Substation site selection method, system and device based on artificial intelligence multi-factor evaluation and dynamic weight adjustment, and medium

Abstract

The invention discloses a transformer substation site selection method, system and device based on artificial intelligence multi-factor evaluation and dynamic weight adjustment and a medium, and belongs to the technical field of transformer substation site selection, and the method comprises the steps: obtaining multi-source image data and laser radar point cloud of a target area through employing multiple sensors of an unmanned plane; performing semantic segmentation and three-dimensional set geometric feature extraction on the image and the laser radar point cloud through a deep convolutional neural network to obtain spatial environment element information; constructing a multi-factor evaluation index system based on the obtained spatial environment element information; a reinforcement learning framework is adopted to output the weight of each index online, and the weight is updated in a rolling manner according to time; constructing a multi-objective optimization model, and searching a Pareto optimal site set under a constraint condition by using a multi-objective particle swarm and ant colony fusion algorithm; and displaying a Pareto solution set, and providing an interactive weight adjustment and site selection report automatic generation function. Closed-loop adjustment is formed, and the site selection period, the social risk and the whole life cycle cost are shortened.

Description

Technical Field

[0001] This invention relates to the field of substation site selection technology, specifically to a substation site selection method, system, equipment, and medium based on artificial intelligence-based multi-factor evaluation and dynamic weight adjustment. Background Technology

[0002] Currently, with the increasing consumption and demand for energy in daily production and life, the construction of various infrastructures is also in full swing, and substations are one of them. The construction of substations involves many processes and procedures, with site selection being the most crucial and initial step.

[0003] With the continuous progress and development of science and technology, various new technical solutions have gradually emerged, such as the rise of GIS-assisted site selection, UAV remote sensing and deep learning technologies, and subsequently the adoption of UAV-assisted site selection.

[0004] However, when these new technical solutions are used, there are still problems such as insufficient data dimensions, static and rigid weights, weak environmental and social impact assessment, lack of three-dimensional spatial constraints, and low level of intelligent decision-making. Summary of the Invention

[0005] In view of the above-mentioned problems, the present invention provides a substation site selection method, system, equipment and medium based on artificial intelligence multi-factor evaluation and dynamic weight adjustment.

[0006] Therefore, the technical problem solved by this invention is: how existing methods suffer from insufficient data dimensions, static and rigid weights, weak environmental and social impact assessment, lack of three-dimensional spatial constraints, and low level of intelligent decision-making.

[0007] To address the aforementioned technical problems, this invention provides the following technical solution: a substation site selection method based on artificial intelligence with multi-factor evaluation and dynamic weight adjustment, comprising: acquiring multi-source image data and lidar point clouds of the target area using multiple sensors from a UAV; performing semantic segmentation and 3D geometric feature extraction on the images and lidar point clouds using a deep convolutional neural network to obtain spatial environmental element information; constructing a multi-factor evaluation index system based on the obtained spatial environmental element information and generating standardized index vectors; employing a reinforcement learning framework, using real-time electricity price, load growth rate, policy parameters, and public sentiment value as state inputs, outputting the weights of each index online, and updating them over time; constructing a multi-objective optimization model with comprehensive cost, environmental impact, and social acceptance as objectives, and using multi-objective particle swarm optimization and ant colony fusion algorithms to search for the Pareto optimal site set under constraints; displaying the Pareto solution set in a 3D geographic information platform and providing interactive weight adjustment and automatic generation of site selection reports.

[0008] As a preferred embodiment of the substation site selection method based on artificial intelligence multi-factor evaluation and dynamic weight adjustment described in this invention, the method of acquiring multi-source image data and lidar point cloud of the target area using multiple sensors of UAV includes: dynamically generating flight routes that meet spatial coverage requirements and flight time constraints based on a three-dimensional terrain model and real-time no-fly zone data; simultaneously acquiring multi-source remote sensing images and lidar point clouds along the flight routes; and performing aerial triangulation and radiometric correction on the acquired image and lidar point cloud data to obtain a registered remote sensing dataset.

[0009] As a preferred embodiment of the substation site selection method based on artificial intelligence and dynamic weight adjustment described in this invention, the step of performing semantic segmentation and three-dimensional geometric feature extraction on images and lidar point clouds through a deep convolutional neural network to obtain spatial environmental element information includes: constructing a deep convolutional network model specifically for substation site selection, embedding a dual attention module and a multi-scale dilated convolution module in the basic network structure; performing semantic segmentation on the collected image data, outputting classification results containing different land feature categories to obtain a two-dimensional semantic map; extracting three-dimensional geometric feature information of the terrain based on lidar point cloud data; and unifying and fusing the coordinates of the two-dimensional semantic map, three-dimensional geometric features, and thermal infrared temperature data from multi-source remote sensing images to construct a multi-dimensional feature tensor containing spatial, channel, and temporal dimensions to generate spatial environmental element information.

[0010] As a preferred embodiment of the substation site selection method based on artificial intelligence and dynamic weight adjustment described in this invention, the step of constructing a multi-factor evaluation index system and generating a standardized index vector based on the obtained spatial environmental element information includes: constructing a multi-factor evaluation index system based on spatial environmental element information; using the extracted spatial environmental element information as input to calculate the attribute values ​​of each index; and normalizing the attribute values ​​of each index to generate a standardized index vector.

[0011] As a preferred embodiment of the substation site selection method based on artificial intelligence and dynamic weight adjustment described in this invention, the method employs a reinforcement learning framework, using real-time electricity price, load growth rate, policy parameters, and public opinion sentiment as state inputs, and outputs the weights of each indicator online and updates them over time. This includes constructing a reinforcement learning network based on an actor-commentator structure, where the actor network takes a state vector as input and outputs the weights of each evaluation indicator, and the commentator network evaluates the long-term returns of the site selection scheme. The state vector includes real-time electricity price, annual load growth rate, carbon trading price, policy subsidy vector, and public opinion sentiment. The reward function is set as a weighted sum of changes in economic returns, carbon emission reduction, and public opinion improvement. Every 24 hours, an experience replay mechanism and a target network mechanism are used to update the weights of each evaluation indicator online.

