Substation site selection intelligent method based on multi-dimensional geographic information and power grid scene constraint

Abstract

The application discloses a substation site selection intelligent method based on multidimensional geographic information and power grid scene constraints, relates to the technical field of substation site selection, and comprises knowledge graph construction, acquisition of feasible site selection, address constraint solving and closed-loop simulation optimization. The application firstly constructs a unified space-time reference geographic power grid space-time knowledge graph; secondly, a multi-modal large model is used to analyze the knowledge graph, automatically identify prohibited construction areas and limited construction areas, and acquire candidate site feasible areas through a self-adaptive spatial clustering algorithm; thirdly, a three-level hierarchical constraint system is constructed, and the feasible areas and the constraint system are taken as inputs to output a Pareto optimal solution set through an improved multi-objective optimization algorithm; and finally, an integrated digital twin is constructed, and a final optimal site selection scheme is determined based on verification and scoring. The application greatly improves the scientificity, safety and economy of substation site selection, and has wide engineering application value.

Description

Technical Field

[0001] This invention relates to the field of substation site selection technology, specifically to an intelligent method for substation site selection based on multidimensional geographic information and power grid scenario constraints. Background Technology

[0002] As a core hub of the power grid, the quality of substation site selection directly determines the safety, stability, and economy of grid operation. Simultaneously, it must consider geographical compliance, policy adaptability, and full life-cycle cost control. With the advancement of new power system construction and the increasing differentiation of load distribution, an intelligent substation site selection method based on multi-dimensional geographic information and power grid scenario constraints is needed.

[0003] Existing substation site selection technologies mainly revolve around the core processes of data acquisition, area screening, scheme optimization, and simple verification, specifically manifesting in the following forms: Traditional site selection methods largely rely on the engineering experience of designers, combined with basic geographic data such as topography and geological conditions obtained from on-site surveys, as well as current power grid data such as power grid topology and load distribution, for preliminary manual screening and scheme comparison; With the application of GIS technology, layer-overlay-based site selection methods have emerged, importing data such as land use, ecological protection red lines, and transportation networks into the GIS system in layer form, visualizing potential site selection areas through layer overlay to assist in manual decision-making; Some technologies introduce simple optimization algorithms, setting fixed constraint thresholds to perform optimization calculations on candidate areas and output preliminary optimized schemes; In the identification of restricted and prohibited construction areas, manual interpretation of remote sensing images and policy texts are often used, or simple numerical threshold screening is used to extract infeasible areas; In the scheme verification stage, basic verification such as power flow calculation and safety distance verification is often performed for single-point rated operating conditions. The above-mentioned solutions have the following technical problems: 1. Data processing is fragmented. Geospatial data and power grid-related data are stored in different systems with heterogeneous formats and inconsistent spatiotemporal benchmarks. There is a lack of effective data association and integration mechanisms, which leads to cumbersome data retrieval, poor correlation, and the accuracy of site selection decisions is easily affected by missing or biased data.

[0004] 2. The identification of restricted and prohibited construction areas is inefficient and inaccurate. Manual interpretation is greatly affected by subjective factors, which is time-consuming, labor-intensive, and prone to missing key constraints. Simple threshold screening cannot accurately identify complex land cover types and policy requirements, resulting in biases in compliance judgment.

[0005] 3. The constraint system is rigid, often using fixed constraint thresholds, and fails to consider the differentiated needs of different site selection scenarios such as urban core areas, new energy bases, and rural power grids. It cannot adapt to the core demands of different regions, resulting in insufficient targeting of the solution.

[0006] 4. Limitations of algorithm performance: Traditional optimization algorithms do not integrate prior spatial knowledge of geography and power grid, generate the initial population randomly, are prone to jumping out of the feasible region, have invalid searches during the optimization process, have slow convergence speed and are prone to getting trapped in local optima, and are difficult to output the global optimal solution.

[0007] 5. The verification mechanism is imperfect, only conducting basic verification for single-point operating conditions, lacking full-scenario simulation verification and closed-loop optimization processes, and unable to comprehensively evaluate the safety and adaptability of the site selection scheme under complex operating conditions such as load fluctuations and changes in new energy output.

[0008] 6. Insufficient adaptability to different scenarios: The design does not take into account the characteristics of new power systems such as high proportion of new energy access and grid topology upgrades. It is difficult to meet the comprehensive requirements of the economic efficiency, safety and robustness of the site selection scheme under the new situation. It focuses more on the initial construction cost and does not fully consider the operation and maintenance cost, decommissioning cost and long-term grid operation efficiency, resulting in poor long-term economic efficiency of the site selection scheme. Summary of the Invention

[0009] To address the aforementioned technical shortcomings, the present invention aims to provide an intelligent method for substation site selection based on multidimensional geographic information and power grid scenario constraints.

[0010] To solve the above technical problems, the present invention adopts the following technical solution: The present invention provides an intelligent method for substation site selection based on multi-dimensional geographic information and power grid scenario constraints, including the following steps: Step 1, knowledge graph construction: collect various types of geospatial data and power grid related data, and construct a geographic power grid spatiotemporal knowledge graph.

[0011] Step 2: Obtain feasible sites: Use a multimodal large model to analyze the spatiotemporal knowledge graph of the geogrid and obtain the feasible domains of each candidate site.

[0012] Step 3: Solving for address constraints: Set up a constraint system, take the feasible region of each candidate site and the constraint system as input, and output the Pareto optimal solution set through a multi-objective optimization algorithm.

[0013] Step 4: Closed-loop simulation optimization: Construct an integrated digital twin, verify the optimal solution set, and obtain the final optimal addressing scheme.

[0014] Preferably, the process of obtaining the feasible region of each candidate site is as follows: In the geographic raster base map of the site selection area, the raster cells corresponding to the restricted construction areas are completely removed to obtain each effective raster cell. An adaptive spatial clustering algorithm is used to perform cluster analysis on each effective raster cell. Based on the spatial continuity and attribute consistency of the raster cells, the clustering neighborhood radius and minimum sample number parameters are automatically adjusted to aggregate each effective raster cell that is spatially continuous and attribute-adapted into a single candidate site patch, thereby obtaining each candidate site patch. Constraint and restriction conditions are marked on the raster cells corresponding to each restricted construction area of ​​each candidate site patch.

