edge stores the minimum safety distance threshold and connection type identifier. Finally, the physical constraint graph data undergoes topology consistency verification, eliminating isolated nodes and redundant edges.

[0031] S3. Spatial Conflict Identification and Operational Accessibility Assessment: This step aims to identify spatial conflicts of equipment and quantitatively assess operational accessibility based on physical constraint graph data. The improved graph convolutional network consists of two graph attention layers (GAL), one pooling layer, and a conflict classification head. First, the physical constraint graph data is input into the improved graph convolutional network, and neighbor node information is aggregated layer by layer through the two graph attention layers. The first layer maps node features from 16 dimensions to 32 dimensions and outputs a primary device node feature vector, while the second layer maps features from 32 dimensions to 64 dimensions and outputs a higher-level device node feature vector. Second, based on the primary and higher-level device node feature vectors, the spatial crossover probability between each pair of adjacent nodes is calculated using the conflict classification head to generate conflict detection result data. The conflict determination condition is that the minimum distance between the node bounding boxes is less than the threshold specified by the electrical safety distance standard data. Finally, an accessibility assessment module is constructed. Based on the maintenance channel specification data, the maintenance channel width data is extracted from the three-dimensional semantic model, and operational accessibility assessment index data is calculated based on channel continuity and turning radius constraints.

[0032] S4. Spatial Iterative Optimization and Equipment Layout Output: This step aims to find the optimal equipment layout scheme that satisfies both electrical safety distance and operational accessibility constraints through heuristic search. First, a spatial distance constraint matrix is ​​constructed based on standard electrical safety distance data. Second, an initial population of candidate layout schemes is randomly generated within the feasible space defined by the spatial distance constraint matrix. Third, for each candidate, the improved graph convolutional network of S3-1 is called to obtain conflict detection results, and the accessibility evaluation module of S3-2 is called to obtain maintenance channel width data and operational accessibility evaluation index data. The fitness value is calculated with conflict quantity and space utilization as optimization objectives and operational accessibility evaluation index data as hard constraints. The population is iteratively evolved through selection, crossover, mutation, and simulated annealing perturbation operations until the maximum number of iterations or convergence condition is reached. Finally, the individual with the best fitness is selected from the final population for decoding, outputting optimized layout scheme data. This scheme includes the optimal three-dimensional coordinates of all mobile devices and the corresponding conflict detection results and operational accessibility evaluation index data.

[0033] The present invention will be further described below with reference to embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0034] S1. Three-dimensional data acquisition and preprocessing at the construction site To comprehensively collect geometric information and organize rule-based data at the microgrid construction site, ensuring the integrity and standardization of data for subsequent 3D modeling, the process involves multi-source data acquisition and standardized preprocessing to form registered point cloud data and a rule entry library. For example... Figure 2 As shown, the specific steps include: S1-1 Laser point cloud acquisition and parameter calibration at the construction site: First, a 3D laser scanner was deployed at the microgrid construction site. A station-based scanning method was used to acquire 3D geometric information of the construction area from different stations, resulting in 3D point cloud data of the construction site. The single-station scanning resolution was set to 3 mm, and the point cloud density was no less than 10,000 points per square meter. Second, target sphere registration was performed on the 3D point cloud data of each station. The Iterative Closest Point (ICP) algorithm was used to calculate the rigid body transformation matrix between stations, unifying the data from multiple stations into the same global coordinate system and outputting the registered point cloud data. Third, statistical filtering and voxel mesh downsampling were performed on the registered point cloud data. The voxel size was set to 2 cm to reduce data redundancy while preserving key geometric features. Finally, the processed registered point cloud data was persistently stored in PCD format, and the acquisition timestamp and coordinate system parameters were recorded.