[0012] This preferred solution uses reinforcement learning to adaptively adjust the weights of evaluation indicators. It can dynamically optimize the weights based on real-time electricity prices, load growth rates, policy parameters, and changes in public opinion, avoiding biases from manual settings and improving the responsiveness and decision-making accuracy of the site selection model to changes in the external environment.

[0013] As a preferred embodiment of the substation site selection method based on artificial intelligence and dynamic weight adjustment described in this invention, the method involves: constructing a multi-objective optimization model with comprehensive cost, environmental impact, and social acceptance as objectives; and using a fusion algorithm of multi-objective particle swarm optimization and ant colony optimization to search for the Pareto optimal site set under constraints. This includes: constructing a multi-objective function with comprehensive cost, environmental impact, and social acceptance as optimization objectives; establishing a three-objective optimization model; setting hard constraints including avoiding earthquake fault zones, airport airspace, military facilities, and core areas of nature reserves, as well as soft constraints such as distance from residential areas, upper limit of slope, and probability of flooding; solving the problem using a fusion optimization method of multi-objective particle swarm optimization and ant colony optimization, where the particle swarm optimization is responsible for global search and the ant colony optimization performs local fine-grained optimization within the particle neighborhood; performing non-dominated sorting and adaptive congestion distance processing on all candidate solutions; and outputting the Pareto optimal site set and comprehensive score results.

[0014] This preferred solution employs a fusion algorithm of multi-objective particle swarm optimization and ant colony optimization to achieve a balanced optimization solution that considers both global and local factors under the three objectives of comprehensive cost, environmental impact, and social acceptability. This enhances the diversity and stability of Pareto solutions and makes the site selection results more scientific and reasonable.

[0015] As a preferred embodiment of the substation site selection method based on artificial intelligence multi-factor evaluation and dynamic weight adjustment described in this invention, the dual attention module includes a channel attention submodule and a spatial attention submodule; the channel attention submodule generates channel weight vectors by performing global average pooling, fully connected layer calculation, and activation function on the input feature map; the spatial attention submodule generates a spatial weight matrix by concatenating max pooling and average pooling, and using convolution operation and activation function; the channel weights and spatial weights are multiplied element-wise and then fused with the original feature map.

[0016] This invention provides a substation site selection system based on artificial intelligence, which involves multi-factor evaluation and dynamic weight adjustment.

[0017] To address the aforementioned technical problems, this invention provides the following technical solution: a substation site selection system based on artificial intelligence with multi-factor evaluation and dynamic weight adjustment, comprising: a multi-source data acquisition module, a feature extraction module, an indicator system construction module, a weight update module, a multi-objective optimization module, and a display module; the multi-source data acquisition module is used to acquire multi-source image data and lidar point clouds of the target area using multiple sensors of a UAV; the feature extraction module is used to perform semantic segmentation and three-dimensional geometric feature extraction on the images and lidar point clouds through a deep convolutional neural network to obtain spatial environmental element information; the indicator system construction module is used to... The system utilizes environmental element information to construct a multi-factor evaluation index system and generate standardized index vectors. The weight update module employs a reinforcement learning framework, using real-time electricity prices, load growth rates, policy parameters, and public sentiment values ​​as state inputs, to output the weights of each index online and update them over time. The multi-objective optimization module constructs a multi-objective optimization model with comprehensive cost, environmental impact, and social acceptance as objectives, and uses a multi-objective particle swarm optimization and ant colony fusion algorithm to search for the Pareto optimal site set under constraints. The display module displays the Pareto solution set in a 3D geographic information platform and provides interactive weight adjustment and automatic site selection report generation functions.

[0018] The present invention provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the substation site selection method based on artificial intelligence multi-factor evaluation and dynamic weight adjustment.

[0019] The present invention provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the steps of the substation site selection method based on artificial intelligence multi-factor evaluation and dynamic weight adjustment are implemented.

[0020] The beneficial effects of this invention are as follows: Based on high-resolution data from multiple sensors of UAVs, this invention utilizes deep learning to achieve centimeter-level recognition and three-dimensional quantization of elements such as terrain, landforms, vegetation, buildings, and sensitive areas.

[0021] We construct a multi-factor indicator system covering four dimensions: technology, economy, environment, and society, and design a dynamic weight adjustment mechanism based on reinforcement learning to achieve real-time response to policies, electricity prices, and public opinion.

[0022] The three-dimensional spatial constraints, soft social indicators, and hard legal constraints are all incorporated into the multi-objective optimization model, and the MOPSO-ACO fusion algorithm is used to solve for the Pareto optimal site set.

[0023] This forms a closed-loop system encompassing data acquisition, feature extraction, dynamic evaluation, intelligent optimization, 3D visualization, and interactive decision-making, significantly shortening the site selection cycle and reducing social risks and life-cycle costs. Attached Figure Description

[0024] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 The overall flowchart of a substation site selection method based on artificial intelligence and dynamic weight adjustment is provided in one embodiment of the present invention. Detailed Implementation

[0026] To make the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0027] Example 1, referring to Figure 1 This is one embodiment of the present invention, which provides a substation site selection method based on artificial intelligence multi-factor evaluation and dynamic weight adjustment, including: S1. Utilize multiple sensors of the UAV to acquire multi-source image data and lidar point cloud of the target area.

[0028] S2. Semantic segmentation and 3D geometric feature extraction of images and LiDAR point clouds are performed using deep convolutional neural networks to obtain spatial environment element information.

[0029] S3. Based on the obtained spatial environment element information, construct a multi-factor evaluation index system and generate standardized index vectors.

[0030] S4. A reinforcement learning framework is adopted, with real-time electricity price, load growth rate, policy parameters and public opinion sentiment value as state inputs, and the weights of each indicator are output online and updated over time.