[0015] Each candidate site patch is verified for land area, shape regularity, and spatial continuity. The verified candidate site patches are recorded as the feasible regions of each candidate site.

[0016] Preferably, the multi-objective optimization algorithm is as follows: Based on the NSGA-III multi-objective optimization algorithm, an improved algorithm framework guided by spatial prior knowledge is built, and three core improvement units are added: a spatial constraint penalty module, an adaptive crossover and mutation module, and a prior knowledge embedding module. The core optimization objectives are the lowest life-cycle cost, the highest power supply reliability, the lowest network loss rate, and the strongest grid topology robustness. A normalized multi-objective function is constructed, and a stepped penalty function is designed for the second-level constraint exceeding the limit, while a zero-tolerance penalty mechanism is set for the first-level constraint exceeding the limit.

[0017] Load centers, weak nodes in the power grid, new energy aggregation points, and high-feasibility areas are extracted from the geographic power grid spatiotemporal knowledge graph as spatial prior knowledge. These are then embedded into the entire process of initial population generation and iterative search in the algorithm to ensure that the initial population is entirely located within the feasible domain and is preferentially generated in high-potential areas.

[0018] The algorithm convergence training was completed using the ZDT and DTLZ standard multi-objective optimization test sets. The adaptive crossover and mutation probability was optimized, and the non-dominated sorting and elite retention strategies were adjusted. When the convergence speed of the improved algorithm reached the preset speed, the algorithm optimization and deployment were completed.

[0019] The beneficial effects of this invention are as follows: 1. This invention first constructs a geographic power grid spatiotemporal knowledge graph with a unified spatiotemporal benchmark; secondly, it uses a multimodal large model to analyze the knowledge graph, automatically identifying prohibited and restricted construction zones, and obtaining feasible regions for candidate sites through an adaptive spatial clustering algorithm; then, it constructs a three-level hierarchical constraint system, using the feasible region and constraint system as input, and outputs a Pareto optimal solution set through an improved multi-objective optimization algorithm; finally, it constructs an integrated digital twin, and determines the final optimal site selection scheme based on the verification score. This invention significantly improves the scientificity, safety, and economy of substation site selection, and has broad engineering application value.

[0020] 2. This invention constructs a geographic power grid spatiotemporal knowledge graph, integrating multi-source heterogeneous data such as topography and power grid topology. It achieves precise data alignment through a combination of spatial coordinate calibration and entity feature matching, solving the problems of fragmented and low-relevance data in traditional site selection. The knowledge graph, based on a graph database, enables millisecond-level relational queries and sub-meter-level spatial accuracy, providing comprehensive and efficient data support for site selection decisions.

[0021] 3. This invention relies on the SAM remote sensing segmentation model and the finely tuned BERT pre-trained model in the field of power grid planning to construct a dual-branch multimodal model, realizing the synergistic linkage of geographic image semantic segmentation and policy text structured parsing, automatically identifying prohibited and restricted construction areas, improving efficiency and segmentation accuracy compared with traditional manual identification methods, and effectively avoiding site selection compliance risks.

[0022] 3. This invention constructs a three-level hierarchical constraint system, presets multiple typical site selection scenario templates, and realizes automatic scenario matching and adaptive correction of constraint parameters through cosine similarity calculation, which can accurately adapt to diverse site selection needs such as urban core areas, new energy bases, and rural power grids.

[0023] 4. Based on the NSGA-III algorithm, this invention adds three major modules: spatial constraint penalty, adaptive crossover mutation, and prior knowledge embedding. It integrates spatial prior knowledge such as load center and weak nodes of the power grid into the optimization process, ensuring that the entire initial population is within the feasible region, improving the convergence speed of the algorithm, effectively avoiding local optima problems, and the output Pareto optimal solution set covers the optimal solutions in multiple dimensions such as life cycle cost and power supply reliability.

[0024] 5. By constructing an integrated digital twin of the site, power grid, and geographical environment, and combining it with power grid simulation software, the optimal solution set is simulated and verified in a full-scenario manner. Three types of scores are generated: compliance, power grid safety, and constraint satisfaction. The final solution is determined based on weighted calculation, forming a complete closed loop of data fusion, feasible domain extraction, intelligent optimization, and simulation verification. This significantly reduces the risk of later changes to the site selection scheme, improves the safety and economy of power grid operation, and reduces the total life cycle cost. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments 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 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.

[0026] Figure 1 This is a schematic diagram of the implementation steps of the method of the present invention. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] according to Figure 1 As shown, this invention provides an intelligent method for substation site selection based on multidimensional geographic information and power grid scenario constraints, including the following steps: Step 1, Knowledge Graph Construction: Collect various types of geospatial data and power grid-related data to construct a geographic power grid spatiotemporal knowledge graph.

[0029] In one specific embodiment, the construction process of the geographic power grid spatiotemporal knowledge graph is as follows: the spatial boundaries of the substation site selection area are defined, and various types of geospatial data and power grid-related data are collected. The various types of geospatial data include topography, geological disasters, meteorology and hydrology, transportation network, land use and ecological protection red line data, and power grid-related data include power grid topology, time-series load, new energy output, line corridors and power grid equipment parameter data.

[0030] It should be noted that the topographic and geomorphological data acquisition subunit uses airborne LiDAR equipment and UAV aerial survey system to collect digital elevation model (DEM) and digital surface model (DSM) data of the selected site area with a spatial resolution of not less than 1m, and simultaneously collects topographic slope, aspect, and elevation difference data; the geological parameter acquisition subunit collects data on soil and rock type, bearing capacity, permeability coefficient, and geological hazard zoning of the selected site area through borehole exploration, in-situ testing, and indoor geotechnical tests.