[0035] S1-2 BIM Model Data Import and Standard Data Processing: First, BIM model data for the microgrid project was obtained from the design unit. The IFC semantic structure was analyzed, and equipment parameter data for energy storage cabinets, grid-connected cabinets, transformers, busbars, and distribution boxes were extracted, specifically including the dimensions of the equipment's 3D bounding box, rated capacity, installation height, and wiring port locations. Second, structural component data in the BIM model data was analyzed, specifically including the geometric boundaries and material properties of walls, columns, beams, and floor slabs, and a mapping table between globally unique identifiers and geometric instances of components was established. Third, electrical safety distance standard data was compiled according to GB 50060 "Design Code for 3-110kV High Voltage Power Distribution Equipment" and DL / T 5352 "Design Code for High Voltage Power Distribution Equipment," forming a rule entry library containing minimum clearance between equipment at different voltage levels and minimum distance from live parts to grounded parts. Finally, according to GB... The 50016 "Code for Fire Protection Design of Buildings" and the power operation and maintenance procedures have compiled and standardized data on maintenance access channels, clearly defining quantitative thresholds such as a minimum width of 1.5 meters for main access channels, a minimum width of 1.2 meters for operating access channels in front of equipment, and a minimum clearance height of 2.2 meters for access channels. The five types of equipment mentioned above—energy storage cabinets, grid-connected cabinets, transformers, busbars, and distribution boxes—are used as equipment examples, and the four types of structural components—walls, columns, beams, and floor slabs—are used as structural component examples, providing a data foundation for instance division and physical constraint diagram node construction in the subsequent 3D semantic model.

[0036] S2. Construction of 3D Semantic Model and Establishment of Physical Constraint Diagram To achieve semantic-level fusion of discrete point cloud data from the construction site with the BIM design model, and to establish unique physical constraints between equipment and structural components, this invention designs a semantic fusion and graph optimization module. For example... Figure 3 As shown, the specific steps include: S2-1 Point Cloud Semantic Segmentation and BIM Semantic Mapping: This module takes the registered point cloud data output from S1-1 and the equipment parameter data and structural component data extracted from S1-2 as input to generate a unified 3D semantic model. First, the registered point cloud data is input into the point cloud semantic segmentation sub-network, which consists of 4 cascaded SA (Set Abstraction) modules. Each SA module performs three operations in sequence: First, it selects a fixed number of key points from the input point set as sub-sampling centers through Farthest Point Sampling (FPS) to achieve layer-by-layer downsampling; then, it collects neighborhood points within a specified search radius centered on each key point using the Ball Query algorithm to form a local point set; finally, it uses a multilayer perceptron (MLP) with shared weights to perform point-by-point feature mapping on each point in each local point set and then performs max pooling to aggregate the local feature vectors of the corresponding key points. The parameter configuration and dimensional changes of the 4-layer module are as follows: Layer 1 downsamples the input point set to one-quarter and outputs 64-dimensional local features; Layer 2 downsamples to one-sixteenth and outputs 128-dimensional features; Layer 3 downsamples to one-sixtieth and outputs 256-dimensional features; Layer 4 downsamples to one-two-hundred-and-fifty-sixth and outputs a 512-dimensional global feature vector. A 3-layer fully connected classification head then classifies each point into one of seven categories: energy storage cabinet, grid-connected cabinet, transformer, busbar trunking, distribution box, structural component, or channel, outputting a point cloud set with semantically labeled data. Secondly, the BIM semantic mapping layer maps the equipment in the BIM model data... The design coordinate system defined by the parameter data is rigidly registered with the global coordinate system of the registered point cloud data, with the registration error controlled within 5 cm. The geometric boundaries, material properties, and globally unique identifiers of the BIM components extracted in S1-2 are then transferred to the point cloud clusters of the corresponding semantic label data to generate a unified 3D semantic model. Next, the semantic instances in the 3D semantic model are distinguished into equipment instances and structural component instances. Axis-aligned bounding boxes are extracted from each instance, and the center coordinates, length, width, and height dimensions, as well as the semantic category code, are recorded as bounding box data. Finally, the bounding box data of the 3D semantic model is serialized and stored in JSON format. S2-2 Spatial Constraint Graph Relaxation Engine and Physical Constraint Graph Construction: This module takes bounding box data from a 3D semantic model as input to construct physical constraint graph data. First, it initializes node attribute vectors for each device instance and structural component instance as graph nodes, where the first three dimensions are the bounding box center coordinates, the fourth dimension is the device category code, and the fifth dimension is the bounding box volume. Second, it constructs an initial topology graph using the spatial adjacency relationships between nodes as edges. The adjacency criterion is that the projection overlap rate of the bounding boxes of two nodes exceeds 30% in any coordinate axis direction. It then queries the minimum safety distance threshold for the corresponding voltage level and device combination based on electrical safety distance standards, assigning this threshold to the weight of the corresponding edge. The edge weight mapping relationship can be expressed as follows:

