Auxiliary decision-making method and system for construction project site selection in urban planning

Abstract

The application discloses an auxiliary decision-making method and system for construction project site selection in urban planning, comprising: acquiring city multi-source heterogeneous data, construction project types and site selection constraint conditions; performing semantic modeling on the city multi-source heterogeneous data based on knowledge graph technology to generate a city element knowledge graph; performing semantic retrieval in the city element knowledge graph based on the construction project types and the site selection constraint conditions to obtain a candidate site selection set meeting the site selection constraint conditions; respectively evaluating each candidate site selection in the candidate site selection set through a subjective and objective double-channel fusion scoring model to obtain a comprehensive inappropriateness score corresponding to each candidate site selection; and iteratively optimizing the candidate site selection set based on the comprehensive inappropriateness score through a greedy take-away heuristic algorithm to obtain an optimal site selection scheme. The application can effectively improve the decision-making efficiency and quality of construction project site selection.

Description

Technical Field

[0001] This application relates to the field of urban planning technology, and in particular to auxiliary decision-making methods and systems for site selection of construction projects in urban planning. Background Technology

[0002] With the acceleration of urbanization, project site selection faces challenges such as multiple data dimensions, strong conflicting objectives, high public participation, and complex policy constraints. Traditional site selection methods mainly rely on expert experience and simple indicator aggregation, which exposes problems such as low efficiency, poor adaptability, and large subjective biases when dealing with complex urban systems. For example, traditional methods struggle to integrate spatiotemporally dynamic socioeconomic data with environmental monitoring information, leading to a disconnect between site selection schemes and actual urban development needs. Furthermore, the lack of a collaborative analysis mechanism for group preferences and objective data makes it easy to overlook public demands or fall into the trap of data-driven bias. In addition, traditional optimization algorithms often get trapped in local optima due to the large search space when dealing with high-dimensional nonlinear site selection problems, failing to guarantee the quality of the global optimal solution. Summary of the Invention

[0003] The main purpose of this application is to provide an auxiliary decision-making method and system for site selection of construction projects in urban planning, aiming to solve the technical problems of low efficiency and poor quality of traditional site selection methods.

[0004] To achieve the above objectives, this application proposes an auxiliary decision-making method for site selection of construction projects in urban planning, including:

[0005] Acquire multi-source heterogeneous data on the city, construction project types, and site selection constraints;

[0006] Based on knowledge graph technology, semantic modeling is performed on the multi-source heterogeneous data of the city to generate a knowledge graph of urban elements.

[0007] Based on the project type and the site selection constraints, a semantic search is performed on the urban element knowledge graph to obtain a set of candidate sites that meet the site selection constraints;

[0008] Each candidate site in the candidate site selection set is evaluated using a dual-channel fusion scoring model that integrates objective and subjective factors to obtain a comprehensive fitness score for each candidate site. The dual-channel fusion scoring model includes an objective data-driven channel and a subjective group preference channel. The objective data-driven channel and the subjective group preference channel dynamically adjust their output contribution ratios through a learnable gating network. The objective data-driven channel is used to output an objective fitness score, and the subjective group preference channel is used to output a group subjective preference score.

[0009] The optimal site selection scheme is obtained by iteratively optimizing the candidate site selection set based on the comprehensive inadequacy score using a greedy take-away heuristic algorithm.

[0010] In one embodiment, the step of semantically modeling the multi-source heterogeneous data of the city based on knowledge graph technology to generate a knowledge graph of urban elements includes:

[0011] The urban multi-source heterogeneous data is spatiotemporally fused to obtain spatiotemporally fused urban multi-source heterogeneous data, which includes at least urban operation data, socioeconomic statistics data, infrastructure data, environmental monitoring data, and social media data.

[0012] Define the city ontology model;

[0013] From the spatiotemporally fused urban multi-source heterogeneous data, entity instance data corresponding to the urban ontology model are extracted to obtain an initial entity set;

[0014] The initial entity set is disambiguated and merged to obtain a deduplicated entity set;

[0015] Based on the deduplicated entity set, relational reasoning is performed to obtain the set of relationships between entities in the urban element knowledge graph;

[0016] A knowledge graph of urban elements is constructed based on the set of relationships between the entities.

[0017] In one embodiment, the step of performing semantic retrieval on the urban element knowledge graph based on the project type and the site selection constraints to obtain a set of candidate sites that meet the site selection constraints includes:

[0018] Based on the project type and site selection constraints, a large language model finely tuned from the field of urban planning is used to analyze user intent and generate a project intent vector.

[0019] Based on the project intent vector, a temporary ontology structure is dynamically expanded in the urban element knowledge graph to construct a context-specific knowledge subgraph.

[0020] The urban space where the construction project is located is converted into multiple geographic analysis units. Multimodal features of each geographic analysis unit are extracted based on the urban element knowledge graph and the urban multi-source heterogeneous data, and the multimodal features are mapped to the joint embedding space.

[0021] A query vector is generated based on the project intent vector, and a nearest neighbor search is performed in the joint embedding space based on the query vector to obtain a first candidate set.

[0022] A formal logical query is performed on the knowledge subgraph based on the query vector to obtain a second candidate set;

[0023] The first candidate set and the second candidate set are weighted and merged to obtain a merged candidate set;

[0024] Acquire urban future development planning data, and perform counterfactual feasibility analysis on each candidate site in the fusion candidate set based on the urban future development planning data to obtain functional sustainability indicators of each candidate site within a future preset time window.

[0025] The fusion candidate set is filtered according to the functional sustainability index to obtain a candidate site set that meets the site selection constraints.

[0026] In one embodiment, the step of evaluating each candidate site in the candidate site selection set using a dual-channel subjective and objective scoring model to obtain a comprehensive inappropriateness score for each candidate site includes:

[0027] Each candidate site in the candidate site set is input into the subjective and objective dual-channel fusion scoring model;

[0028] Objective data features are extracted from each candidate site through the objective data-driven channel, and an objective fitness score is output based on the extracted objective data features.

[0029] The subjective group preference channel is used to collect subjective group preference information for each candidate site, and a subjective group preference score is output based on the collected subjective group preference information.

[0030] A learnable gating network is used to dynamically fuse the objective fitness score and the group's subjective preference score to obtain a comprehensive inappropriateness score for each candidate site.

[0031] In one embodiment, the step of extracting objective data features from each candidate site through the objective data-driven channel and outputting an objective fitness score based on the extracted objective data features includes:

[0032] Construct a causal graph of the city domain, and generate a causal adjacency matrix based on the causal graph of the city domain;

[0033] A multi-domain teacher-student knowledge distillation architecture is constructed, wherein the multi-domain teacher-student knowledge distillation architecture consists of multiple teacher models and a shared student model. Each teacher model corresponds to a specific professional domain, and the student model is used to learn the knowledge of each teacher model.

[0034] Based on the causal adjacency matrix and the teacher model, a knowledge distillation loss function for causal perception is designed.

[0035] The student model is trained based on the knowledge distillation loss function to obtain the trained student model;

[0036] The trained student model extracts objective data features related to each candidate site from the multi-source heterogeneous data of the city, and embeds the extracted objective data features into a preset feature space to obtain objective feature embedding.

[0037] Based on the objective feature embedding, Bayesian deep learning uncertainty modeling is performed to obtain the objective fitness score corresponding to each candidate site selection.

[0038] In one embodiment, the step of collecting subjective group preference information for each candidate site through the subjective group preference channel and outputting a subjective group preference score based on the collected subjective group preference information includes:

[0039] The system receives subjective evaluation data on each candidate site from multiple decision-makers in the form of a probabilistic language terminology set, and determines the corresponding authority weight based on the professional title, years of experience, and degree of interest of each decision-maker.