[0031] The meteorological and hydrological data acquisition subunit connects to the meteorological science data center and local hydrological stations to collect data on extreme precipitation, extreme temperatures, lightning strike frequency, flood inundation range, and groundwater depth in the selected area over the past 30 years.

[0032] The traffic network data collection subunit collects data on the distribution, network density, and accessibility of highways, national roads, provincial roads, and county and township roads in the selected area through open-source OSM maps and local transportation bureau road network data.

[0033] The land use and ecological red line data collection sub-unit connects to the local natural resources spatial planning database to collect vector data on current land use classification, permanent basic farmland area, ecological protection red line, cultural relic protection area area, and drinking water source protection area.

[0034] Geospatial data is acquired through a dedicated data acquisition unit, while power grid-related data is extracted from the power grid dispatching system and equipment management system. The collected raw data is cleaned, denoised, and completed. After removing abnormal data, the data format and spatiotemporal reference are unified, and standardized calculations are performed. All standardized data is then imported into the data processing module. A combination of spatial coordinate calibration and entity feature matching is used to integrate and align multi-source data. Using a graph database as the storage and computation carrier, the aligned standardized data is imported into the graph database to obtain a geographic power grid spatiotemporal knowledge graph.

[0035] It should be noted that the spatial coordinate benchmark is unified: the coordinate system of all collected geospatial data and power grid related data is converted into the geodetic coordinate system, and the elevation benchmark is unified to the elevation benchmark, thus completing the coarse calibration of the spatial coordinates.

[0036] Same-name entity recognition: For the same geographical entity or power grid entity in different data sources, extract the spatial contour features, attribute features and topological relationship features of the entity, and construct the entity feature vector.

[0037] Entity feature matching: A bidirectional maximum matching algorithm is used to calculate the cosine similarity of entity feature vectors from different data sources. When the similarity is ≥0.95, they are determined to be entities with the same name.

[0038] Precise entity alignment: For entities with the same name, the least squares method is used to perform precise calibration of spatial coordinates, eliminating coordinate deviations from different data sources. At the same time, the attribute data of the entity in different data sources are correlated and merged to complete the integration and alignment of multi-source data.

[0039] Alignment result verification: Randomly select no less than 10% of entities for manual verification. When the entity alignment accuracy is ≥99%, the alignment result is deemed qualified.

[0040] Step 2: Obtain feasible sites: Use a multimodal large model to analyze the spatiotemporal knowledge graph of the geogrid and obtain the feasible domains of each candidate site.

[0041] In one specific embodiment, the parsing process of the geogrid spatiotemporal knowledge graph is as follows: extract geospatial raster images, vector boundary data, entity attribute data and policy constraint text data of the site selection area from the geogrid spatiotemporal knowledge graph, convert them into a standardized input format that can be recognized by a multimodal large model, and divide them into three types of parallel input units: image tensors, text vectors and spatial topology matrices.

[0042] It should be noted that the conversion to a standardized input format recognizable by a multimodal large model is divided into three types of parallel input units: image tensors, text vectors, and spatial topology matrices. The specific content is as follows: Image tensor generation: The geospatial raster image is normalized to a uniform scale of 1024×1024 pixels. Radiometric calibration, atmospheric correction, and cloud removal preprocessing are performed on the image. The PyTorch framework is used to convert the preprocessed image into a floating-point tensor with dimensions [batch_size, 3, 1024, 1024], where batch_size is the batch size and 3 represents the RGB three channels, which serve as the input unit for the image channels.

[0043] Text vector generation: The policy constraint text data is preprocessed by sentence segmentation, word segmentation, and stop word removal. The same word segmenter as the BERT pre-trained model is used to convert the text into a token sequence. The sequence length is uniformly padded to 512, generating a text vector with dimension [batch_size, 512], which serves as the input unit of the text channel.

[0044] Spatial topology matrix generation: Geographic entities and power grid entities in the selected area are abstracted as topology nodes, and spatial adjacency, inclusion, and intersection relationships between entities are abstracted as topology edges. An undirected topology graph is constructed, and an adjacency matrix is ​​used to represent topology relationships. The matrix dimension is consistent with the number of entities. A matrix element value of 1 indicates that two entities have a topology association, and a value of 0 indicates that there is no topology association. A standardized spatial topology matrix is ​​generated as the input unit of the topology channel.

[0045] The three types of input units after transformation are simultaneously input into the multimodal large model. The semantic segmentation and label recognition of ground objects are completed through the image channel, the structured parsing of policy constraint rules is completed through the text channel, and the spatial correlation and compliance verification are completed through the topology channel.

[0046] Based on the multi-channel analysis results, the feature segmentation labels and policy constraint rules are integrated to obtain the prohibited construction zones and restricted construction zones within the site selection area.

[0047] In one specific embodiment, the multimodal large model is constructed as follows: a SAM remote sensing segmentation large model is used as the basic model for the visual branch, and a finely tuned BERT pre-trained model from the power grid planning domain is used as the basic model for the text branch. A dual-branch multimodal fusion architecture with cross-attention mechanism fusion is built. Historical high-resolution remote sensing images of substation site selection scenarios, land and space planning texts, environmental impact assessment and safety assessment documents, restricted and prohibited construction zone annotation samples, and compliance verification rule documents are collected. Data cleaning, annotation, and data augmentation are completed to construct a site selection scenario-specific fine-tuned dataset and test set.

[0048] It should be noted that the specific content of the construction of the cross-attention mechanism fusion dual-branch multimodal fusion architecture is as follows: The dual-branch multimodal fusion architecture includes four parts: visual branch, text branch, cross-attention fusion module and output head module.

[0049] The visual branch uses a fine-tuned SAM large model image encoder, removes the original mask decoder, retains the ViT-H backbone network, and outputs a visual feature vector with a dimension of 768.

[0050] The text branch uses a finely tuned BERT pre-trained model from the power grid planning domain, retaining a 12-layer Transformer encoder, and outputting a text feature vector with a dimension of 768.