[0037] in For nodes With nodes The weight of the edges between them. The category encoding is used as the semantic category identifier of node i. Next, a spatial constraint graph relaxation engine is introduced to perform global optimization on the initial topology graph. The optimization objective is to minimize the sum of squared residuals of all edges, where the residual is defined as the absolute value of the difference between the actual spatial distance and the edge weight threshold. The optimal configuration of node poses is solved iteratively by the Levenberg-Marquardt (LM) algorithm, and the optimized physical constraint graph data is output. Finally, the topology consistency of the physical constraint graph data is checked, and isolated nodes with a degree of 0 and redundant long edges with a weight greater than 5 meters are removed to ensure the connectivity and constraint compactness of the graph structure.

[0038] S3, Spatial Conflict Identification and Operational Accessibility Assessment To achieve high-precision identification of device spatial conflicts and quantitative assessment of accessibility for maintenance channels on physical constraint graph data, thus ensuring the safety and compliance of layout schemes, this invention designs an improved graph convolutional network and accessibility assessment module. For example... Figure 4 As shown, the specific steps include: S3-1 Improved Graph Convolutional Network for Feature Extraction and Collision Detection: This module takes physical constraint graph data as input to generate collision detection results. The improved graph convolutional network consists of two graph attention layers, one global average pooling layer, and a fully connected collision classification head. First, the node attribute vectors and edge weight matrices of the physical constraint graph data are input into the first graph attention layer. This layer uses an 8-head attention mechanism, where each head maps 16-dimensional input features to a 4-dimensional attention space. The concatenated vectors output a 32-dimensional primary device node feature vector. The attention coefficients can be calculated as follows:

[0039] in For nodes For nodes Attention weights For nodes Feature representation, First, a learnable attention parameter vector is generated. Second, the primary device node feature vector and edge weight matrix are input into the second graph attention layer, outputting a 64-dimensional advanced device node feature vector. The edge weights are attenuated by a gating mechanism, with larger weights resulting in lower message passing strength. Third, the advanced device node feature vector is globally averaged and concatenated with the spatial relative position encoding of node pairs. This concatenation is then input into a fully connected (FC) conflict classification head, which consists of two 128-dimensional FC layers and one sigmoid output layer, outputting the spatial crossover probability between each pair of adjacent nodes. Finally, node pairs with a spatial crossover probability greater than 0.5 are marked as conflict states, generating conflict detection result data. Each conflict record includes the conflict node number, the actual spatial distance, and the threshold specified by the corresponding electrical safety distance standard.

[0040] S3-2 Accessibility Assessment Module: This module takes a 3D semantic model and maintenance channel specification data as input to generate operational accessibility assessment index data. First, it extracts semantically labeled point cloud clusters representing channel categories from the 3D semantic model generated in S2-1. Using the Euclidean clustering algorithm, connected regions are clustered into independent channel instances, and the central path skeleton of each channel instance is extracted, outputting maintenance channel width data. The width is defined as the length of the 95% confidence interval of the point cloud distribution perpendicular to the central path direction. Second, based on the minimum width threshold of 1.5 meters and the minimum clearance height threshold of 2.2 meters in the maintenance channel specification data, compliance is assessed for each channel instance. Channels with widths below the thresholds or with insufficient height sections are marked as accessibility defect areas. Third, the shortest path distance from each device instance to the nearest compliant channel using the A* algorithm is calculated, and the weighted sum of this path distance and the channel width is used as the individual accessibility score for that device. The weighting formula can be expressed as:

[0041] in Score the accessibility of device k. Let k be the shortest path distance from device k to the nearest compliant channel. The width of the nearest channel. and The weighting coefficients and add The value is equal to 1. Finally, the individual accessibility scores of all devices are aggregated by mean and normalized to the range of 0 to 1, and the operation and maintenance accessibility evaluation index data is output, where the closer the value is to 1, the better the overall accessibility.