[0031] S5. With comprehensive cost, environmental impact and social acceptance as objectives, a multi-objective optimization model is constructed, and a Pareto optimal site set is searched under constraints using a multi-objective particle swarm optimization and ant colony fusion algorithm.

[0032] S6. Display the Pareto solution set in the 3D geographic information platform and provide interactive weight adjustment and automatic generation of site selection reports.

[0033] This invention integrates remote sensing, semantic recognition, weighted adaptive learning, and multi-objective optimization methods to achieve comprehensive quantitative evaluation and scheme selection of candidate substation sites. It has good adaptability in dealing with complex scenarios and multi-source heterogeneous data, and can provide intelligent and interpretable site selection decision support for power grid planning.

[0034] Example 2, an embodiment of the present invention, provides a substation site selection method based on artificial intelligence multi-factor evaluation and dynamic weight adjustment, based on the previous embodiment, including: In this embodiment of the application, the multi-source image data in step S1 can be visible light images, multispectral images and thermal infrared images acquired synchronously by multi-sensor devices carried by the UAV. The visible light images have a resolution better than 2 cm, the multispectral images contain no less than 6 bands, and the thermal infrared images are used to reflect the surface temperature distribution. All three types of data have undergone aerial triangulation and radiometric correction processing, and have uniform georegistration accuracy, which is used to support surface feature identification and thermal anomaly analysis.

[0035] In one alternative implementation, multi-source image data can also be obtained by combining multi-temporal remote sensing images acquired by remote sensing satellites with supplementary ground-based images to improve coverage over a wide area and serve as an auxiliary data source for UAV aerial surveys.

[0036] In another alternative implementation, multi-source image data can also be fused from datasets of synthetic aperture radar (SAR) images and airborne infrared scan images to enhance target recognition stability and imaging integrity under complex terrain and weather interference.

[0037] This invention can effectively support the accurate identification and classification of surface elements such as topography, vegetation, water bodies, and buildings. In particular, based on the fusion of multi-source heterogeneous data, it helps to improve the accuracy of subsequent semantic segmentation, enhance the detection capability of small targets, and provide comprehensive and reliable spatial data support for the quantification of multi-factor evaluation indicators.

[0038] Furthermore, in step S1, the UAV's multi-sensor system acquires multi-source image data and lidar point cloud data of the target area, including the following steps A1-A3: A1. Based on the three-dimensional terrain model and real-time no-fly zone data, dynamically generate flight routes that meet spatial coverage requirements and flight time constraints.

[0039] A2. Simultaneously acquire multi-source remote sensing images and lidar point clouds along the flight path.

[0040] A3. Perform aerial triangulation and radiometric correction on the acquired imagery and lidar point cloud data to obtain a registered remote sensing dataset.

[0041] Specifically, in step A1, based on the three-dimensional terrain model and real-time no-fly zone data, an improved genetic algorithm is used to generate a flight route that balances spatial coverage of ≥98% and the shortest flight time.

[0042] Specifically, in step A2, four types of data are acquired simultaneously: red, green, and blue primary color models (Red, Green, Blue, RGB), multispectral images, thermal infrared images, and LiDAR (Light Detection and Ranging) point clouds. The RGB ground resolution is ≤2cm, the LiDAR point cloud density is ≥50pts / m², and there are ≥6 multispectral bands.

[0043] Specifically, in step A3, the aerial triangulation is completed by fusing data from the Position and Orientation System (POS) / Inertial Measurement Unit (IMU), and the root mean square (RMS) of the aerial triangulation is ≤0.5 pixels.

[0044] In this embodiment, the spatial environmental element information in step S2 can be obtained by using a deep convolutional neural network to perform semantic segmentation and geometric feature extraction on remote sensing images and lidar point clouds to obtain spatial distribution information of terrain, landforms, buildings, roads, vegetation, water bodies and environmentally sensitive areas. The expression form is a two-dimensional semantic map and a three-dimensional point cloud model with a unified coordinate system. After fusing the thermal infrared temperature map, a multi-dimensional feature tensor containing spatial dimension, channel dimension and time dimension is constructed to support subsequent multi-factor evaluation modeling.

[0045] In one alternative implementation, spatial environmental element information can also be used to generate elevation distribution maps, slope maps, and view analysis maps by fusing multispectral images and digital surface models, which can be used to assist in assessing construction difficulty, geological stability, and visual landscape sensitivity.

[0046] In another alternative implementation, spatial environmental element information can also be based on vector layers of building boundary lines, road center lines, green space distribution belts, and thermal anomaly areas automatically extracted from lidar point clouds, used to describe the functional area structure and environmental constraint pattern around the candidate site.

[0047] The spatial environmental element information constructed by this invention has the ability to fuse semantic, geometric and thermal multimodal data, and can accurately characterize the natural attributes and human intervention status of the site selection area. It provides high-resolution and structured spatial basic information support for the comprehensive construction of evaluation indicators, the fine setting of constraints and the spatial adaptability of optimization models.

[0048] Furthermore, in step S2, a deep convolutional neural network is used to perform semantic segmentation and 3D geometric feature extraction on the image and lidar point cloud to obtain spatial environment element information, including the following steps B1-B4: B1. Construct a deep convolutional network model specifically for substation site selection, embedding dual attention modules and multi-scale dilated convolutional modules into the basic network structure.

[0049] B2. Perform semantic segmentation on the collected image data, output classification results containing different land feature categories, and obtain a two-dimensional semantic map.

[0050] B3. Extracting three-dimensional geometric features of terrain based on lidar point cloud data.

[0051] B4. Unify and fuse the coordinates of the two-dimensional semantic map, three-dimensional geometric features and thermal infrared temperature data in multi-source remote sensing images to construct a multi-dimensional feature tensor containing spatial, channel and temporal dimensions, and generate spatial environmental element information.