[0040] The subjective evaluation data is standardized and aligned with a language scale to obtain an aligned subjective evaluation data matrix. The rows of the aligned subjective evaluation data matrix represent different decision-makers, the columns represent different candidate sites, and the matrix element values ​​are the subjective evaluation quantification values ​​given by each decision-maker for the corresponding candidate site.

[0041] Construct a group consensus matrix based on the aligned subjective evaluation data matrix;

[0042] The authority weights are dynamically adjusted based on the group consensus matrix to obtain the adjusted authority weight matrix.

[0043] The group subjective preference score is calculated using a weighted probability language averaging operator based on the aligned subjective evaluation data matrix and the adjusted authority weight matrix.

[0044] In one embodiment, the dynamic fusion of the objective fitness score and the group's subjective preference score using a learnable gating network to obtain a comprehensive inappropriateness score for each candidate site selection includes:

[0045] Obtain the causal interferability index for each candidate site;

[0046] Construct a learnable gated network, which includes an input layer, a hidden layer, and an output layer;

[0047] The causal interventionability index, objective fitness score, and group subjective preference score of each candidate site are input into the input layer as input data.

[0048] The input data is processed by a non-linear activation function in the hidden layer to generate dynamic fusion weights;

[0049] In the output layer, the objective fitness score and the group subjective preference score are weighted and summed according to the dynamic fusion weight to obtain the comprehensive inappropriateness score corresponding to each candidate site.

[0050] In one embodiment, the step of iteratively optimizing the candidate site selection set based on the comprehensive inadequacy score using a greedy take-away heuristic algorithm to obtain the optimal site selection scheme includes:

[0051] The candidate sites in the candidate site selection set are sorted in descending order according to the comprehensive inadequacy score to obtain the sorted candidate site selection sequence;

[0052] Starting from the head of the sorted candidate address sequence, candidate addresses are selected and removed sequentially. Each time a candidate address is selected, it is determined whether the number of remaining candidate addresses in the candidate address set is less than the minimum number to retain or whether the proportion of candidate addresses currently being removed is less than the maximum removal proportion. If either condition is met, the removal operation is stopped, and a simplified candidate set is obtained.

[0053] Based on the simplified candidate set, a local optimization search is performed to obtain the optimal addressing scheme.

[0054] In one embodiment, the step of performing a local optimization search based on the simplified candidate set to obtain the optimal addressing scheme includes:

[0055] Adjust the boundary range of each candidate site in the simplified candidate set or replace some geographical units to generate multiple neighborhood candidate sites;

[0056] Calculate the comprehensive inappropriateness score for each neighborhood candidate site, select the neighborhood candidate site with the lowest comprehensive inappropriateness score to replace the original candidate site, and obtain a locally optimized simplified candidate set;

[0057] Perform a swap search operation on the locally optimized and simplified candidate set, randomly select two candidate locations and swap their positions, and calculate the comprehensive inadequacy score after the swap.

[0058] If the overall inadequacy score after the swap is lower than the overall inadequacy score before the swap, the swap operation is accepted, and the locally optimized simplified candidate set is updated. The swap search operation is repeated until the preset maximum number of iterations is reached or the locally optimized simplified candidate set has not been updated after several consecutive iterations, thus obtaining the target location set.

[0059] The candidate site with the lowest overall inadequacy score in the target site selection set is determined as the optimal site selection scheme.

[0060] Furthermore, to achieve the above objectives, this application also proposes an auxiliary decision-making system for site selection of construction projects in urban planning. The auxiliary decision-making system for site selection of construction projects in urban planning includes:

[0061] The acquisition module is used to acquire multi-source heterogeneous data on the city, construction project types, and site selection constraints.

[0062] The modeling module is used to perform semantic modeling on the multi-source heterogeneous data of the city based on knowledge graph technology, and generate a knowledge graph of urban elements.

[0063] The retrieval module is used to perform semantic retrieval in the urban element knowledge graph based on the construction project type and the site selection constraints to obtain a set of candidate sites that meet the site selection constraints.

[0064] The evaluation module is used to evaluate each candidate site in the candidate site selection set using a dual-channel fusion scoring model of objective and subjective factors, and obtain a comprehensive fitness score for each candidate site. The dual-channel fusion scoring model of objective and subjective factors includes an objective data-driven channel and a subjective group preference channel. The objective data-driven channel and the subjective group preference channel dynamically adjust their output contribution ratio through a learnable gating network. The objective data-driven channel is used to output an objective fitness score, and the subjective group preference channel is used to output a group subjective preference score.

[0065] The optimization module is used to iteratively optimize the candidate site selection set based on the comprehensive inadequacy score using a greedy heuristic algorithm to obtain the optimal site selection scheme.

[0066] This application proposes one or more technical solutions to acquire multi-source heterogeneous urban data, construction project types, and site selection constraints; perform semantic modeling on the multi-source heterogeneous urban data based on knowledge graph technology to generate an urban element knowledge graph; perform semantic retrieval on the urban element knowledge graph based on the construction project types and site selection constraints to obtain a set of candidate sites that meet the site selection constraints; evaluate each candidate site in the candidate site set using a dual-channel fusion scoring model of objective and subjective factors to obtain a comprehensive inappropriateness score for each candidate site, wherein the dual-channel fusion scoring model of objective and subjective factors includes an objective data-driven channel and a subjective group preference channel, the objective data-driven channel and the subjective group preference channel dynamically adjust their output contribution ratio through a learnable gating network, the objective data-driven channel is used to output an objective fitness score, and the subjective group preference channel is used to output a group subjective preference score; and iteratively optimize the candidate site set based on the comprehensive inappropriateness score using a greedy take-away heuristic algorithm to obtain the optimal site selection scheme. By adopting the above approach, a probabilistic language attribute group decision-making system is introduced based on a dual-channel subjective and objective scoring model to accurately integrate diverse subjective opinions. Combined with a greedy heuristic algorithm, non-optimal candidate areas are quickly eliminated, which effectively improves the decision-making efficiency and quality of project site selection. Attached Figure Description

[0067] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0068] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0069] Figure 1 This is a flowchart illustrating an example of an auxiliary decision-making method for site selection of construction projects in urban planning, as provided in this application.

[0070] Figure 2 This is a schematic diagram of the module structure of an auxiliary decision-making system for site selection of construction projects in urban planning, as described in an embodiment of this application.

[0071] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0072] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0073] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0074] This application provides a solution that introduces a probabilistic language attribute group decision-making method to accurately integrate multiple subjective opinions based on a dual-channel subjective and objective scoring model, and combines a greedy heuristic algorithm to quickly eliminate non-optimal candidate areas, effectively improving the decision-making efficiency and quality of construction project site selection.

[0075] It should be noted that the executing entity in this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or an electronic device capable of performing the above functions, or an auxiliary decision-making device for site selection of construction projects in urban planning. The following description uses an auxiliary decision-making device for site selection of construction projects in urban planning as an example to illustrate this embodiment and the subsequent embodiments.

[0076] Based on this, embodiments of this application provide an auxiliary decision-making method for site selection of construction projects in urban planning, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the auxiliary decision-making method for site selection of construction projects in urban planning, as described in this application.

[0077] In this embodiment, the auxiliary decision-making method for site selection of construction projects in urban planning includes steps S10 to S50:

[0078] Step S10: Obtain multi-source heterogeneous data of the city, construction project types, and site selection constraints.

[0079] It should be noted that urban multi-source heterogeneous data refers to data from different data sources with different structures and formats, such as urban operation data, socio-economic statistics, infrastructure data, environmental monitoring data, and social media data. This embodiment does not specifically limit this. Project types can be public facilities projects such as hospitals, schools, subway stations, logistics centers, waste incineration plants, fire stations, and sewage treatment plants, or civil building projects such as commercial complexes and residential communities, determined based on actual planning needs. Site selection constraints include legal constraints, technical constraints, and social constraints. Legal constraints include building red lines and protected areas; technical constraints include geological bearing capacity and terrain slope; and social constraints include transportation accessibility and population coverage. These can be specifically set according to project characteristics and urban planning requirements.