[0042] S4, Spatial Iterative Optimization and Equipment Layout Output To achieve automatic optimization of equipment layout with the goals of minimizing conflicts and maximizing space utilization, while satisfying the dual constraints of electrical safety distance and operational accessibility, this invention designs a hybrid heuristic spatial iterative optimization module. For example... Figure 5 As shown, the specific steps include: S4-1 Constraint Matrix Construction and Candidate Solution Initialization: First, based on the electrical safety distance standard data and maintenance passage specification data compiled in S1-2, a spatial distance constraint matrix is ​​constructed, where the matrix elements... Indicates device With equipment exist The minimum permissible displacement increment in the direction is obtained by subtracting the current actual distance from the standard electrical safety distance between the two devices and then dividing by 2. Similarly, constraint matrices are constructed in the Y and Z directions. and Secondly, an initial candidate layout scheme population is randomly generated within the feasible space defined by the constraint matrix. The population size is set to 50 individuals, and each individual is composed of the three-dimensional coordinate offset encoding of all mobile devices. The encoding dimension is M×3, where M is the number of mobile devices. Mobile devices refer to energy storage cabinets, grid-connected cabinets, transformers, busbar trunking, and distribution boxes. Finally, boundary verification is performed on each candidate individual to ensure that the equipment does not exceed the physical boundary of the construction site after being moved and does not intrude into the space occupied by structural components. S4-2 Hybrid Heuristic Search and Optimal Solution Output: First, for each candidate individual, the improved graph convolutional network of S3-1 is called to obtain collision detection results data. Then, the reachability evaluation module of S3-2 is called to obtain maintenance channel width data and operation and maintenance reachability evaluation index data. The fitness value is calculated, and the fitness function is defined as:

[0043] in For fitness value, The number of collisions in the collision detection results data. This is the sum of the volume occupied by the equipment and the volume of the maintenance passage. The total volume of the construction site. and These are conflict weights and space utilization weights, respectively. add The fitness value is set to 1. In the fitness calculation, the operational reachability assessment index is introduced as a hard constraint: if the current individual's operational reachability assessment index is below 0.6, the individual's fitness value is directly set to zero to ensure its elimination during the selection phase. Secondly, a tournament selection strategy is used to select superior individuals from the current population to enter the mating pool, where offspring individuals are generated through a single-point crossover operation with a crossover probability of 0.8. Thirdly, a Gaussian mutation operation is performed on the offspring individuals with a mutation probability of 0.15, and the mutation amplitude adaptively decreases with the current iteration number. Simultaneously, the fitness ranking is adjusted. The last 20% of individuals undergo simulated annealing perturbation, accepting inferior solutions with a certain probability to maintain population diversity. Then, the parent and offspring generations are merged, and elites are selected based on fitness values, retaining the top 50 individuals to form the next generation. Finally, the above evolutionary process is repeated until the maximum number of iterations of 200 generations is reached or the optimal fitness change is less than 0.001 for 20 consecutive generations. The individual with the best fitness is selected from the final population for decoding, and the optimized layout scheme data is output. This scheme includes the optimal three-dimensional coordinates of all devices, as well as the corresponding conflict detection results and operation and maintenance accessibility evaluation index data.

[0044] Among them, the maintenance channel volume The calculation formula is ,in For the first The maintenance channel width data for each channel is represented by the 95% confidence interval width extracted from the point cloud by S3-2; For the first The center path length of each channel is extracted from the skeleton by S3-2; To standardize clearance height, the minimum clearance height threshold of 2.2m from the maintenance access standard data is adopted.