[0051] The cross-attention fusion module adopts a cross-attention mechanism, setting the query vector Q to come from the visual feature vector, and the key vector K and value vector V to come from the text feature vector. At the same time, a reverse cross-attention layer is set, where the query vector Q comes from the text feature vector, and the key vector K and value vector V come from the visual feature vector, realizing bidirectional interactive fusion of visual features and text features, and outputting the fused multimodal feature vector.

[0052] The output head module includes a fully connected layer, an activation function, and a classification layer. It takes the fused multimodal feature vector as input and outputs the land cover segmentation result, the policy constraint rule parsing result, and the compliance verification result, respectively, thus completing the architecture construction of the multimodal large model.

[0053] The underlying feature extraction parameters of the base model are frozen. Transfer learning is used to fine-tune the high-level feature layer and cross-modal fusion layer of the model based on a dedicated fine-tuning dataset. A step-adaptive learning rate is set, and the model convergence effect is jointly optimized by the cross-entropy loss function and the contrastive learning loss function.

[0054] It should be noted that the step-adaptive learning rate setting, which uses a combination of cross-entropy loss function and contrastive learning loss function to optimize model convergence, is as follows: Step-adaptive learning rate setting: A cosine annealing step-adaptive learning rate strategy is adopted, with the initial learning rate set to 2e-5. Every 10 training epochs, the learning rate decays to 90% of the original learning rate, and the minimum learning rate lower limit is set to 1e-6. At the same time, a learning rate warm-up mechanism is set, with the first 5 epochs as a warm-up period, during which the learning rate linearly increases from 1e-7 to the initial learning rate of 2e-5.

[0055] The abbreviation for e in scientific notation means multiplying by 10 to the power of 10. For example, 2e-5 is 2 multiplied by 10 to the power of -5.

[0056] Construction of joint loss function: total loss function Where α is the cross-entropy loss weight, set by the staff, typically 0.7; β is the contrastive learning loss weight, set by the staff, typically 0.3; and the formula for calculating the cross-entropy loss function Loss_ce is: Where N is the sample size. For the true labels of the samples, The model predicts probabilities, where i is the sample number, i is a positive integer, and the maximum value of i is n. The contrastive learning loss function Loss_cl adopts the InfoNCE loss, and the calculation formula is as follows: z_i and z_j are the feature vectors of positive sample pairs, z_k is the feature vector of negative samples, sim is the cosine similarity function, and τ is the temperature coefficient, which is set by the staff and is generally set to 0.07.

[0057] Model convergence optimization: The AdamW optimizer is used to update the model parameters through backpropagation based on the total loss function mentioned above. When the training set loss decreases by less than 1e-5 for 10 consecutive rounds and the validation set accuracy no longer improves, the model is considered to have converged and training is stopped.

[0058] The cross-modal feature fusion effect is optimized by using a dynamic attention mechanism, the inference speed is improved by using model quantization compression technology, and the accuracy is verified by using a test set.

[0059] Step 2: Obtain feasible sites: Use a multimodal large model to analyze the spatiotemporal knowledge graph of the geogrid and obtain the feasible domains of each candidate site.

[0060] In one specific embodiment, the process of obtaining the feasible domain of each candidate site is as follows: In the geographic raster base map of the site selection area, the raster cells corresponding to the restricted construction areas are completely removed to obtain each effective raster cell. An adaptive spatial clustering algorithm is used to perform cluster analysis on each effective raster cell. Based on the spatial continuity and attribute consistency of the raster cells, the clustering neighborhood radius and minimum sample number parameters are automatically adjusted to aggregate each effective raster cell that is spatially continuous and attribute-adapted into a single candidate site patch, thereby obtaining each candidate site patch. Constraint and restriction conditions are marked on the raster cells corresponding to each restricted construction area of ​​each candidate site patch.

[0061] It should be noted that the adaptive spatial clustering algorithm is an improved DBSCAN algorithm, and the specific execution steps are as follows: The spatial coordinates and attribute features of the effective grid cells in the selected area are used as input. Each grid cell corresponds to a two-dimensional spatial coordinate (x, y) and an attribute feature vector. Based on the spatial density of the grid cells, the adaptive neighborhood radius eps is calculated using the k-nearest neighbor distance method. Specifically, the distance from each grid cell to its k-th nearest neighbor grid cell is calculated, where k is set to 4. All k-nearest neighbor distances are sorted, and a k-distance curve is plotted. The distance value at the inflection point of the curve is taken as the adaptive neighborhood radius eps. Based on the attribute consistency of raster cells, for spatially continuous raster cell clusters with attribute feature similarity ≥ 0.9, the minimum sample size (MinPts) is set to the number of raster cells corresponding to the minimum land area of ​​the substation; for scattered raster cells with attribute feature similarity < 0.9, MinPts is set to twice the number of raster cells corresponding to the minimum land area. Based on the adaptive eps and MinPts parameters, DBSCAN clustering is performed to aggregate raster cells with achievable density into the same cluster, with each cluster corresponding to a candidate site patch; noisy raster cells with insufficient density are removed to complete the clustering analysis.

[0062] Each candidate site patch is verified for land area, shape regularity, and spatial continuity. The verified candidate site patches are recorded as the feasible regions of each candidate site.

[0063] Step 3: Address Constraint Solution: Set up a constraint system, take the feasible region of each candidate site and the constraint system as input, and output the Pareto optimal solution set through a multi-objective optimization algorithm.

[0064] In one specific embodiment, the constraint system is set up as follows: a three-level hierarchical constraint system is established, wherein the first level constraint is a veto-type constraint, the second level constraint is a flexible constraint with dynamically adjustable thresholds, and the third level constraint is the core evaluation index for multi-objective optimization.

[0065] Various typical site selection scenarios are preset. For the core requirements of each scenario, corresponding constraint priorities, threshold ranges and weight coefficient templates are pre-configured. The cosine similarity of various feature parameters of the current site selection area with various feature parameters of various typical site selection scenarios is calculated to obtain the similarity of the current typical site selection scenarios. The typical site selection scenario with the highest similarity is recorded as the current typical site selection scenario. Then, the initial constraint threshold, initial weight and initial constraint conditions of the current typical site selection scenario are obtained.