[0052] Specifically, in step B1, a dedicated deep convolutional network (Self-Supervised UNet++, SS-UNet++) for substation site selection is constructed. This network embeds a dual-attention module and a multi-scale dilated convolutional module into the UNet++ backbone to improve the accuracy of small target and boundary recognition. The deep convolutional neural network uses the SS-UNet++ structure, which embeds a dual-channel attention mechanism and a multi-scale dilated convolutional module into the backbone network to achieve high-precision segmentation of complex terrain and building boundaries, with an intersection-over-union (IoU) ratio ≥ 92% for the segmentation results.

[0053] The dual attention module includes a channel attention submodule and a spatial attention submodule. The channel attention submodule generates channel weight vectors by performing global average pooling on the input feature map, calculating fully connected layers, and using activation functions. The spatial attention submodule generates a spatial weight matrix by concatenating max pooling and average pooling, and using convolution operations and activation functions. The channel weights and spatial weights are then multiplied element-wise and fused with the original feature map.

[0054] Specifically, the channel attention submodule performs global average pooling → fully connected layer → Sigmoid activation on the input feature map to obtain the channel weight vector; the spatial attention submodule performs max pooling and average pooling concatenation → 7×7 convolution → Sigmoid activation on the input feature map to obtain the spatial weight matrix; the channel weights and spatial weights are multiplied element-wise and then multiplied with the original feature map to achieve adaptive feature enhancement.

[0055] In this embodiment of the application, the classification result in step B2 can be a two-dimensional semantic map output by the constructed deep convolutional neural network after performing semantic segmentation on the aerial survey image. According to the preset scene classification system, the region is divided into 12 categories: vegetation, farmland, water body, road, building, education, medical, cultural relics, industry, storage, open space, and others. Each pixel corresponds to a category label and confidence value, forming a multi-channel image output result.

[0056] In an alternative implementation, the classification results can also combine the semantic label map with the 3D laser point cloud data, and the grouping results of ground point class, vertical structure point class and edge feature point class generated by spatial consistency correction can be used to improve the boundary recognition accuracy and structural expression ability.

[0057] In another alternative implementation, the classification results can also be used to help identify heat source facilities, energy channels and unsuitable site areas by generating a temperature anomaly distribution map after performing temperature partitioning and clustering analysis on thermal infrared images.

[0058] The classification results obtained by this invention have the advantages of high accuracy, high confidence and spatial consistency. It can realize pixel-level recognition and layer generation of multiple types of surface elements. Especially in complex scenes, it maintains clear boundaries and good detail preservation, providing a stable and accurate data foundation for subsequent index quantification, constraint extraction and site selection evaluation, and significantly improving the automation and intelligence level of the overall evaluation process.

[0059] Specifically, in step B2, the scene is divided into 12 categories: vegetation, farmland, water bodies, roads, buildings, education, medical care, cultural relics, industry, warehousing, open space, and others, with an intersection-union ratio (IoU) of ≥92%.

[0060] In this embodiment of the application, the three-dimensional geometric feature information in step B3 can be spatial geometric attributes such as slope, aspect, building height, surface roughness and thermal anomaly area calculated based on lidar point cloud data. These features have a clear location and shape in geographic space and are used to reflect the topographic relief, structural features and environmental changes of the site area.

[0061] In one alternative implementation, the three-dimensional geometric feature information can also be used to analyze the constructability space, construction convenience, and structural integrity by extracting ground feature height difference maps, building edge extraction results, and cross-sectional views after constructing digital surface models and digital elevation models.

[0062] In another optional implementation, the three-dimensional geometric feature information can also be combined with contour maps, cross-sectional change maps and terrain slope heat maps obtained by data fusion of multi-view stereo images and lidar point clouds, which are used to describe the trend of large-scale geomorphic feature changes and the complexity of the terrain around the candidate site.

[0063] The three-dimensional geometric feature information extracted by this invention has the characteristics of high precision, high density and complete spatial representation. It can truly reflect the geometric morphological changes of the target area in three-dimensional space, effectively support the assessment of construction difficulty, analysis of route corridors and identification of environmental impacts, provide a high-quality geometric input basis for the construction of multi-factor indicators and site optimization, and improve the engineering adaptability and risk identification capability of the overall site selection decision.

[0064] Specifically, in step B3, the slope, aspect, building height, surface roughness, and area of ​​thermal anomaly region are calculated based on LiDAR point cloud computing.

[0065] Specifically, in step B4, the two-dimensional semantic graph, the three-dimensional geometric model, and the thermal infrared temperature map are unified into the UTM coordinate system to form a 4D tensor. ,in, To represent the number of pixels in the vertical direction of the data after unification to the UTM coordinate system, To unify the number of horizontal pixels in the data after converting to the UTM coordinate system, The number of feature channels, Let T be the set of real numbers, and let T be the time dimension.

[0066] Specifically, the steps for semantic segmentation and 3D feature extraction of images and point clouds using deep convolutional networks to obtain information on terrain, landforms, vegetation, buildings, roads, and environmentally sensitive areas are as follows: A deep convolutional network SS-UNet++ suitable for substation site selection is constructed. A dual-attention module and a multi-scale dilated convolutional module are embedded on top of UNet++ to enhance the accuracy of small target recognition and boundary determination. The aerial survey scene is divided into 12 categories: vegetation, farmland, water bodies, roads, buildings, education, medical facilities, cultural relics, industry, warehousing, open space, and others. Then, a semantic label map is output with an IoU ≥ 92% model accuracy requirement, where each pixel corresponds to a category label and a confidence matrix. Based on LiDAR point computing, slope, aspect, building height, surface roughness, and the area of ​​thermal anomaly regions are calculated. The system generates elevation and slope distribution maps to provide a basis for terrain feasibility for subsequent site selection. It unifies the two-dimensional semantic segmentation results, three-dimensional point cloud features, and thermal infrared temperature maps into the UTM coordinate system to form a 4D tensor F containing spatial dimension (H×W), channel dimension (C), and time dimension (T), achieving multimodal feature alignment of image, point cloud, and thermal infrared. After morphological filtering and connected component analysis, the fusion results output vector layers of terrain, landforms, vegetation, buildings, roads, and environmentally sensitive areas. Each layer is quantified into attributes such as area, minimum distance, slope range, and surface temperature, providing input data for subsequent multi-factor evaluation.