[0080] It is understandable that multi-source heterogeneous data in a city can be obtained through various channels such as data interfaces of urban planning departments, third-party data service platforms, or sensor networks.

[0081] Step S20: Based on knowledge graph technology, perform semantic modeling on the multi-source heterogeneous data of the city to generate a knowledge graph of urban elements.

[0082] It should be noted that knowledge graph technology, by constructing a semantic network between entities, attributes, and relationships, can integrate scattered, multi-source, heterogeneous urban data into a structured knowledge system, achieving cross-domain data semantic association and unified representation.

[0083] In this embodiment, multi-source heterogeneous urban data undergoes preprocessing, including data cleaning, format conversion, and missing value imputation, to ensure data quality meets modeling requirements. Subsequently, natural language processing (NLP) techniques are used to extract entities and relationships from unstructured data, such as identifying entities like "school" and "hospital" from planning texts, as well as relationships like "adjacent" and "service area." Simultaneously, Geographic Information System (GIS) technology is used to convert spatial data into geographic entities and establish a spatial relationship model. Finally, through knowledge fusion and reasoning techniques, the extracted entities, relationships, and geographic entities are integrated into a unified urban element knowledge graph. This network not only reflects the static relationships between various urban elements but also contains dynamic interactive information, providing rich semantic support for subsequent site selection decisions.

[0084] In one feasible implementation, step S20 may include: spatiotemporally fusing the urban multi-source heterogeneous data to obtain spatiotemporally fused urban multi-source heterogeneous data, wherein the urban multi-source heterogeneous data includes at least urban operation data, socioeconomic statistics data, infrastructure data, environmental monitoring data, and social media data; defining an urban ontology model; extracting entity instance data corresponding to the urban ontology model from the spatiotemporally fused urban multi-source heterogeneous data to obtain an initial entity set; performing disambiguation and merging processing on the initial entity set to obtain a deduplicated entity set; performing relational reasoning based on the deduplicated entity set to obtain a set of inter-entity relationships in the urban element knowledge graph; and constructing the urban element knowledge graph based on the set of inter-entity relationships.

[0085] It should be noted that, in this embodiment, the multi-source heterogeneous urban data includes at least urban operation data, socioeconomic statistics, infrastructure data, environmental monitoring data, and social media data. Urban operation data includes geographic information data, traffic data, dynamic pedestrian flow data, and remote sensing imagery data. Geographic information data includes topography, land use, and existing building distribution. Traffic data includes road networks and public transportation coverage. Dynamic pedestrian flow data includes mobile phone signaling trajectories and public transport card swipe records. Remote sensing imagery data refers to high-resolution satellite or aerial images used to extract features such as green spaces and water bodies. Socioeconomic statistics include population density, commercial area distribution, and land costs. Infrastructure data includes POI data for public service facilities. Environmental monitoring data includes air quality, water body distribution, and ecologically sensitive areas. Social media data refers to public opinions and sentiment feedback from social media platforms, reflecting public views and attitudes towards the urban environment, facilities, and services. Through spatiotemporal fusion technology, these data from different times and spaces can be organically integrated, eliminating spatiotemporal heterogeneity and forming a unified and coherent dataset.

[0086] Understandably, defining a city ontology model allows for the construction of domain-specific upper-level and application ontology, clarifying entity types, attributes, and relationships within the city, and providing a standardized framework for subsequent data extraction and relationship reasoning. The spatiotemporally fused, multi-source heterogeneous urban data is then analyzed to extract entity instance data conforming to the ontology model definition, forming an initial entity set. When analyzing different data sources within the spatiotemporally fused, multi-source heterogeneous urban data, appropriate analysis strategies must be adopted based on the characteristics of each data source to ensure the accuracy and completeness of the extracted entity instance data. For example, for public service facility POI data, a combination of regular expression matching and classification mapping can be used to extract entity instances; for mobile signaling data, a combination of cluster analysis and dwell point detection can be used; for remote sensing image data, deep learning segmentation can be used; and for social media data, a combination of NLP named entity recognition and sentiment analysis can be used to extract entity instances.

[0087] It is worth noting that the purpose of disambiguation merging of the initial entity set is to eliminate the problem of entity duplication caused by different data sources or differences in expression. Similarity algorithms, such as Levenshtein distance and GeoHash nearest neighbor, can be used to determine whether two entities point to the same real object. Then, a rule engine or machine learning model can be applied to perform disambiguation merging to obtain the deduplicated entity set.

[0088] It is worth noting that by utilizing spatial analysis, statistical correlation, and logical rules, the set of relationships between entities in the urban element knowledge graph can be automatically derived. For example, by analyzing the connectivity between schools and surrounding roads, the "adjacent" relationship between schools and roads can be derived; by analyzing the overlap between the service area of ​​a hospital and the surrounding population distribution, the "service coverage" relationship between the hospital and the population can be derived.

[0089] A database of urban entities and their semantic relationships stored in a graph structure forms an urban element knowledge graph. This urban element knowledge graph supports efficient semantic querying and reasoning, and can quickly respond to various information needs in the site selection decision-making process.

[0090] Step S30: Based on the project type and the site selection constraints, perform semantic retrieval in the urban element knowledge graph to obtain a set of candidate sites that meet the site selection constraints.

[0091] It should be noted that by performing semantic retrieval based on the project type and site selection constraints in the urban element knowledge graph, the rich semantic association information in the knowledge graph can be fully utilized to accurately filter out candidate sites that meet various constraints. For example, when the project type is a hospital and the site selection constraints include a certain distance from the existing hospital, convenient transportation, and dense population coverage, the system will search for areas in the knowledge graph that meet these semantic requirements and summarize these areas to form a set of candidate sites.

[0092] Understandably, based on the type of construction project and site selection constraints, an intelligent semantic retrieval method that integrates contextual understanding, dynamic ontology expansion, multimodal embedding-assisted retrieval, and counterfactual feasibility verification can efficiently identify a set of candidate sites that meet the needs of the construction project from the urban element knowledge graph, thereby further improving the accuracy and efficiency of site selection.

[0093] In one feasible implementation, step S30 may include: based on the project type and site selection constraints, using a language model fine-tuned in the urban planning domain to analyze user intent and generate a project intent vector; dynamically expanding a temporary ontology structure in the urban element knowledge graph based on the project intent vector to construct a context-specific knowledge subgraph; converting the urban space where the project is located into multiple geographic analysis units, extracting multimodal features of each geographic analysis unit based on the urban element knowledge graph and the urban multi-source heterogeneous data, and mapping the multimodal features to a joint embedding space; generating a query vector based on the project intent vector, and according to... The query vector performs a nearest neighbor search in the joint embedding space to obtain a first candidate set; a formal logical query is performed on the knowledge subgraph based on the query vector to obtain a second candidate set; the first and second candidate sets are weighted and fused to obtain a fused candidate set; urban future development planning data is obtained, and counterfactual feasibility analysis is performed on each candidate site in the fused candidate set based on the urban future development planning data to obtain the functional sustainability index of each candidate site within a future preset time window; the fused candidate set is filtered based on the functional sustainability index to obtain a set of candidate sites that meet the site selection constraints.