[0066] It should be noted that the cosine similarity calculation of various feature parameters of the current site selection area with various feature parameters of various typical site selection scenarios is as follows: For five types of typical site selection scenarios, namely high-density load in urban core areas, new energy base aggregation, rural power grids in counties, ultra-high voltage supporting substations, and emergency temporary substations, a 6-dimensional standard feature vector is constructed. The feature dimensions include: regional load density, proportion of new energy installed capacity, power grid voltage level, regional urbanization rate, construction period requirements, and site land area. The feature values ​​of each dimension are normalized by min-max to obtain the standard feature vector V_s of various typical scenarios, where s is the number of each typical scenario, and s is: 1, 2, 3, 4, 5.

[0067] The above six types of feature parameters of the current location area are extracted and normalized in the same way as the standard feature vector to obtain the current scene feature vector V_c. The cosine similarity between the current scene feature vector and the standard feature vectors of various typical scenes is calculated using the cosine similarity calculation formula. The five similarity values ​​are compared and the typical location scene corresponding to the maximum similarity value is determined as the matching scene of the current location area.

[0068] In another specific embodiment, the three-level hierarchical constraint system is as follows: The three-level hierarchical constraint system is divided into a rigid constraint layer, an elastic constraint layer, and an optimization target layer. The specific constraint indicators of each layer are as follows: The first layer of rigid constraints is a veto-type constraint: The site shall not occupy permanent basic farmland, ecological protection red line, primary protection zone of drinking water source, or protection scope of cultural relics protection unit; the safety distance between the site and surrounding flammable and explosive sites and civil explosive warehouses shall meet the requirements of the building design fire protection code; the protection distance between the site and electromagnetically sensitive targets such as residential areas, schools, and hospitals shall meet the requirements of the technical specification for electromagnetic radiation environmental impact assessment of 500kV ultra-high voltage power transmission and transformation projects; the slope of the site terrain shall be ≤15°, and the geological disaster risk level shall be low or below.

[0069] The second layer of elastic constraints is a flexible constraint with dynamically adjustable thresholds: the power supply radius of a 110kV substation is ≤5km, the power supply radius of a 220kV substation is ≤15km, and the power supply radius of a 500kV substation is ≤50km. The thresholds can be dynamically adjusted by ±20% according to the regional load density; the land acquisition cost of the substation site does not exceed 120% of the average cost of similar land use in the region; the total length of the incoming and outgoing line corridors does not exceed 110% of the average corridor length of substations of the same capacity; and the accessibility distance from the substation site to county-level or above highways is ≤2km.

[0070] The third layer of optimization constraints consists of core evaluation indicators for multi-objective optimization: lowest substation lifecycle cost; regional power supply reliability ≥ 99.99%; reduction of regional power grid line loss rate ≥ 0.5 percentage points; optimal power grid topology robustness; and improvement of node betweenness centrality balance ≥ 10%.

[0071] In one specific embodiment, the output of the Pareto optimal solution set is specifically output as follows: the spatial parameters of the feasible region of each candidate site, the rule parameters of the hierarchical constraint system, and the objective function parameters of the multi-objective optimization are determined as the input parameters of the optimization model. The multi-objective optimization algorithm is started, and a global iterative search is completed within the feasible region. The first-level constraint verification is performed on the candidate solutions generated by each iteration to obtain each effective solution that satisfies the first-level constraint conditions.

[0072] Based on the non-dominated sorting criterion, all valid solutions are sorted hierarchically, and non-dominated solutions located on the Pareto front are selected. Dominated inferior solutions are eliminated to form the Pareto optimal solution set.

[0073] It should be noted that the non-dominated ranking criterion is used to perform hierarchical ranking of all valid solutions and select the non-dominated solutions located on the Pareto front. The specific content is as follows: For two candidate solutions A and B, if A is not inferior to B on all optimization objectives and is superior to B on at least one optimization objective, then A is said to dominate B; if A is not dominated by any other candidate solution, then A is said to be a non-dominated solution.

[0074] For all valid solutions that satisfy the rigid constraints, iterate through the domination relationships between each solution and other solutions, count the number of times each solution is dominated, and the set of solutions dominated by that solution; solutions with a domination count of 0 are classified into the first front layer, denoted as the Pareto front, and these solutions are removed from the solution set; for the remaining solutions, repeat the above steps, successively classifying them into the second front layer, the third front layer, and so on, until all solutions are classified.

[0075] All non-dominated solutions in the first frontier layer are used to form a Pareto optimal solution set. For the solutions in the Pareto optimal solution set, a crowding degree calculation method is used to remove solutions with a crowding degree less than a preset threshold, so as to ensure the diversity and uniformity of the solution set.

[0076] The output Pareto optimal solution set consists of each candidate solution, and each candidate solution contains various core parameters. The multi-objective function values ​​and advantages and disadvantages of each solution are output simultaneously.

[0077] In one specific embodiment, the multi-objective optimization algorithm is as follows: Based on the NSGA-III multi-objective optimization algorithm, an improved algorithm framework guided by spatial prior knowledge is built, and three core improvement units are added: a spatial constraint penalty module, an adaptive crossover and mutation module, and a prior knowledge embedding module. The core optimization objectives are the lowest life cycle cost, the highest power supply reliability, the lowest network loss rate, and the strongest grid topology robustness. A normalized multi-objective function is constructed, and a stepped penalty function is designed for the second-level constraint exceeding the limit, while a zero-tolerance penalty mechanism is set for the first-level constraint exceeding the limit.