[0045] Simulation experiments and analysis: To verify the effectiveness of the proposed method for detecting conflicts in the construction layout of microgrid power generation, grid, load, and storage equipment based on 3D vision and graph optimization algorithms, this experiment uses a microgrid project construction site in an industrial park as the research object. The site includes 6 energy storage cabinets, 3 grid-connected cabinets, 2 transformers, and several supporting power distribution devices. The construction site dimensions are 15m × 12m × 4m. The experimental environment is based on the Ubuntu 20.04 operating system, using the PyTorch 2.0 deep learning framework and the Open3D point cloud processing library. Training and testing were conducted on a workstation equipped with an NVIDIA RTX 4090 GPU and 64GB of memory. Point cloud data acquisition was performed using a FARO Focus S350 laser scanner, and BIM model data was exported as IFC 4.0 format files using Revit 2023.

[0046] The experimental design includes three comparative dimensions: First, the collision detection method of this invention is compared with the baseline method based on traditional axis-aligned bounding box collision detection to verify the improvement in collision recognition accuracy of the improved graph convolutional network; second, the accessibility assessment results of this invention are compared with manual on-site measurement data to verify the accuracy of maintenance channel width data extraction and operation and maintenance accessibility assessment index data calculation; finally, the layout optimization results of this invention are compared with empirically designed layouts to verify the comprehensive advantages of the hybrid heuristic search algorithm in terms of space utilization and compliance.

[0047] analyze Figure 6 The proposed method achieves a precision of 94.2%, a recall of 91.8%, and an F1 score of 93.0% on the collision detection task. This represents a 12.5 percentage point improvement in precision and an 8.3 percentage point improvement in recall compared to the baseline method based on traditional axis-aligned bounding boxes. The main reason for this is that the improved graph convolutional network encodes higher-order topological relationships in the physical constraint graph data through an attention mechanism, enabling it to identify non-adjacent device pairs with potential electromagnetic interference constraints. In contrast, traditional bounding box methods can only detect direct spatial intersections.

[0048] analyze Figure 7 The operational accessibility assessment index data output by this method is distributed within the range of 0.72 to 0.91 across 20 test scenarios, with a mean of 0.82 and a standard deviation of 0.05. Compared to the accessibility scores measured manually on-site, the Pearson correlation coefficient reaches 0.89, and the mean absolute error is 0.04, indicating that the extraction of maintenance channel width data and the quantitative assessment of accessibility by this method have relatively small errors compared to manual measurement. In scenarios where the channel has local narrowness or bends, this method can still accurately capture accessibility bottlenecks through the A* algorithm (A-Star Algorithm) path search.

[0049] analyze Figure 8 It can be seen that the fitness value of the hybrid heuristic search algorithm increases rapidly in the first 80 iterations, from an initial average of 0.42 to 0.78. Subsequently, it enters a fine search phase between generations 80 and 160, where the fitness value slowly converges to around 0.85, finally reaching a stable optimal value of 0.862 in the 187th generation. Compared with the single genetic algorithm, this hybrid algorithm improves the convergence speed by about 35% and the final fitness value by 4.2 percentage points. This is attributed to the simulated annealing perturbation mechanism effectively avoiding premature convergence.

[0050] Table 1. Conflict Detection, Accessibility, and Layout Optimization Assessment

[0051] Analysis of Table 1 shows that the proposed method achieves a precision and recall rate exceeding 90% in the conflict detection dimension, indicating that the improved graph convolutional network has good balance in identifying spatial conflicts. In the accessibility assessment dimension, the correlation coefficient with manual on-site scoring reaches 0.89, verifying the effectiveness of the maintenance passage width data extraction method based on the 3D semantic model. In the layout optimization dimension, the space utilization rate reaches 78.5%, and the compliance rate with regulations reaches 100%, demonstrating that the hybrid heuristic search algorithm can achieve high space utilization efficiency while meeting electrical safety distance standards and maintenance passage specifications. Based on the combined experimental results across these three dimensions, the method proposed in this invention can provide reliable intelligent decision support for the construction and layout of microgrid power-grid-load-storage equipment.