[0067] Furthermore, in step S3, based on the obtained spatial environment element information, a multi-factor evaluation index system is constructed, and a standardized index vector is generated, including the following steps C1-C3: C1. Construct a multi-factor evaluation index system based on spatial environmental element information.

[0068] C2. Using the extracted spatial environment element information as input, calculate the attribute values ​​of each indicator.

[0069] C3. Normalize the attribute values ​​of each indicator to generate a standardized indicator vector.

[0070] Specifically, the technical indicators include: geological stability, line corridor length, ease of access to and from the main grid, construction difficulty, distance from the main electrical grid, and short-circuit capacity margin.

[0071] Economic indicators: land acquisition cost, demolition compensation cost, crop compensation cost, construction access road cost, operation and maintenance cost, transmission loss revenue, and total life cycle NPV.

[0072] Environmental indicators: vegetation destruction, carbon emission increment, area exceeding noise standards, water encroachment, landscape visual sensitivity, and area in conflict with ecological red lines.

[0073] Social indicators: minimum distance from schools / hospitals / cultural relics / residential areas, probability of resident complaints, policy incentives, public sentiment, and landscape harmony.

[0074] Quantification Interface: Each indicator is automatically calculated using a geospatial analysis script based on the spatial / semantic / geometric / thermal features extracted in step S2, and written into the indicator vector in real time. ,in, For index vectors, As the first indicator, For the first One indicator, This refers to the number of indicators.

[0075] Demolition compensation is calculated using the following formula: in, For demolition compensation, For the first Building volume The unit price is the compensation price per unit volume. This is a floor correction factor, with a value ranging from 0.7 to 1.3. This refers to the number of buildings.

[0076] Carbon emission increments are calculated using the Life Cycle Assessment (LCA) approach: in, Including carbon emissions from construction, operation, and demolition; To calculate carbon emissions during operation, the calculation is: annual grid loss × regional power grid carbon emission factor × 30 years; Carbon emissions from dismantling are estimated at 15% of those from construction. To build carbon emissions.

[0077] The public opinion sentiment value is obtained through the following sub-steps: using web crawlers to crawl social media texts containing the keyword "substation"; using the BERT-BiLSTM hybrid model to classify the sentiment polarity of the text and output the sentiment value in the range of [-1,1]; and taking the exponentially weighted moving average of the sentiment values ​​over 90 consecutive days as the public opinion sentiment value for the day.

[0078] Furthermore, in step S4, a reinforcement learning framework is employed, using real-time electricity price, load growth rate, policy parameters, and public sentiment value as state inputs. The weights of each indicator are output online and updated over time, including the following steps D1-D4: D1. Construct a reinforcement learning network based on an actor-commentator structure, where the actor network takes the state vector as input and outputs the weights of each evaluation index, and the commentator network evaluates the long-term reward of the site selection scheme.

[0079] D2. The state vector includes real-time electricity price, annual load growth rate, carbon trading price, policy subsidy vector, and public sentiment value.

[0080] D3. The reward function is set as a weighted sum of changes in economic returns, changes in carbon emission reduction, and improvements in public opinion.

[0081] D4. The weights of each evaluation indicator are updated online every 24 hours using an experience playback mechanism and a target network mechanism.

[0082] Specifically, in step D1, an actor-critic deep reinforcement learning framework is constructed, where: the actor network... by state As input, output the weights of each indicator. Critic Network Evaluate the site selection options in status Long-term returns.

[0083] Specifically, the reward function in step D3 is expressed as: in, ,and Adaptive adjustment This is the economic benefit coefficient. This is the carbon emission reduction factor. This is the public opinion improvement coefficient. For the reward function; This refers to the change in economic returns, specifically the increase in economic benefits compared to the previous period. This refers to the change in carbon emission reduction, that is, the change in emission reduction in the current period compared to the baseline period. This is a change in public sentiment value, used to reflect the degree of improvement in public attitudes toward the project.

[0084] Adaptive adjustment coefficient The update rules are as follows: in, The learning rate, ranging from 0.01 to 0.05; The economic return coefficient before the update. The updated economic return coefficient, As the benchmark return, The rate of change in carbon trading prices. The carbon emission reduction factor before the update. For the updated carbon emission reduction factor, This represents the updated public opinion improvement coefficient.

[0085] Furthermore, in step S5, a multi-objective optimization model is constructed with comprehensive cost, environmental impact, and social acceptance as objectives. A multi-objective particle swarm optimization and ant colony optimization algorithm is then used to search for the Pareto optimal site set under constraints, including the following steps E1-E4: E1. Construct a multi-objective function with comprehensive cost, environmental impact and social acceptance as optimization objectives, and establish a three-objective optimization model.

[0086] E2. Set hard constraints including avoiding earthquake fault zones, airport airspace, military facilities and core areas of nature reserves, as well as soft constraints such as distance from residential areas, maximum slope and probability of flooding.

[0087] E3. An optimization method combining multi-objective particle swarm optimization and ant colony optimization is used to solve the problem. The particle swarm optimization is responsible for global search, while the ant colony optimization performs local fine-tuning within the particle neighborhood.

[0088] E4. Perform non-dominated sorting and adaptive congestion distance processing on all candidate solutions, and output the Pareto optimal site set and comprehensive score results.

[0089] Specifically, in step E1, a three-objective optimization model is constructed with comprehensive cost, environmental impact, and social acceptance as objectives: in, For overall cost, For environmental impact, For social acceptance.

[0090] Specifically, in step E2, hard constraints include avoiding earthquake fault zones, airport airspace, military facilities, and the core area of ​​national nature reserves.