[0078] It should be noted that the three core improvement units—the spatial constraint penalty module, the adaptive crossover and mutation module, and the prior knowledge embedding module—are specifically as follows: Load centers, weak nodes in the power grid, new energy aggregation points, and high-feasibility areas are extracted from the geographic power grid spatiotemporal knowledge graph as high-potential prior regions; during initial population generation, 100% of the initial individuals are generated within the feasible region, with 80% generated within the high-potential prior regions and 20% randomly generated within the feasible region; during the iterative search process, a search weight of 1.5 is set for the high-potential prior regions, and a search weight of 0.8 is set for other regions within the feasible region, guiding the algorithm to prioritize the search of high-potential regions.

[0079] Spatial constraint penalty module: For individuals that jump out of the feasible region during the iteration process, a zero-tolerance penalty is set, and their objective function value is set to infinity, and they are directly eliminated; For individuals that exceed the elastic constraint limit, a stepped penalty function is set: the penalty coefficient is 1.2 for exceeding the limit by ≤10%; the penalty coefficient is 1.5 for exceeding the limit by 10%-30%; and the penalty coefficient is 5 for exceeding the limit by >30%. The penalty coefficient is multiplied by the individual's objective function value to update the individual's fitness value and guide the algorithm to converge to the region that satisfies the constraints.

[0080] Adaptive crossover and mutation module: Employs a simulated binary crossover operator, where the crossover probability Pc is adaptively adjusted based on the number of iterations. P, where and These are the preset maximum and minimum probabilities, preset by staff, and are generally set as follows: The value is 0.9. The value is: 0.6. For the current iteration round, The maximum number of iterations is determined by a polynomial mutation operator, with the mutation probability Pm adaptively adjusted according to the non-dominated level of the individual: Pm for individuals in the first leading edge layer is 0.01, Pm for individuals in the second leading edge layer is 0.05, and Pm for individuals in the third and subsequent leading edge layers is 0.1. In the early stages of iteration, a high crossover probability and a low mutation probability are used to ensure the algorithm's global search capability; in the later stages of iteration, a low crossover probability and a high mutation probability are used to improve the algorithm's local optimization capability.

[0081] Load centers, weak nodes in the power grid, new energy aggregation points, and high-feasibility areas are extracted from the geographic power grid spatiotemporal knowledge graph as spatial prior knowledge. These are then embedded into the entire process of initial population generation and iterative search in the algorithm to ensure that the initial population is entirely located within the feasible domain and is preferentially generated in high-potential areas.

[0082] It should be noted that the multi-objective function contains four optimization sub-objectives, all of which are min-max normalized. All sub-objectives are minimization objectives, as follows: Full life cycle cost sub-objective f1(x): obtained by performing min-max normalization on the full life cycle cost of each candidate site, including construction cost, operation and maintenance cost, and decommissioning cost, where x is the candidate site number.

[0083] The power supply reliability sub-objective f2(x) is obtained by performing min-max normalization on the average power supply availability of the area after each candidate site is put into operation.

[0084] The sub-objective of network loss rate f3(x) is obtained by performing min-max normalization on the theoretical line loss rate of the regional power grid after each candidate site is put into operation.

[0085] The sub-objective of topological robustness of the power grid is f4(x): obtained by performing min-max normalization on the coefficient of variation of the betweenness centrality of the power grid nodes after each candidate site is put into operation. The smaller the coefficient of variation, the stronger the topological robustness.

[0086] The final objective of the multi-objective optimization is minF(x): minF(x) is: [f1(x),f2(x),f3(x),f4(x)].

[0087] The algorithm convergence training was completed using the ZDT and DTLZ standard multi-objective optimization test sets. The adaptive crossover and mutation probability was optimized, and the non-dominated sorting and elite retention strategies were adjusted. When the convergence speed of the improved algorithm reached the preset speed, the algorithm optimization and deployment were completed.

[0088] Step 4: Closed-loop simulation optimization: Construct an integrated digital twin, verify the optimal solution set, and obtain the final optimal addressing scheme.

[0089] In one specific embodiment, the verification of the optimal solution set is carried out as follows: Based on a 3D GIS engine, an integrated digital twin of the site, power grid and geographical environment is constructed, and all element data of the site selection area are entered to obtain a 3D digital model of each site selection scheme.

[0090] It should be noted that the construction of an integrated digital twin of the substation site, power grid, and geographical environment specifically involves the following: loading the DEM digital elevation model, high-resolution remote sensing imagery, and 3D terrain data of the selected site area to build a 1:1 scale 3D geographical environment base, achieving 3D visualization of terrain slope, elevation, and feature distribution; based on the land area, main transformer capacity, and overall layout scheme of the candidate site, using BIM technology to build a 3D civil engineering model of the substation, including detailed models of the main transformer, distribution equipment building, main control building, grounding grid, and fire protection facilities, with a model accuracy reaching LOD300. Level; Construct a 3D power grid model consistent with the actual power grid, including transmission lines, towers, adjacent substations, and load nodes connected to substations, restoring the topological connections, electrical parameters, and operating conditions of the power grid; Perform coordinate matching and spatial fusion of the 3D models of the geographical environment, site civil engineering, and power grid structure, and associate the geological parameters, environmental parameters, power grid electrical parameters, and constraint rule data of the site to achieve bidirectional linkage between the 3D model and attribute data; Develop spatial distance measurement, compliance verification, power flow simulation visualization, and environmental impact simulation functions in the digital twin to complete the construction of an integrated digital twin.

[0091] The site coordinates, main transformer capacity, and transmission corridor planning parameters of each candidate scheme in the Pareto optimal solution set are automatically imported into the digital twin and power grid simulation software to obtain the compliance verification score, power grid safety verification score, and constraint satisfaction verification score of each candidate scheme.

[0092] It should be noted that the compliance verification score, power grid safety verification score, and constraint satisfaction verification score of each candidate scheme are as follows: Compliance verification score: 100 points in total, with 4 primary indicators, each worth 25 points, including land use compliance, safety distance compliance, environmental protection compliance, and geological condition compliance; the score for each indicator is: indicator weight × (1 - exceedance range) × 100, and the scores of all indicators are added together to obtain the compliance verification score.