[0094] It should be noted that a language model fine-tuned in the field of urban planning is used to perform semantic analysis on project types and site selection constraints, generating a project intent vector that accurately captures the core needs of users. This vector serves as the benchmark for subsequent searches, ensuring a high degree of consistency between the search direction and user intent. The project intent vector includes core functional category tags, target audience characteristics, spatial accessibility preferences, environmental atmosphere preferences, and commercial competition sensitivity levels. This fine-tuned language model in the field of urban planning has been trained through comparative learning from historical successful and unsuccessful cases, possessing contextual awareness capabilities. It can accurately identify implicit semantics and potential relationships in project types and site selection constraints. For example, when processing "waste incineration plant site selection," it can not only identify explicit constraints such as "environmental standards" and "transportation distance," but also capture implicit factors such as public acceptance and community impact.

[0095] Based on the generated project intent vector, a temporary ontology structure is dynamically expanded in the urban element knowledge graph to construct a knowledge subgraph highly matched with the current construction project context. Its nodes include: newly created entity classes for the current project type; related entities such as facilities, population, transportation, and environment; and dynamically derived relationship paths, such as "areas attracting young customers are usually near subway stations and cultural and creative parks." Simultaneously, a graph neural network is used to embed the newly created entity classes into the space of the urban element knowledge graph, ensuring semantic consistency. The dynamic ontology expansion process in this implementation supports zero-sample transfer and is applicable to novel facility types that are not predefined. For example, when a new public facility such as a "community elderly care center" appears, its functional attributes and related elements can be automatically parsed, and a corresponding context-specific knowledge subgraph can be quickly constructed in the knowledge graph.

[0096] A geographic analysis unit is the smallest geographic area unit that divides urban space according to certain rules. These units can be regular grids, such as square or hexagonal grids, or irregular areas divided based on road networks, administrative divisions, or natural geographic boundaries. Converting the urban space where a construction project is located into multiple geographic analysis units helps to analyze urban spatial characteristics more precisely, providing a more accurate basis for site selection decisions. Based on urban element knowledge graphs and multi-source heterogeneous urban data, multimodal features of each geographic analysis unit are extracted, including knowledge graph embedding vectors, remote sensing image features, street view image encoding, and social media text sentiment vectors. This embodiment does not impose specific limitations on these features and maps them to a joint embedding space, enabling the comparison and fusion of features from different modalities within the same semantic space.

[0097] A query vector is generated based on the project intent vector. This query vector integrates the core needs and preferences of the construction project. A nearest neighbor search is performed in the joint embedding space based on the query vector, i.e., the cosine similarity between all geographic analysis units and the query vector is calculated. This quickly identifies the geographic analysis units that best match the project intent, forming the first candidate set. These geographic analysis units have high similarity to the project requirements in terms of spatial features, functional attributes, or environmental conditions. Simultaneously, a formal logical query is performed on the knowledge subgraph based on the query vector. Utilizing the semantic relationships and logical rules in the knowledge graph, deeper information related to the project intent is mined, forming the second candidate set. These candidate sets may contain geographic analysis units that are not directly reflected in the joint embedding space but meet the conditions through semantic association. The first and second candidate sets are then weighted and fused, comprehensively considering spatial similarity and semantic association, to obtain the fused candidate set, as shown in the following equation:

[0098]

[0099] in, To merge candidate sets, As the second candidate set, As the first candidate set, This is a weighting coefficient that can be automatically adjusted based on the project type. For example, infrastructure projects tend to emphasize logic, while commercial projects tend to emphasize perception.

[0100] It is worth noting that the city's future development planning data includes control plans, rail transit construction plans, and lists of old city renovation and demolition projects. This data reflects the city's future spatial layout adjustments, infrastructure construction plans, and land use change trends. After obtaining the city's future development planning data, counterfactual feasibility analysis is performed on each candidate site in the integrated candidate set based on this data. This predicts the changes in their status over the next 3-5 years, including: whether they face changes in land use, whether there are plans to construct similar competing facilities, and whether road expansion will lead to increased noise. A causal reasoning graph model is then constructed to assess their survival index, i.e., functional sustainability indicators. Candidate sites with survival indices below a preset threshold are filtered out, ensuring that the final set of candidate sites not only meets current site selection constraints but also has the potential to continue functioning in the future. This forward-looking analysis effectively avoids the problem of unreasonable site selection caused by changes in urban development, improving the long-term stability and adaptability of site selection decisions.

[0101] Step S40: Evaluate each candidate site in the candidate site selection set using a dual-channel fusion scoring model to obtain a comprehensive fitness score for each candidate site. The dual-channel fusion scoring model includes an objective data-driven channel and a subjective group preference channel. The objective data-driven channel and the subjective group preference channel dynamically adjust their output contribution ratios through a learnable gating network. The objective data-driven channel is used to output an objective fitness score, and the subjective group preference channel is used to output a group subjective preference score.

[0102] It should be noted that the dual-channel fusion scoring model includes an objective data-driven channel and a subjective group preference channel. The objective data-driven channel outputs an objective fitness score, quantitatively evaluating the objective conditions of candidate sites based on multi-source heterogeneous urban data. The subjective group preference channel outputs a group subjective preference score, capturing and quantifying the subjective cognition and social perception of diverse stakeholders based on a probabilistic language group decision-making mechanism. To avoid either the subjective or objective channel unilaterally dominating the decision-making process, the objective data-driven channel and the subjective group preference channel adaptively adjust their contribution ratios according to project type, stage, and development environment through a learnable gating network.

[0103] Step S50: The candidate site selection set is iteratively optimized based on the comprehensive inadequacy score using a greedy take-away heuristic algorithm to obtain the optimal site selection scheme.

[0104] It's worth noting that the greedy removal heuristic algorithm can efficiently eliminate suboptimal solutions within a large pool of candidate regions, quickly converging to a set of high-quality site selection solutions, thus solving the problem of high computational cost in traditional evolutionary algorithms. In site selection optimization scenarios, this algorithm first sorts each candidate site in the candidate site set according to its comprehensive inappropriateness score. The higher the score, the more unsuitable the site, and the higher it ranks in the sort. The algorithm then sequentially selects the candidate site with the highest inappropriateness score from the sorted candidate site set and removes it from the candidate set. This process simulates the "greedy" removal of the most unsuitable site. Through continuous iteration, suboptimal sites are gradually eliminated, narrowing the candidate range. As candidate sites with high inappropriateness scores are removed, the overall suitability of the remaining candidate sites continuously improves.

[0105] Furthermore, to avoid getting stuck in local optima due to prematurely eliminating potential high-quality sites during algorithm iteration, a backtracking mechanism can be introduced. Specifically, after a certain number of candidate sites are eliminated consecutively, the algorithm backtracks to the previous round's candidate set state, re-evaluates the overall inadequacy score of the remaining sites, and dynamically corrects the score by combining dynamic correlation information in the city element knowledge graph. For example, if a candidate site's score decreases due to short-term surrounding planning and construction, but the knowledge graph shows its significant long-term development potential, its score can be improved through the correction mechanism to avoid premature elimination. Through the synergistic effect of backtracking and dynamic correction, the algorithm can achieve a balance between global and local considerations, ensuring that the final site selection scheme is both optimal in the current context and adaptable to the future. In addition, to improve the robustness of the algorithm, a dynamic termination condition can be set. When candidate sites are not eliminated in multiple consecutive iterations or the difference in the overall inadequacy scores of the remaining candidate sites is less than a preset threshold, the optimization process automatically terminates, and the current optimal site selection scheme is output.

[0106] In one feasible implementation, step S50 may include: sorting each candidate location in the candidate location set in descending order according to the comprehensive inadequacy score to obtain a sorted candidate location sequence; starting from the head of the sorted candidate location sequence, sequentially selecting candidate locations for removal, and determining whether the number of remaining candidate locations in the candidate location set is less than the minimum retention number or whether the proportion of candidate locations currently being removed is less than the maximum removal proportion when each selection is made; if either condition is met, stopping the removal operation to obtain a simplified candidate set; and performing a local optimization search based on the simplified candidate set to obtain the optimal location scheme.