[0091] Soft constraints include distance from residential areas ≥ minimum safe distance threshold dmin, slope < upper limit threshold θmax, and flood inundation probability < 1 / 100. These represent the safety requirements of the site selection area under a once-in-a-century flood and are used to limit hydrological risks.

[0092] If a candidate site falls into the hard constraint region, a maximum penalty value M=10 is assigned. 6 If a candidate site violates a soft constraint, it is directly eliminated; if the candidate site violates the soft constraint, the constraint violation degree is calculated. and treat it as a secondary penalty item Add the objective function, where It increases exponentially with the number of iterations.

[0093] Specifically, in step E3, a fusion algorithm of Multi-Objective Particle Swarm Optimization-Ant Colony Optimization (MOPSO-ACO) is proposed, in which the outer particle swarm is responsible for global search, and the inner ant colony performs fine optimization in the particle neighborhood.

[0094] Set particle swarm size Ant colony size Maximum number of iterations Randomly generate candidate site location vectors and initialize the velocity vector. pheromone matrix and heuristic functions .

[0095] In each iteration, based on the individual optimal solution and the group optimal solution Update particle velocity and position: in, For inertial weights, , As a learning factor, , A random number between 0 and 1 For the first Each particle at time The velocity vector, For the first Each particle at time The velocity vector, For the first Each particle at time The position vector.

[0096] Within the neighborhood solution set obtained from particle swarm optimization, ant colony optimization is used for local optimization. Ants select paths based on transition probabilities. in, For ants from nodes Transfer to node The probability of; For nodes To the node The intensity of pheromones between them reflects the priority of this path being chosen by historical experience. For nodes To the node The heuristic value is related to path distance, cost, or benefit; The pheromone importance coefficient; For the importance coefficient of the heuristic function, For the current ant from node After departure, any node in the set of all reachable candidate nodes.

[0097] Once a feasible path is completed, updates are made according to the pheromone increment rule: in, The evaporation coefficient is... This is to increase the amount of pheromones.

[0098] Calculate the comprehensive cost for all candidate solutions. Environmental impact Social acceptance And form the target vector as follows: in, The target vector.

[0099] Next, an adaptive crowding distance strategy is used to sort and filter the non-dominated solution set in order to maintain the uniformity and diversity of the Pareto front solution distribution; If the current iteration count has not been reached If the Pareto solution set is not yet stable, continue with steps S2 to S4; if the convergence condition is met, output the Pareto optimal site set. And its overall score.

[0100] The specific process includes: initializing the particle swarm, where each particle's position represents the coordinates and shape of a candidate site; calculating the objective function value of each particle and updating the individual optimal and global optimal values; and starting the ant colony algorithm within the neighborhood of the globally optimal particle, using a rasterized neighborhood as the graph and initializing the pheromone value. ,in, This is the initial value of the pheromone. The number of objectives for multi-objective optimization. The maximum value among all objective function values ​​is used; after the ant colony iteration ends, the optimal solution is fed back to the particle swarm to achieve information collaboration; an adaptive crowding distance is used to maintain the diversity of the Pareto front, and the crowding distance threshold decreases linearly with the number of iterations.

[0101] Furthermore, in step S6, the Pareto solution set is displayed in the 3D geographic information platform, and interactive weight adjustment and automatic site selection report generation functions are provided, including the following steps: Load DOM, DSM, architectural white model, and route corridor onto the Cesium 3D Earth platform to enable roaming, sectioning, and solar radiation analysis; provide interactive weight sliders that allow users to drag and the system to instantly reorder candidate sites; automatically generate PDF site selection reports containing indicator values, weights, sensitivity analysis, and risk matrices.

[0102] Specifically, the front end uses React+D3.js to render the slider and real-time radar chart; the back end calls the Python Flask service through a RESTful API and recalculates the sorting using the cached Pareto solution set; the response time is <500ms and it supports ≥100 concurrent users.

[0103] Example 3 is an embodiment of the present invention, which provides a substation site selection method based on artificial intelligence multi-factor evaluation and dynamic weight adjustment. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through experiments.

[0104] Platform: DJI M300RTK, CERNET PSDK102S five-lens camera (2.1cm GSD), Velodyne VLP-32CLiDAR.

[0105] The flight routes were generated using a genetic algorithm, resulting in a double-layer grid with 80% overlap in heading and 75% overlap in lateral directions. A total of 38 flights were conducted, covering an area of ​​48 km².

[0106] After POS / IMU fusion, the horizontal accuracy is ±2cm and the vertical accuracy is ±3cm.

[0107] Data volume: 117,000 original photos, 2.4TB point cloud.

[0108] Preprocessing: Aerial triangulation: ContextCapture 3D modeling, aerial triangulation RMS 0.48 pixels; Orthophoto DOM: 2cm resolution, GeoTIFF format.

[0109] Point cloud classification uses the SS-UNet++ network of this invention (PyTorch 1.13, RTX 4090×2 training, input 512×512×4 channels), and the IoU of 12 land cover classes is shown in Table 1.

[0110] Table 1. IoU Table for 12 Types of Land Features

[0111] The indicator system is quantified based on DOM / DSM / semantic segmentation results. The ArcPy+GDAL script automatically calculates 18 indicators (examples) as shown in Table 2.

[0112] Table 2 18 Indicators

[0113] Reinforcement learning dynamic weight training: Real-time electricity price: The average marginal electricity price in the provincial electricity spot market over the previous 30 days.

[0114] Load growth rate: The EMS system's rolling forecast of GDP elasticity coefficient for the next three years.

[0115] Policy Vector: Extract the texts of government-issued policies on carbon subsidies, green corridors, and NIMBY compensation (TF-IDF → 32-dimensional vector).

[0116] Sentiment score: Crawled posts containing the keyword "substation" from Sina Weibo and local forums over the past 90 days, and scored using SnowNLP sentiment scores [-1, 1].

[0117] Multi-objective optimization solution: The coding uses the center point of a 100m×100m grid as the decision variable, resulting in a total of 12,480 feasible grids.