[0093] Power grid safety verification score: The full score is 100 points, with 5 primary indicators, including power flow rationality (20 points), N-1 verification pass rate (25 points), voltage stability (20 points), short-circuit current compliance (20 points), and renewable energy absorption capacity (15 points). The N-1 verification pass rate is calculated as the number of verified components / total number of verified components × 100, and the score for this item is 25 × N-1 verification pass rate. The remaining indicators are scored on a percentage basis based on simulation results. The scores of all indicators are added together to obtain the power grid safety verification score.

[0094] Constraint Satisfaction Verification Score: The maximum score is 100 points, based on the candidate solutions' satisfaction of the three-level constraint system. Rigid constraints are scored as follows: 40 points for all rigid constraints are satisfied, and 0 points for any constraint not satisfied. Flexible constraints are scored as follows: 40 points for each constraint, with each constraint scored as 10 × (1 - the extent of exceeding the limit). Optimization constraints are scored as follows: 20 points for each constraint, based on the degree of achievement of the objective function. The three scores are added together to obtain the constraint satisfaction verification score.

[0095] In one specific embodiment, the process of obtaining the final optimal addressing scheme is as follows: Obtain the standard score and weight factor corresponding to the compliance verification score, power grid safety verification score and constraint satisfaction verification score from the database; divide the compliance verification score, power grid safety verification score and constraint satisfaction verification score of each candidate scheme by the corresponding standard score, then multiply by the corresponding weight factor, and finally add them together to obtain the total score of each candidate scheme; and record the candidate scheme with the largest total score as the final optimal addressing scheme.

[0096] It should be noted that the standard scores and weighting factors corresponding to the compliance verification score, power grid safety verification score, and constraint satisfaction verification score are all set by the staff.

[0097] The database is used to store the standard scores and weighting factors corresponding to compliance verification scores, power grid safety verification scores, and constraint satisfaction verification scores.

[0098] The cosine similarity, BERT pre-trained model, SAM remote sensing segmentation large model, BERT pre-trained model, string annealing step learning rate strategy, AdamW optimizer, DBSCAN algorithm and least squares calculation method described in this invention are all existing technologies and can be found on the Internet, so they will not be described in detail here.

[0099] The examples described in this invention are not limited to the specific embodiments listed above. The examples are merely illustrative to facilitate understanding of the invention and do not constitute a limitation on the scope of protection of this invention. Any modifications, equivalent substitutions, etc., made within the spirit and principles of this invention should be included within the scope of protection.

[0100] The above description is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described or use similar methods to replace them, as long as they do not deviate from the concept of the invention or exceed the scope defined in this specification, they should all fall within the protection scope of the present invention.

Claims

1. An intelligent substation site selection method based on multi-dimensional geographic information and power grid scenario constraints, characterized in that, Includes the following steps: Step 1: Knowledge Graph Construction: Collect various geospatial data and power grid-related data to construct a geospatial-power grid spatiotemporal knowledge graph; Step 2: Obtain feasible sites: Use a multimodal large model to analyze the spatiotemporal knowledge graph of the geogrid and obtain the feasible domains of each candidate site; Step 3: Address Constraint Solving: Set up a constraint system, take the feasible region of each candidate site and the constraint system as input, and output the Pareto optimal solution set through a multi-objective optimization algorithm; Step 4: Closed-loop simulation optimization: Construct an integrated digital twin, verify the optimal solution set, and obtain the final optimal addressing scheme.

2. The intelligent substation site selection method based on multi-dimensional geographic information and power grid scenario constraints according to claim 1, characterized in that, The specific construction process for building the geographic power grid spatiotemporal knowledge graph is as follows: Define the spatial boundaries of the substation site selection area, and collect various geospatial data and power grid-related data. The geospatial data includes topography, geological hazards, meteorology and hydrology, transportation network, land use and ecological protection red line data. The power grid-related data includes power grid topology, time-series load, new energy output, line corridors and power grid equipment parameter data. Geospatial data is acquired through a dedicated data acquisition unit, while power grid-related data is extracted from the power grid dispatching system and equipment management system. The collected raw data is cleaned, denoised, and completed. After removing abnormal data, the data format and spatiotemporal reference are unified, and standardized calculations are performed. All standardized data is then imported into the data processing module. A combination of spatial coordinate calibration and entity feature matching is used to integrate and align multi-source data. Using a graph database as the storage and computation carrier, the aligned standardized data is imported into the graph database to obtain a geographic power grid spatiotemporal knowledge graph.

3. The intelligent substation site selection method based on multi-dimensional geographic information and power grid scenario constraints as described in claim 1, characterized in that, The specific parsing process of the geographic power grid spatiotemporal knowledge graph is as follows: From the geospatial knowledge graph of the geogrid, geospatial raster images, vector boundary data, entity attribute data and policy constraint text data of the site selection area are extracted and converted into a standardized input format that can be recognized by the multimodal large model. The input units are divided into three types: image tensors, text vectors and spatial topology matrices. The three types of input units after transformation are simultaneously input into the multimodal large model. The semantic segmentation and label recognition of ground objects are completed through the image channel, the structured parsing of policy constraint rules is completed through the text channel, and the spatial correlation and compliance verification are completed through the topology channel. Based on the multi-channel analysis results, the feature segmentation labels and policy constraint rules are integrated to obtain the prohibited construction zones and restricted construction zones within the site selection area.

4. The intelligent substation site selection method based on multi-dimensional geographic information and power grid scenario constraints according to claim 3, characterized in that, The specific construction process of the multimodal large model is as follows: We adopted the SAM remote sensing segmentation model as the basic model for the visual branch and the BERT pre-trained model fine-tuned in the field of power grid planning as the basic model for the text branch. We built a dual-branch multimodal fusion architecture that integrates attention mechanisms. We collected historical high-resolution remote sensing images, land and space planning texts, environmental impact assessment and safety assessment documents, restricted and prohibited construction zone annotation samples, and compliance verification rule documents for substation site selection scenarios. We completed data cleaning, annotation, and data augmentation to construct a fine-tuned dataset and test set specifically for the site selection scenario. The underlying feature extraction parameters of the base model are frozen. Transfer learning is used to fine-tune the high-level feature layer and cross-modal fusion layer of the model based on a dedicated fine-tuning dataset. A step-adaptive learning rate is set, and the model convergence effect is jointly optimized by the cross-entropy loss function and the contrastive learning loss function. The cross-modal feature fusion effect is optimized by using a dynamic attention mechanism, the inference speed is improved by using model quantization compression technology, and the accuracy is verified by using a test set.