[0107] It should be noted that the candidate sites in the candidate site set are sorted in descending order according to their comprehensive inappropriateness scores. After sorting, candidate sites with high comprehensive inappropriateness scores will be at the front of the sequence. These sites have lower priority in site selection decisions due to issues such as inconvenient transportation, environmental conflicts, or incompatibility with future urban planning. By sorting in descending order, the algorithm can prioritize the processing of sites with the most significant inappropriateness, thereby efficiently narrowing down the candidate range during the iteration process. For example, if a candidate site causes excessive environmental noise due to its proximity to an industrial area, its comprehensive inappropriateness score will be significantly higher, placing it at the top of the ranking, and the algorithm will remove it first.

[0108] During the removal process, minimum retention quantity and maximum removal ratio are set as constraints to prevent the candidate set from becoming too small or the optimization process from ending prematurely due to excessive removal. The minimum retention quantity ensures that there are enough remaining candidate sites for subsequent local optimization searches, avoiding the impact of insufficient samples on the accuracy of the results; the maximum removal ratio limits the proportion of candidate sites removed in a single iteration, preventing the loss of potential high-quality sites due to excessive removal at once. When the condition for stopping the removal operation is met, a simplified candidate set is obtained. The candidate sites in this set are more representative in terms of spatial distribution, functional attributes, or environmental conditions, providing high-quality input for subsequent local optimization searches.

[0109] In one feasible implementation, the step of performing a local optimization search based on the simplified candidate set to obtain the optimal location scheme includes: adjusting the boundary range of each candidate location in the simplified candidate set or replacing some geographical units to generate multiple neighborhood candidate locations; calculating the comprehensive inadequacy score of each neighborhood candidate location, selecting the neighborhood candidate location with the lowest comprehensive inadequacy score to replace the original candidate location, and obtaining a locally optimized simplified candidate set; performing a swap search operation on the locally optimized simplified candidate set, randomly selecting two candidate locations to swap locations, and calculating the comprehensive inadequacy score after the swap; if the comprehensive inadequacy score after the swap is lower than the comprehensive inadequacy score before the swap, then accepting the swap operation and updating the locally optimized simplified candidate set, repeating the swap search operation until a preset maximum number of iterations is reached or the locally optimized simplified candidate set has not been updated after several consecutive iterations, thus obtaining a target location set; and determining the candidate location with the lowest comprehensive inadequacy score in the target location set as the optimal location scheme.

[0110] It should be noted that adjusting the boundary range of each candidate site in the simplified candidate set or replacing some geographical units aims to explore potentially better areas around the candidate sites. This fine-tuning method generates multiple neighboring candidate sites, expanding the site search scope and increasing the likelihood of discovering better sites. Calculating the comprehensive suitability score of each neighboring candidate site allows for a quantitative assessment of its suitability. Replacing the original candidate site with the neighboring candidate site that has the lowest comprehensive suitability score can gradually optimize the simplified candidate set and improve the overall site selection quality.

[0111] The algorithm performs a swap search operation on the locally optimized, simplified candidate set, randomly selecting two candidate sites for location swapping. By calculating the overall suitability score before and after the swap, it determines whether the swap improves site suitability. If the score is lower after the swap, it indicates that the swap improved site suitability, and the swap is accepted, updating the simplified candidate set. This process, by randomly swapping candidate site positions, breaks the limitations of the original site layout, providing the possibility of discovering better site combinations. The swap search operation is repeated until a preset maximum number of iterations is reached or the simplified candidate set is not updated after several consecutive iterations. At this point, the algorithm considers that the site quality cannot be further improved through swapping, thus obtaining the target site set. Finally, the candidate site with the lowest overall suitability score in the target site set is determined as the optimal site solution. This solution, considering both objective conditions and subjective preferences, and balancing current suitability with future adaptability, provides a scientific and reasonable decision-making basis for site selection in urban planning projects.

[0112] This embodiment provides an auxiliary decision-making method for project site selection in urban planning. It acquires multi-source heterogeneous urban data, project types, and site selection constraints; performs semantic modeling on the multi-source heterogeneous urban data based on knowledge graph technology to generate an urban element knowledge graph; performs semantic retrieval on the urban element knowledge graph based on the project types and site selection constraints to obtain a set of candidate sites that meet the constraints; evaluates each candidate site in the set using a dual-channel fusion scoring model (objective and subjective), obtaining a comprehensive inadequacy score for each candidate site. The dual-channel fusion scoring model includes an objective data-driven channel and a subjective group preference channel. The objective data-driven channel and the subjective group preference channel dynamically adjust their output contribution ratios through a learnable gating network. The objective data-driven channel outputs an objective fitness score, and the subjective group preference channel outputs a group subjective preference score. Finally, iteratively optimizes the candidate site set based on the comprehensive inadequacy score using a greedy heuristic algorithm to obtain the optimal site selection scheme. By adopting the above approach, a probabilistic language attribute group decision-making system is introduced based on a dual-channel subjective and objective scoring model to accurately integrate diverse subjective opinions. Combined with a greedy heuristic algorithm, non-optimal candidate areas are quickly eliminated, which effectively improves the decision-making efficiency and quality of project site selection.

[0113] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in the first embodiment described above can be referred to the above description and will not be repeated hereafter. Based on this, step S20 includes steps S401 to S404:

[0114] Step S201: Input each candidate location in the candidate location set into the subjective and objective dual-channel fusion scoring model.

[0115] It should be noted that each candidate site in the candidate site selection set is input into the dual-channel fusion scoring model of objective and subjective factors, which simultaneously triggers the operation of the objective data-driven channel and the subjective group preference channel.

[0116] Step S402: Extract objective data features from each candidate site through the objective data-driven channel, and output an objective fitness score based on the extracted objective data features.

[0117] It should be noted that the objective data-driven channel utilizes a dynamic knowledge distillation-causal reasoning joint architecture and an adaptive uncertainty quantification mechanism to extract objective data features related to candidate site selection from multi-source heterogeneous urban data, and outputs an objective fitness score based on these features. The dynamic knowledge distillation-causal reasoning joint architecture can deeply mine the potential causal relationships in the data, while the adaptive uncertainty quantification mechanism can accurately assess the uncertainty in the data, thereby ensuring that the output objective fitness score has high accuracy and reliability.

[0118] In one feasible implementation, step S402 may include: constructing a causal graph of the urban domain and generating a causal adjacency matrix based on the causal graph; constructing a multi-domain teacher-student knowledge distillation architecture, wherein the multi-domain teacher-student knowledge distillation architecture consists of multiple teacher models and a shared student model, each teacher model corresponding to a specific professional domain, and the student model being used to learn the knowledge of each teacher model; designing a causal perception knowledge distillation loss function based on the causal adjacency matrix and the teacher models; training the student model based on the knowledge distillation loss function to obtain a trained student model; extracting objective data features related to each candidate site from multi-source heterogeneous data of the city through the trained student model, and embedding the extracted objective data features into a preset feature space to obtain objective feature embedding; performing Bayesian deep learning uncertainty modeling based on the objective feature embedding to obtain an objective fitness score corresponding to each candidate site.

[0119] It should be noted that the urban domain causal graph is a causal network between the objective characteristics of candidate site selection and project success, established through an unsupervised causal discovery algorithm. It acquires operational performance data of historically completed projects and their site selection environmental characteristics, uses an unsupervised causal discovery algorithm to learn the directed dependencies between variables, and introduces expert prior rules to filter illegal causal paths, outputting a weighted directed acyclic graph, namely the urban domain causal graph G=(V,E,W), where V is the node set, such as population density, green space ratio, public transportation coverage, etc.; E is the edge set, representing the causal direction; and W is the weight set, quantifying the causal strength. The causal adjacency matrix generated based on the urban domain causal graph can explicitly express the causal dependencies between various site selection features in numerical form.