[0118] Objective function: .

[0119] Hard constraints: avoid earthquake fault zones (>500m), airport airspace (>2km), and the core area of ​​provincial wetlands.

[0120] Soft constraints: distance from residential areas ≥150m, slope <10°, probability of flooding <1 / 100.

[0121] Algorithm parameters: MOPSO particle count 100, maximum iterations 150; ACO pheromone heuristic factor. , Evaporation coefficient .

[0122] Operating environment: Intel Xeon Gold 6248R×2, 128GB RAM, time taken: 2 hours and 17 minutes.

[0123] The output yields 25 non-dominated solutions to the Pareto front, with the top 5 combined as shown in Table 3.

[0124] Table 3 Comprehensive Scoring Table

[0125] Results verification and implementation: Simultaneous verification with the feasibility study unit: The demolition area within the P3 site boundary is only 0.42 million m² (reduced by 89% compared to the original plan), and the length of the 110kV cable outgoing line is shortened by 1.8km; Environmental impact assessment passed on the first attempt: The predicted noise level is 48dB during the day and 41dB at night, meeting the requirements of Class 2 area of ​​the "Environmental Quality Standard for Noise"; Approval cycle: Only 8 months from the start of site selection to obtaining approval, a 55% reduction compared to the same period last year; Investment savings: Static investment reduced by 31 million yuan, and the total life cycle cost NPV reduced by 220 million yuan; Social benefits: Carbon emission intensity decreased by 18%, and there were 0 resident complaints.

[0126] To verify the advantages of this invention over the traditional static weighted method (AHP fixed weight), a unified data source was selected for evaluation within the same planning area: aerial DOM, DSM, LiDAR point cloud, thermal infrared, basic geographic and control zones, with a unified coordinate system, consistent spatial resolution and time window to ensure consistency in data collection.

[0127] The baseline method is static AHP with fixed weights (frozen after expert scoring and not updated over time).

[0128] This invention utilizes dynamic weighting (Actor–Critic), MOPSO+ACO multi-objective search, and crowd distance to preserve diversity.

[0129] Ablation settings: Remove dynamic weights (replace with fixed weights), everything else is the same; remove ACO local fine search; remove crowding distance (only for non-dominated sorting).

[0130] Evaluation indicators (all statistics are based on Top-K schemes, with K being 5 or 10) include: environmental impact assessment pass rate (%); approval cycle (months) – from project initiation and site selection to obtaining approval documents; total investment NPV (in RMB 100 million, with a uniform discount rate and term); demolition area (10,000 m²); total outgoing line length (km); public negative sentiment index (0–100, the lower the better); and number of complaints after operation (cases / period).

[0131] The improvement rate is calculated as follows: Improvement % = (Invention - Baseline) / Baseline × 100% (for indicators where "higher is better," such as pass rate); Decrease % = (Baseline - Invention) / Baseline × 100% (for indicators where "lower is better," such as cycle time, NPV, demolition area, etc.). The results are shown in Table 4.

[0132] Table 4 Comparison Results

[0133] Example 4 is an embodiment of the present invention. This embodiment provides a substation site selection system based on artificial intelligence for multi-factor evaluation and dynamic weight adjustment, including a multi-source data acquisition module, a feature extraction module, an indicator system construction module, a weight update module, a multi-objective optimization module, and a display module.

[0134] The multi-source data acquisition module is used to acquire multi-source image data and lidar point clouds of the target area using multiple sensors of the UAV.

[0135] The feature extraction module is used to perform semantic segmentation and 3D geometric feature extraction on images and LiDAR point clouds through deep convolutional neural networks to obtain spatial environment element information.

[0136] The indicator system construction module is used to construct a multi-factor evaluation indicator system based on the obtained spatial environment element information and generate standardized indicator vectors.

[0137] The weight update module uses a reinforcement learning framework to take real-time electricity price, load growth rate, policy parameters and public opinion sentiment as state inputs, outputs the weights of each indicator online, and updates them over time.

[0138] The multi-objective optimization module is used to construct a multi-objective optimization model with the objectives of comprehensive cost, environmental impact and social acceptance, and to search for the Pareto optimal site set under constraints using a multi-objective particle swarm optimization and ant colony fusion algorithm.

[0139] The display module is used to display Pareto solutions in a 3D geographic information platform and provides interactive weight adjustment and automatic generation of site selection reports.

[0140] This embodiment also provides an electronic device applicable to the substation site selection method based on artificial intelligence multi-factor evaluation and dynamic weight adjustment, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the substation site selection method based on artificial intelligence multi-factor evaluation and dynamic weight adjustment as proposed in the above embodiment.

[0141] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the substation site selection method based on artificial intelligence and dynamic weight adjustment as proposed in the above embodiments.

[0142] The storage medium proposed in this embodiment and the substation site selection method based on artificial intelligence for multi-factor evaluation and dynamic weight adjustment proposed in the above embodiments belong to the same inventive concept. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0143] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0144] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A substation site selection method based on multi-factor evaluation and dynamic weight adjustment using artificial intelligence, characterized by: include, Utilize multiple sensors of drones to acquire multi-source image data and lidar point clouds of the target area; By using deep convolutional neural networks to perform semantic segmentation and 3D geometric feature extraction on images and LiDAR point clouds, spatial environment element information is obtained. Based on the obtained spatial environment element information, a multi-factor evaluation index system is constructed, and a standardized index vector is generated. Using a reinforcement learning framework, real-time electricity prices, load growth rates, policy parameters, and public sentiment values ​​are used as state inputs. The weights of each indicator are output online and updated over time. With the goals of comprehensive cost, environmental impact and social acceptance, a multi-objective optimization model is constructed, and a multi-objective particle swarm optimization and ant colony fusion algorithm is used to search for the Pareto optimal site set under constraints. Display Pareto solutions in a 3D geographic information platform and provide interactive weight adjustment and automatic generation of site selection reports.