5. The intelligent substation site selection method based on multi-dimensional geographic information and power grid scenario constraints according to claim 4, characterized in that, The specific process for obtaining the feasible domains of each candidate site is as follows: In the geographic raster base map of the site selection area, the raster cells corresponding to the restricted construction areas are completely removed to obtain each effective raster cell. An adaptive spatial clustering algorithm is used to perform cluster analysis on each effective raster cell. Based on the spatial continuity and attribute consistency of the raster cells, the cluster neighborhood radius and minimum sample number parameters are automatically adjusted to aggregate each effective raster cell that is spatially continuous and attribute-adapted into a single candidate site patch. Thus, each candidate site patch is obtained. Constraint and restriction conditions are marked on the raster cells corresponding to each restricted construction area of ​​each candidate site patch. Each candidate site patch is verified for land area, shape regularity, and spatial continuity. The verified candidate site patches are recorded as the feasible regions of each candidate site.

6. The intelligent substation site selection method based on multi-dimensional geographic information and power grid scenario constraints according to claim 1, characterized in that, The specific process for setting up the constraint system is as follows: A three-level hierarchical constraint system is established, in which the first level of constraint is a veto-type constraint, the second level of constraint is a flexible constraint with dynamically adjustable thresholds, and the third level of constraint is the core evaluation index for multi-objective optimization. Various typical site selection scenarios are preset. For the core requirements of each scenario, corresponding constraint priorities, threshold ranges and weight coefficient templates are pre-configured. The cosine similarity of various feature parameters of the current site selection area with various feature parameters of various typical site selection scenarios is calculated to obtain the similarity of the current typical site selection scenarios. The typical site selection scenario with the highest similarity is recorded as the current typical site selection scenario. Then, the initial constraint threshold, initial weight and initial constraint conditions of the current typical site selection scenario are obtained.

7. The intelligent substation site selection method based on multi-dimensional geographic information and power grid scenario constraints according to claim 1, characterized in that, The output Pareto optimal solution set is produced as follows: The spatial parameters of the feasible region of each candidate site, the rule parameters of the hierarchical constraint system, and the objective function parameters of the multi-objective optimization are determined as the input parameters of the optimization model. The multi-objective optimization algorithm is started, and a global iterative search is completed within the feasible region. The first-level constraint verification is performed on the candidate solutions generated by each iteration to obtain each effective solution that satisfies the first-level constraint conditions. Based on the non-dominated sorting criterion, all valid solutions are sorted hierarchically, and the non-dominated solutions located on the Pareto front are selected. Dominated inferior solutions are eliminated to form the Pareto optimal solution set. The output Pareto optimal solution set consists of each candidate solution, and each candidate solution contains various core parameters. The multi-objective function values ​​and advantages and disadvantages of each solution are output simultaneously.

8. The intelligent substation site selection method based on multi-dimensional geographic information and power grid scenario constraints according to claim 1, characterized in that, The multi-objective optimization algorithm is described in detail below: Based on the NSGA-III multi-objective optimization algorithm, an improved algorithm framework guided by spatial prior knowledge is built. Three core improvement units are added: a spatial constraint penalty module, an adaptive crossover and mutation module, and a prior knowledge embedding module. The core optimization objectives are the lowest life cycle cost, the highest power supply reliability, the lowest network loss rate, and the strongest grid topology robustness. A normalized multi-objective function is constructed. At the same time, a stepped penalty function is designed for the second-level constraint limit violation, and a zero-tolerance penalty mechanism is set for the first-level constraint limit violation. Load centers, weak nodes in the power grid, new energy aggregation points, and high-feasibility areas are extracted from the spatiotemporal knowledge graph of the geographic power grid as spatial prior knowledge. These are then embedded into the entire process of initial population generation and iterative search in the algorithm to ensure that the initial population is located entirely within the feasible domain and is preferentially generated in high-potential areas. The algorithm convergence training was completed using the ZDT and DTLZ standard multi-objective optimization test sets. The adaptive crossover and mutation probability was optimized, and the non-dominated sorting and elite retention strategies were adjusted. When the convergence speed of the improved algorithm reached the preset speed, the algorithm optimization and deployment were completed.

9. The intelligent substation site selection method based on multi-dimensional geographic information and power grid scenario constraints according to claim 1, characterized in that, The optimal solution set is verified, and the specific verification process is as follows: Based on a 3D GIS engine, an integrated digital twin of the site, power grid, and geographical environment is constructed. All element data of the site selection area are entered to obtain a 3D digital model of each site selection scheme. The site coordinates, main transformer capacity, and transmission corridor planning parameters of each candidate scheme in the Pareto optimal solution set are automatically imported into the digital twin and power grid simulation software to obtain the compliance verification score, power grid safety verification score, and constraint satisfaction verification score of each candidate scheme.

10. The intelligent substation site selection method based on multi-dimensional geographic information and power grid scenario constraints according to claim 1, characterized in that, The specific process for obtaining the final optimal addressing scheme is as follows: Obtain the standard scores and weighting factors corresponding to the compliance verification score, power grid safety verification score, and constraint satisfaction verification score from the database. Divide the compliance verification score, power grid safety verification score, and constraint satisfaction verification score of each candidate scheme by the corresponding standard score, then multiply by the corresponding weighting factor, and finally add them together to obtain the total score of each candidate scheme. The candidate scheme with the highest total score is recorded as the final optimal location scheme.

Images