[0120] In the multi-domain teacher-student knowledge distillation architecture, multiple specialized sub-models serve as teacher models, sharing a single student model to achieve cross-domain knowledge transfer. That is, the multi-domain teacher-student knowledge distillation architecture consists of multiple teacher models and a shared student model. Each teacher model corresponds to a specific professional domain, and the student model learns the knowledge from each teacher model. The multiple teacher models are deep learning sub-models based on different professional domains, including: a transportation accessibility model T1, an ecological carrying capacity model T2, an economic vitality model T3, and a social inclusion model T4. Specifically, the input of the transportation accessibility model T1 is road network topology, and the output is 15-minute coverage rate; the input of the ecological carrying capacity model T2 is remote sensing and environmental monitoring data, and the output is a sustainability index; the input of the economic vitality model T3 is POIs and consumption trajectories, and the output is business potential; and the input of the social inclusion model T4 is population structure, and the output is a fairness score. These teacher models perform deep modeling for different professional domains, capturing the unique characteristics of each domain. The student model, through the knowledge distillation process, comprehensively learns the knowledge from each teacher model, forming a cross-domain integrated understanding.

[0121] Based on the causal adjacency matrix and the teacher model, a causal-aware knowledge distillation loss function is designed. This function guides the student model to focus on features with significant causal relationships during the learning process, ensuring that the extracted objective data features are both domain-specific and conform to causal logic. The knowledge distillation loss function should include three terms: the first term is the distillation loss term, used to allow the student to imitate the output of multiple teacher models; the second term is the causal regularization term, which introduces urban causal knowledge to improve interpretability and robustness; and the third term is the regularization term, which prevents overfitting and ensures the model's generalization ability, as shown in the following equation:

[0122]

[0123] in, Total loss value, Embedding unified objective features for student models. The projection head weights for the k-th task are... For the k-th teacher model, the input... The predicted output, This is the causal regularization strength coefficient. Output the gradient of the student model relative to the input x. It is a causal adjacency matrix. The Hadamard product of gradient and causal graph. This represents the loss value for the regularization term. These are the weight coefficients for the distillation loss term, the causal regularization term, and the regularization term, respectively.

[0124] Through this loss function, the student model can comprehensively consider the output of the teacher model, causal relationships, and the model's generalization ability during training, thereby extracting more professional and causally logical objective data features. The trained student model possesses cross-domain comprehensive understanding capabilities, enabling it to accurately extract objective data features related to each candidate site from multi-source heterogeneous urban data, and embed these features into a pre-defined feature space to obtain objective feature embeddings.

[0125] Understandably, objective feature embeddings serve as input for Bayesian deep learning uncertainty modeling, enabling multiple forward samplings and estimation of cognitive uncertainty. Through Bayesian deep learning uncertainty modeling, the fitness score and uncertainty level of candidate site selection at the objective data level can be quantified.

[0126] The uncertainty modeling process of Bayesian deep learning includes constructing a Bayesian neural network, embedding objective features as input data into the network, using Bayesian inference methods to estimate the posterior distribution of network parameters, predicting the performance of each candidate site under different conditions based on the posterior distribution of parameters, and combining these performances to obtain an objective fitness score.

[0127] Step S403: Collect subjective preference information for each candidate site through the subjective group preference channel, and output a subjective preference score based on the collected subjective preference information.

[0128] It should be noted that the subjective group preference channel is constructed based on a probabilistic language group decision-making mechanism. This mechanism introduces a set of probabilistic language terms to describe the decision-maker's subjective evaluation of candidate sites, effectively solving the problems of information loss and ambiguity in traditional language evaluation.

[0129] Specifically, decision-makers subjectively evaluate candidate sites using probabilistic terminology, such as "highly likely suitable" or "moderately likely average," based on their own preferences and experience. This evaluation information is collected and transformed into a probabilistic evaluation matrix. Subsequently, a probabilistic group decision-making algorithm processes the evaluation matrix, calculating a comprehensive score for each candidate site at the level of group subjective preference, thus obtaining a group subjective preference score. In this process, the probabilistic group decision-making algorithm fully considers the uncertainty and ambiguity of decision-makers' evaluations, aggregating the subjective opinions of multiple decision-makers to form a more representative and reliable group subjective preference result. Simultaneously, this mechanism allows decision-makers to dynamically adjust evaluation terms and weights according to actual circumstances, further enhancing the flexibility and adaptability of decision-making. Through the operation of the subjective group preference channel, collective wisdom can be fully explored and integrated, providing strong support for the comprehensive evaluation of candidate sites.

[0130] In one feasible implementation, step S403 may include: receiving subjective evaluation data on each candidate site from multiple decision-makers in the form of a probabilistic language terminology set, and determining corresponding authority weights based on the professional titles, years of experience, and degree of interest of each decision-maker; standardizing and aligning the subjective evaluation data with a language scale to obtain an aligned subjective evaluation data matrix, wherein the rows of the aligned subjective evaluation data matrix represent different decision-makers, the columns represent different candidate sites, and the matrix element values ​​are the quantitative values ​​of the subjective evaluations given by each decision-maker for the corresponding candidate site; constructing a group consensus matrix based on the aligned subjective evaluation data matrix; dynamically adjusting the authority weights according to the group consensus matrix to obtain an adjusted authority weight matrix; and calculating the group subjective preference score using a weighted probabilistic language averaging operator based on the aligned subjective evaluation data matrix and the adjusted authority weight matrix.

[0131] It should be noted that the system automatically identifies the decision-making entities involved based on the project type. Each decision-maker inputs their subjective evaluation data for each candidate site through a specific interactive interface, using a probabilistic linguistic terminology set. Based on the decision-maker's professional title, years of experience in the field, and degree of stake in the project, a corresponding authority weight is assigned to each decision-maker. Decision-makers with higher titles, more experience, and closer stake in the project are given higher weights to reflect their professionalism and importance in the decision-making process.

[0132] Different decision-makers may use slightly different linguistic scales; therefore, a unified mapping is required. This includes converting all subjective evaluation data in probabilistic linguistic terminology sets into a standard 7-level semantic scale, performing approximate matching using a semantic similarity function, smoothing out missing or extreme distributions, and finally generating an aligned subjective evaluation data matrix. The rows of this matrix represent different decision-makers, the columns represent different candidate sites, and the matrix elements are the quantitative subjective evaluation values ​​given by each decision-maker for the corresponding candidate site.

[0133] Based on the aligned subjective evaluation data matrix, a group consensus matrix is ​​constructed by defining the generalized Jensen-Shannon distance between two subjective evaluation data points, as follows:

[0134]

[0135] in, For two subjective evaluation data and Generalized Jensen-Shannon distance between and relative to the mean The Kullback-Leibler divergence.

[0136] For each candidate site The generalized Jensen-Shannon distance between expert k and k′ is calculated as follows:

[0137]

[0138] in, For experts k and k′ regarding candidate site selection The generalized Jensen-Shannon distance between subjective evaluation data. and Experts k and k′ respectively select candidate sites. The provided subjective evaluation data.

[0139] The divergence index between experts k and k′ can be obtained by averaging the generalized Jensen-Shannon distance over the entire candidate set, as shown in the following formula:

[0140]

[0141] in, Let k be the divergence index between experts k and k′, and N be the total number of candidate sites.

[0142] By calculating the divergence index among all expert pairs, a group consensus matrix can be constructed, which reflects the degree of consensus among different experts on each candidate site. The smaller the element value of the group consensus matrix, the higher the degree of consensus among the corresponding expert pairs on that candidate site.