2. The substation site selection method based on artificial intelligence multi-factor evaluation and dynamic weight adjustment as described in claim 1, characterized in that: The method of acquiring multi-source image data and lidar point cloud of the target area using multiple sensors of a drone includes, Based on the 3D terrain model and real-time no-fly zone data, flight routes that meet spatial coverage requirements and flight time constraints are dynamically generated. Simultaneously acquire multi-source remote sensing images and lidar point clouds along the flight path; Aerial triangulation and radiometric correction were performed on the acquired imagery and lidar point cloud data to obtain a registered remote sensing dataset.

3. The substation site selection method based on artificial intelligence multi-factor evaluation and dynamic weight adjustment as described in claim 2, characterized in that: The process involves semantic segmentation and 3D geometric feature extraction of images and LiDAR point clouds using a deep convolutional neural network to obtain spatial environment element information, including: A deep convolutional network model specifically for substation site selection is constructed, embedding dual attention modules and multi-scale dilated convolutional modules into the basic network structure; Semantic segmentation is performed on the collected image data, and classification results containing different land feature categories are output to obtain a two-dimensional semantic map; Extracting 3D geometric features of terrain based on lidar point cloud data; By unifying and fusing coordinates of two-dimensional semantic graphs, three-dimensional geometric features, and thermal infrared temperature data from multi-source remote sensing images, a multi-dimensional feature tensor containing spatial, channel, and temporal dimensions is constructed to generate spatial environmental element information.

4. The substation site selection method based on artificial intelligence multi-factor evaluation and dynamic weight adjustment as described in claim 3, characterized in that: Based on the obtained spatial environment element information, a multi-factor evaluation index system is constructed, and a standardized index vector is generated, including: Based on spatial environmental element information, a multi-factor evaluation index system is constructed; The extracted spatial environment element information is used as input to calculate the attribute values ​​of each indicator. The attribute values ​​of each indicator are normalized to generate a standardized indicator vector.

5. The substation site selection method based on artificial intelligence multi-factor evaluation and dynamic weight adjustment as described in claim 4, characterized in that: The aforementioned approach employs a reinforcement learning framework, using real-time electricity prices, load growth rates, policy parameters, and public sentiment values ​​as state inputs. It outputs the weights of each indicator online and updates them over time. include, Construct a reinforcement learning network based on an actor-commentator structure, where the actor network takes the state vector as input and outputs the weights of each evaluation index, and the commentator network evaluates the long-term reward of the site selection scheme. The state vector includes real-time electricity price, annual load growth rate, carbon trading price, policy subsidy vector, and public sentiment value; The reward function is set as a weighted sum of changes in economic gains, changes in carbon emission reductions, and improvements in public opinion; The weights of each evaluation indicator are updated online every 24 hours using an experience playback mechanism and a target network mechanism.

6. The substation site selection method based on artificial intelligence multi-factor evaluation and dynamic weight adjustment as described in claim 5, characterized in that: The aforementioned multi-objective optimization model, with comprehensive cost, environmental impact, and social acceptance as objectives, constructs a multi-objective optimization model and utilizes a multi-objective particle swarm optimization and ant colony optimization algorithm to search for a Pareto optimal site set under constraints, including: Construct a multi-objective function with comprehensive cost, environmental impact and social acceptance as optimization objectives, and establish a three-objective optimization model; The system includes hard constraints such as avoiding earthquake fault zones, airport airspace, military facilities and core areas of nature reserves, as well as soft constraints such as distance from residential areas, maximum slope and probability of flooding. The solution is obtained by using an optimization method that combines multi-objective particle swarm optimization and ant colony optimization. The particle swarm optimization is responsible for global search, while the ant colony optimization is used for local fine-tuning within the particle neighborhood. All candidate solutions are subjected to non-dominated sorting and adaptive congestion distance processing, and the Pareto optimal site set and comprehensive score results are output.

7. The substation site selection method based on artificial intelligence multi-factor evaluation and dynamic weight adjustment as described in claim 6, characterized in that: The dual attention module includes a channel attention submodule and a spatial attention submodule; The channel attention submodule generates channel weight vectors by performing global average pooling on the input feature map, calculating fully connected layers, and using activation functions. The spatial attention submodule generates a spatial weight matrix by concatenating max pooling and average pooling, and using convolution operations and activation functions. The channel weights and spatial weights are multiplied element-wise and then fused with the original feature map.

8. A substation site selection system based on artificial intelligence multi-factor evaluation and dynamic weight adjustment, employing the substation site selection method based on artificial intelligence multi-factor evaluation and dynamic weight adjustment as described in any one of claims 1 to 7, characterized in that, include: Multi-source data acquisition module, feature extraction module, indicator system construction module, weight update module, multi-objective optimization module, and display module; The multi-source data acquisition module is used to acquire multi-source image data and lidar point cloud of the target area using multiple sensors of the UAV. The feature extraction module is used to perform semantic segmentation and three-dimensional geometric feature extraction on images and lidar point clouds through a deep convolutional neural network to obtain spatial environment element information. The indicator system construction module is used to construct a multi-factor evaluation indicator system based on the obtained spatial environment element information and generate standardized indicator vectors. The weight update module is used to employ a reinforcement learning framework, taking real-time electricity price, load growth rate, policy parameters and public opinion sentiment value as state inputs, outputting the weights of each indicator online, and updating them on a rolling basis over time. The multi-objective optimization module is used to construct a multi-objective optimization model with the objectives of comprehensive cost, environmental impact and social acceptance, and to search for the Pareto optimal site set under constraints using a multi-objective particle swarm and ant colony fusion algorithm. The display module is used to display Pareto solutions in a 3D geographic information platform and provides interactive weight adjustment and automatic generation of site selection reports.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the substation site selection method based on artificial intelligence and dynamic weight adjustment as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the substation site selection method based on artificial intelligence and dynamic weight adjustment as described in any one of claims 1 to 7.

Images