[0052] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0053] While the specific embodiments of the present invention have been described above, they are not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for detecting conflicts in the construction layout of microgrid power generation, grid-load, and energy storage equipment based on 3D vision and graph optimization algorithms, characterized in that, The process includes the following: S1 collects 3D point cloud data of the construction site, outputs registered point cloud data after preprocessing, extracts equipment parameter data and structural component data from BIM model data, and obtains electrical safety distance standard data and maintenance passage specification data. S2, input the registered point cloud data into the point cloud semantic segmentation sub-network and output the point cloud set with semantically labeled data. After the BIM semantic mapping layer is fused with the equipment parameter data, a three-dimensional semantic model is generated, and bounding box data is constructed. Physical constraint graph data is constructed with bounding box data as graph nodes and electrical safety distance standard data as edge weights. S3 inputs the physical constraint graph data into the improved graph convolutional network to output the conflict detection result data, extracts the maintenance channel width data through the accessibility assessment module, and calculates the operation and maintenance accessibility assessment index data based on the channel continuity and turning radius constraints; The improved graph convolutional network consists of two graph attention layers, one global average pooling layer, and a fully connected conflict classification head. First, the node attribute vectors and edge weight matrices of the physical constraint graph data are input into the first graph attention layer. This layer employs an 8-head attention mechanism, where each head maps 16-dimensional input features to a 4-dimensional attention space. The concatenated vectors output a 32-dimensional primary device node feature vector. The attention coefficients are calculated as follows: in For nodes For nodes Attention weights For nodes Feature representation, First, a learnable attention parameter vector is generated. Second, the primary device node feature vector and edge weight matrix are input into the second graph attention layer, outputting a 64-dimensional advanced device node feature vector. The edge weights are attenuated by a gating mechanism, with larger weights resulting in lower message passing strength. Third, the advanced device node feature vector is globally averaged and concatenated with the spatial relative position encoding of node pairs. This concatenation is then input into a fully connected conflict classification head, which consists of two 128-dimensional FC layers and one Sigmoid output layer, outputting the spatial crossover probability between each pair of adjacent nodes. Finally, node pairs with a spatial crossover probability greater than 0.5 are marked as conflict states, generating conflict detection result data. Each conflict record includes the conflict node number, the actual spatial distance, and the threshold specified by the corresponding electrical safety distance standard. The specific data processing procedure of the accessibility assessment module is as follows: First, semantic label data point cloud clusters labeled as channel categories are extracted from the generated 3D semantic model. Connected regions are clustered into independent channel instances using Euclidean clustering algorithm, and the central path skeleton of each channel instance is extracted. The maintenance channel width data is output, where the width is defined as the length of the 95% confidence interval of the point cloud distribution perpendicular to the central path direction. Second, based on the minimum width threshold of 1.5 meters and the minimum clearance height threshold of 2.2 meters in the maintenance channel specification data, compliance is judged for each channel instance. Channels with widths below the thresholds or with insufficient height sections are marked as accessibility defect areas. Next, calculate the shortest path distance from each device instance to the nearest compliant channel using the A* algorithm, and use the weighted sum of this path distance and channel width as the individual reachability score for that device; finally, aggregate the mean of the individual reachability scores of all devices, normalize them to the range of 0 to 1, and output the operation and maintenance reachability evaluation index data, where the closer the value is to 1, the better the overall reachability. S4 constructs a spatial distance constraint matrix using electrical safety distance standard data. Within the feasible space defined by the constraint matrix, it uses operation and maintenance accessibility evaluation index data as hard constraints and aims to minimize the number of conflicts and maximize space utilization. It then iteratively optimizes the layout by using a heuristic search algorithm to output an optimized layout scheme.