[0143] Based on the group consensus matrix, the initially assigned authority weights can be dynamically adjusted. Specifically, for expert pairs with high consensus, their authority weights can be enhanced to a certain extent; while for expert pairs with low consensus, their authority weights are correspondingly weakened. This dynamic adjustment mechanism helps improve the consistency and reliability of group decision-making, ensuring that the final group subjective preference score is more accurate and objective. The calculation is performed by introducing a weighted probability linguistic averaging operator, combined with the adjusted authority weight matrix and the aligned subjective evaluation data matrix, as shown in the following equation:

[0144]

[0145] in, Let m be the group's subjective preference score for candidate location i, and m be the number of decision-makers. The adjusted authority weight of the k-th decision-maker. Let be the subjective evaluation quantified by the k-th decision-maker for candidate location i, and PLWA be the weighted probabilistic language average operator.

[0146] The weighted probabilistic linguistic averaging operator obtains the group subjective preference score for each candidate site by weighted summation of the quantified subjective evaluation values ​​of each decision-maker, thus achieving effective integration and quantitative evaluation of group subjective preferences. This subjective group preference channel based on a probabilistic linguistic group decision-making mechanism not only fully considers the uncertainty and ambiguity of decision-makers' evaluations, but also improves the accuracy and reliability of group decision-making by dynamically adjusting authority weights and introducing the weighted probabilistic linguistic averaging operator, providing strong support for the comprehensive evaluation of candidate sites.

[0147] Step S404: Dynamically fuse the objective fitness score and the group subjective preference score using a learnable gating network to obtain a comprehensive inappropriateness score for each candidate site.

[0148] It should be noted that a learnable gated network is a neural network structure with adaptive adjustment capabilities. It can dynamically allocate the weights of objective fitness scores and group subjective preference scores based on the characteristics and importance of the input data.

[0149] Specifically, the gating network receives objective fitness scores and group subjective preference scores as inputs, and generates a dynamic fusion weight through nonlinear transformation and parameter learning. This weight determines the proportion of objective data and subjective preferences in the comprehensive evaluation process.

[0150] In one feasible implementation, step S404 may include: obtaining the causal interventionability index of each candidate site; constructing a learnable gating network, the gating network comprising an input layer, a hidden layer, and an output layer; inputting the causal interventionability index, objective fitness score, and group subjective preference score of each candidate site as input data into the input layer, processing the input data in the hidden layer using a nonlinear activation function to generate dynamic fusion weights; and performing a weighted summation of the objective fitness score and the group subjective preference score according to the dynamic fusion weights in the output layer to obtain a comprehensive inappropriateness score corresponding to each candidate site.

[0151] It should be noted that the causal interventionability index is calculated based on the causal adjacency matrix G and is used to characterize the strength of the causal effect that may occur when intervening in candidate site selection, as shown in the following formula:

[0152]

[0153] in, Let i be the causal interferability index of candidate site i. Let d be the set of variables related to candidate site i, such as green space ratio and number of bus stops, and let d be the dimension of the causal adjacency matrix G. Let be the element in the k-th row and j-th column of the causal adjacency matrix G, representing the strength of the causal influence of variable k on variable j. The weighting coefficient for variable k can be set according to actual needs or prior knowledge.

[0154] In the hidden layers of a gated network, nonlinear activation functions perform complex transformations on the input data to capture nonlinear relationships within it. Commonly used nonlinear activation functions include ReLU, sigmoid, and tanh, which introduce nonlinear characteristics, giving the gated network stronger expressive power and adaptability. Through the processing in the hidden layers, the gated network can generate dynamic fusion weights that match the features of the input data, and these weights can adaptively adjust according to the actual situation of different candidate sites.

[0155] In the output layer, the gating network performs a weighted sum of the objective fitness score and the group's subjective preference score based on the generated dynamic fusion weights. For each candidate location, its overall fitness score can be calculated using the following formula:

[0156]

[0157] Where S(i) is the comprehensive inadequacy score of candidate location i, ω(i) is the dynamic fusion weight, S_obj(i) is the objective fitness score of candidate location i, and S_sub(i) is the group subjective preference score of candidate location i.

[0158] Gated networks can comprehensively consider objective data and subjective preferences to obtain a more comprehensive and accurate overall inappropriateness score. They not only make full use of information from objective data and subjective preferences, but also improve the flexibility and accuracy of the assessment by adaptively adjusting weights.

[0159] This embodiment achieves a multi-dimensional and comprehensive evaluation of candidate sites by dynamically fusing objective fitness scores and subjective preference scores of the group. It not only overcomes the limitations of single data sources in traditional methods, but also effectively quantifies data uncertainty and decision-maker evaluation uncertainty by introducing Bayesian deep learning uncertainty modeling and probabilistic language group decision-making mechanisms, thereby effectively improving the scientificity and reliability of site selection evaluation.

[0160] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the auxiliary decision-making method for site selection of construction projects in urban planning. Any simple modifications based on this technical concept are within the scope of protection of this application.

[0161] This application also provides an auxiliary decision-making system for site selection of construction projects in urban planning. Please refer to [reference needed]. Figure 2 The auxiliary decision-making system for site selection of construction projects in urban planning includes:

[0162] The acquisition module 10 is used to acquire multi-source heterogeneous data of the city, construction project types, and site selection constraints.

[0163] Modeling module 20 is used to perform semantic modeling on the multi-source heterogeneous data of the city based on knowledge graph technology, and generate a knowledge graph of urban elements.

[0164] The retrieval module 30 is used to perform semantic retrieval in the urban element knowledge graph based on the construction project type and the site selection constraints to obtain a set of candidate sites that meet the site selection constraints.

[0165] Evaluation module 40 is used to evaluate each candidate site in the candidate site selection set using a dual-channel fusion scoring model of objective and subjective factors, and obtain a comprehensive fitness score corresponding to each candidate site. The dual-channel fusion scoring model of objective and subjective factors includes an objective data-driven channel and a subjective group preference channel. The objective data-driven channel and the subjective group preference channel dynamically adjust the output contribution ratio through a learnable gating network. The objective data-driven channel is used to output an objective fitness score, and the subjective group preference channel is used to output a group subjective preference score.

[0166] The optimization module 50 is used to iteratively optimize the candidate site selection set based on the comprehensive inadequacy score using a greedy heuristic algorithm to obtain the optimal site selection scheme.

[0167] The auxiliary decision-making system for site selection of construction projects in urban planning provided in this application adopts the auxiliary decision-making method for site selection of construction projects in urban planning in the above embodiments, which can solve the technical problems of low efficiency and poor quality of traditional site selection methods. Compared with the prior art, the beneficial effects of the auxiliary decision-making system for site selection of construction projects in urban planning provided in this application are the same as the beneficial effects of the auxiliary decision-making method for site selection of construction projects in urban planning provided in the above embodiments, and other technical features of the auxiliary decision-making system for site selection of construction projects in urban planning are the same as the features disclosed in the methods of the above embodiments, and will not be repeated here.