2. The method for detecting conflicts in the construction layout of microgrid power generation, grid-load, and energy storage equipment based on three-dimensional vision and graph optimization algorithms as described in claim 1, characterized in that: The equipment parameter data includes the equipment's three-dimensional bounding box dimensions, rated capacity, installation height, and wiring port locations; the structural component data includes the geometric boundaries and material properties of walls, columns, beams, and floor slabs, and establishes a mapping table between globally unique identifiers and geometric instances of components; the electrical safety distance standard data forms a rule entry library containing the minimum clearance between equipment at different voltage levels and the minimum distance from live parts to grounded parts; the maintenance passage specification data includes quantitative thresholds for a minimum width of 1.5 meters for the main passage, a minimum width of 1.2 meters for the operating passage in front of the equipment, and a clearance height of not less than 2.2 meters.

3. The method for detecting conflicts in the construction layout of microgrid power generation, grid-load, and energy storage equipment based on three-dimensional vision and graph optimization algorithms as described in claim 1, characterized in that: The point cloud semantic segmentation subnetwork consists of four cascaded SA modules. Each SA module performs three operations sequentially: First, it selects a fixed number of keypoints from the input point set as sub-sampling centers through farthest point sampling, and performs downsampling layer by layer; then, it collects neighborhood points within a specified search radius using a ball query algorithm, centered on each keypoint, to form a local point set; finally, it uses a multilayer perceptron with shared weights to perform point-by-point feature mapping on each point in each local point set, and then performs max pooling to aggregate the local feature vectors of the corresponding keypoints. The parameter configuration and dimensional changes of the block are as follows: The first layer downsamples the input point set to one-quarter and outputs 64-dimensional local features; the second layer downsamples to one-sixteenth and outputs 128-dimensional features; the third layer downsamples to one-sixty-fourth and outputs 256-dimensional features; the fourth layer downsamples to one-two-hundred-and-fifty-sixth and outputs a 512-dimensional global feature vector. Each point is then classified into one of seven categories—energy storage cabinet, grid-connected cabinet, transformer, busbar trunking, distribution box, structural component, or channel—through a three-layer fully connected classification head, and the output is a point cloud set with semantically labeled data.

4. The method for detecting conflicts in the construction layout of microgrid power generation, grid-load, and storage equipment based on three-dimensional vision and graph optimization algorithms as described in claim 3, characterized in that: The BIM semantic mapping layer rigidly registers the design coordinate system defined by the equipment parameter data in the BIM model data with the global coordinate system of the registered point cloud data. It also extracts the geometric boundaries, material properties, and globally unique identifiers of BIM components and passes them to the point cloud clusters of the corresponding semantic label data to generate a unified three-dimensional semantic model. Furthermore, it distinguishes the semantic instances in the three-dimensional semantic model into equipment instances and structural component instances, extracts axis-aligned bounding boxes for each, and records the center coordinates, length, width, height, and three-dimensional dimensions, as well as the semantic category code, as bounding box data.

5. The method for detecting conflicts in the construction layout of microgrid power generation, grid-load, and energy storage equipment based on three-dimensional vision and graph optimization algorithms as described in claim 4, characterized in that: The data used to construct the physical constraint graph specifically includes: Each device instance and structural component instance is used as a graph node. The node attribute vector is initialized, where the first three dimensions are the bounding box center coordinates, the fourth dimension is the device category code, and the fifth dimension is the bounding box volume. Next, the initial topology graph is constructed using the spatial adjacency relationship between nodes as edges. The adjacency determination condition is that the projection overlap rate of the bounding boxes of two nodes in any coordinate axis direction exceeds 30%. The minimum safety distance threshold for the corresponding voltage level and equipment combination is queried based on the electrical safety distance standard data, and the threshold is assigned to the weight of the corresponding edge. A spatial constraint graph relaxation engine is introduced to perform global optimization of the initial topology graph. The optimization objective is to minimize the sum of squared residuals of all edges, where the residual is defined as the absolute value of the difference between the actual spatial distance and the edge weight threshold. The optimal configuration of node poses is solved iteratively by the LM algorithm, and the optimized physical constraint graph data is output.

6. The method for detecting conflicts in the construction layout of microgrid power-grid-load-storage equipment based on three-dimensional vision and graph optimization algorithms as described in claim 1, characterized in that: The specific process of S4 includes: First, a spatial distance constraint matrix is ​​constructed based on electrical safety distance standard data. Second, an initial candidate layout scheme population is randomly generated within the feasible space defined by the spatial distance constraint matrix. Third, for each candidate individual, an improved graph convolutional network is called to obtain conflict detection result data, and a reachability assessment module is called to obtain maintenance channel width data and operation and maintenance reachability assessment index data. The fitness value is calculated with the number of conflicts and space utilization rate as optimization objectives and the operation and maintenance reachability assessment index data as hard constraints. The population is iteratively evolved through selection, crossover, mutation, and simulated annealing perturbation operations until the maximum number of iterations or convergence condition is reached. Finally, the individual with the best fitness is selected from the final population for decoding, and the optimized layout scheme data is output. The scheme includes the optimal three-dimensional coordinates of all mobile devices and the corresponding conflict detection result data and operation and maintenance reachability assessment index data.

7. The method for detecting conflicts in the construction layout of microgrid power generation, grid-load, and storage equipment based on three-dimensional vision and graph optimization algorithms as described in claim 6, characterized in that: Construct the spatial distance constraint matrix and initial candidate layout schemes as follows: A spatial distance constraint matrix is ​​constructed using electrical safety distance standard data and maintenance access specification data, where the matrix elements... Indicates device With equipment exist The minimum permissible displacement increment in the direction is obtained by subtracting the current actual distance from the standard electrical safety distance between the two devices and then dividing by 2. Similarly, constraint matrices are constructed in the Y and Z directions. and Secondly, an initial candidate layout scheme population is randomly generated within the feasible space defined by the constraint matrix. The population size is set to 50 individuals, and each individual is composed of the three-dimensional coordinate offset encoding of all mobile devices. The encoding dimension is M×3, where M is the number of mobile devices. Mobile devices refer to energy storage cabinets, grid-connected cabinets, transformers, busbar trunking, and distribution boxes. Finally, boundary verification is performed on each candidate individual to ensure that the equipment does not exceed the physical boundary of the construction site after being moved and does not intrude into the space occupied by structural components.

8. The method for detecting conflicts in the construction layout of microgrid power generation, grid, load, and storage equipment based on three-dimensional vision and graph optimization algorithms as described in claim 6, characterized in that: The iterative optimization process using a heuristic search algorithm is as follows: First, for each candidate individual, an improved graph convolutional network is called to obtain conflict detection results data. Then, the reachability assessment module is called to obtain maintenance channel width data and operation and maintenance reachability assessment index data. Finally, its fitness value is calculated. The fitness function is defined as: in For fitness value, The number of collisions in the collision detection results data. This is the sum of the volume occupied by the equipment and the volume of the maintenance passage. The total volume of the construction site. and These are conflict weights and space utilization weights, respectively. add The fitness value is set to 1. In the fitness calculation, the operational accessibility assessment index is introduced as a hard constraint: if the current individual's operational accessibility assessment index is below 0.6, the individual's fitness value is directly set to zero to ensure it is eliminated during the selection process. Secondly, a tournament selection strategy is used to select superior individuals from the current population to enter the mating pool, where offspring individuals are generated through single-point crossover with a crossover probability of 0.

8. Thirdly, Gaussian mutation is performed on the offspring individuals with a mutation probability of 0.15, and the mutation amplitude is adaptively reduced by the current iteration number. Simultaneously, simulated annealing perturbation is performed on the bottom 20% of individuals in fitness ranking to accept inferior solutions with a certain probability, thus maintaining population diversity. Then, the parent and offspring generations are merged, and elite retention is performed based on fitness values, retaining the top 50 individuals to form the next generation population. Finally, the evolutionary process is repeated until the maximum number of iterations is reached or the optimal fitness change is less than 0.001 for 20 consecutive generations. The individual with the best fitness is selected from the final population for decoding, and the optimized layout scheme data is output.

Images