[0168] The above are only some embodiments of this application and do not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the content of this application specification and drawings, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A method for assisting decision-making in the site selection of construction projects in urban planning, characterized in that, The method includes: Acquire multi-source heterogeneous data on the city, construction project types, and site selection constraints; Based on knowledge graph technology, semantic modeling is performed on the multi-source heterogeneous data of the city to generate a knowledge graph of urban elements. Based on the project type and the site selection constraints, a semantic search is performed on the urban element knowledge graph to obtain a set of candidate sites that meet the site selection constraints; Each candidate site in the candidate site selection set is evaluated using a dual-channel fusion scoring model that integrates objective and subjective factors to obtain a comprehensive fitness score for each candidate site. The dual-channel fusion scoring model includes an objective data-driven channel and a subjective group preference channel. The objective data-driven channel and the subjective group preference channel dynamically adjust their output contribution ratios through a learnable gating network. The objective data-driven channel is used to output an objective fitness score, and the subjective group preference channel is used to output a group subjective preference score. The optimal location scheme is obtained by iteratively optimizing the candidate location set based on the comprehensive inadequacy score using a greedy take-away heuristic algorithm. The evaluation of each candidate site in the candidate site selection set is performed using a dual-channel subjective and objective scoring model to obtain a comprehensive inappropriateness score for each candidate site, including: Each candidate site in the candidate site set is input into the subjective and objective dual-channel fusion scoring model; Objective data features are extracted from each candidate site through the objective data-driven channel, and an objective fitness score is output based on the extracted objective data features. The subjective group preference channel is used to collect subjective group preference information for each candidate site, and a subjective group preference score is output based on the collected subjective group preference information. A learnable gating network is used to dynamically fuse the objective fitness score and the group's subjective preference score to obtain a comprehensive inappropriateness score for each candidate site.

2. The method as described in claim 1, characterized in that, The process of semantically modeling the multi-source heterogeneous data of the city based on knowledge graph technology to generate a knowledge graph of city elements includes: The urban multi-source heterogeneous data is spatiotemporally fused to obtain spatiotemporally fused urban multi-source heterogeneous data. The urban multi-source heterogeneous data includes at least urban operation data, socio-economic statistics, infrastructure data, environmental monitoring data, and social media data. Define the city ontology model; From the spatiotemporally fused urban multi-source heterogeneous data, entity instance data corresponding to the urban ontology model are extracted to obtain an initial entity set; The initial entity set is disambiguated and merged to obtain a deduplicated entity set; Based on the deduplicated entity set, relational reasoning is performed to obtain the set of relationships between entities in the urban element knowledge graph; A knowledge graph of urban elements is constructed based on the set of relationships between the entities.

3. The method as described in claim 1, characterized in that, The process involves semantic retrieval of the urban element knowledge graph based on the project type and the site selection constraints to obtain a set of candidate sites that meet the site selection constraints, including: Based on the project type and site selection constraints, a large language model finely tuned from the field of urban planning is used to analyze user intent and generate a project intent vector. Based on the project intent vector, a temporary ontology structure is dynamically expanded in the urban element knowledge graph to construct a context-specific knowledge subgraph. The urban space where the construction project is located is converted into multiple geographic analysis units. Multimodal features of each geographic analysis unit are extracted based on the urban element knowledge graph and the urban multi-source heterogeneous data. The multimodal features are then mapped to a joint embedding space. A query vector is generated based on the project intent vector, and a nearest neighbor search is performed in the joint embedding space based on the query vector to obtain a first candidate set. A formal logical query is performed on the knowledge subgraph based on the query vector to obtain a second candidate set; The first candidate set and the second candidate set are weighted and merged to obtain a merged candidate set; Acquire urban future development planning data, and perform counterfactual feasibility analysis on each candidate site in the fusion candidate set based on the urban future development planning data to obtain functional sustainability indicators of each candidate site within a future preset time window. The fusion candidate set is filtered according to the functional sustainability index to obtain a candidate site set that meets the site selection constraints.

4. The method as described in claim 1, characterized in that, The step of extracting objective data features from each candidate site through the objective data-driven channel and outputting an objective fitness score based on the extracted objective data features includes: Construct a causal graph of the city domain, and generate a causal adjacency matrix based on the causal graph of the city domain; A multi-domain teacher-student knowledge distillation architecture is constructed, wherein the multi-domain teacher-student knowledge distillation architecture consists of multiple teacher models and a shared student model. Each teacher model corresponds to a specific professional domain, and the student model is used to learn the knowledge of each teacher model. Based on the causal adjacency matrix and the teacher model, a knowledge distillation loss function for causal perception is designed. The student model is trained based on the knowledge distillation loss function to obtain the trained student model; The trained student model extracts objective data features related to each candidate site from the multi-source heterogeneous data of the city, and embeds the extracted objective data features into a preset feature space to obtain objective feature embedding. Based on the objective feature embedding, Bayesian deep learning uncertainty modeling is performed to obtain the objective fitness score corresponding to each candidate site selection.

5. The method as described in claim 1, characterized in that, The step of collecting subjective group preference information for each candidate site through the subjective group preference channel and outputting a subjective group preference score based on the collected subjective group preference information includes: The system receives subjective evaluation data on each candidate site from multiple decision-makers in the form of a probabilistic language terminology set, and determines the corresponding authority weight based on the professional title, years of experience, and degree of interest of each decision-maker. The subjective evaluation data is standardized and aligned with a language scale to obtain an aligned subjective evaluation data matrix, wherein the rows of the aligned subjective evaluation data matrix represent different decision-makers, the columns represent different candidate sites, and the matrix element values ​​are the subjective evaluation quantification values ​​given by each decision-maker for the corresponding candidate site. Construct a group consensus matrix based on the aligned subjective evaluation data matrix; The authority weights are dynamically adjusted based on the group consensus matrix to obtain the adjusted authority weight matrix. The group subjective preference score is calculated using a weighted probability language averaging operator based on the aligned subjective evaluation data matrix and the adjusted authority weight matrix.

6. The method as described in claim 4, characterized in that, The method utilizes a learnable gating network to dynamically fuse the objective fitness score and the group's subjective preference score to obtain a comprehensive inadequacy score for each candidate site, including: Obtain the causal interferability index for each candidate site. The causal interferability index is calculated based on the causal adjacency matrix and is used to characterize the intensity of the causal effect that may occur when intervening in the candidate site. Construct a learnable gated network, which includes an input layer, a hidden layer, and an output layer; The causal interventionability index, objective fitness score, and group subjective preference score of each candidate site are input into the input layer as input data. The input data is processed by a non-linear activation function in the hidden layer to generate dynamic fusion weights; In the output layer, the objective fitness score and the group subjective preference score are weighted and summed according to the dynamic fusion weight to obtain the comprehensive inappropriateness score corresponding to each candidate site.

7. The method as described in claim 1, characterized in that, The step of iteratively optimizing the candidate site selection set based on the comprehensive inadequacy score using a greedy heuristic algorithm to obtain the optimal site selection scheme includes: The candidate sites in the candidate site selection set are sorted in descending order according to the comprehensive inadequacy score to obtain the sorted candidate site selection sequence; Starting from the head of the sorted candidate address sequence, candidate addresses are selected and removed sequentially. Each time a candidate address is selected, it is determined whether the number of remaining candidate addresses in the candidate address set is less than the minimum number to retain or whether the proportion of candidate addresses currently being removed is less than the maximum removal proportion. If either condition is met, the removal operation is stopped, and a simplified candidate set is obtained. Based on the simplified candidate set, a local optimization search is performed to obtain the optimal addressing scheme.

8. The method as described in claim 7, characterized in that, The step of performing a local optimization search based on the simplified candidate set to obtain the optimal addressing scheme includes: Adjust the boundary range of each candidate site in the simplified candidate set or replace some geographical units to generate multiple neighborhood candidate sites; Calculate the comprehensive inappropriateness score for each neighborhood candidate site, select the neighborhood candidate site with the lowest comprehensive inappropriateness score to replace the original candidate site, and obtain a locally optimized simplified candidate set; Perform a swap search operation on the locally optimized and simplified candidate set, randomly select two candidate locations and swap their positions, and calculate the comprehensive inadequacy score after the swap. If the overall inadequacy score after the swap is lower than the overall inadequacy score before the swap, the swap operation is accepted, and the locally optimized simplified candidate set is updated. The swap search operation is repeated until the preset maximum number of iterations is reached or the locally optimized simplified candidate set has not been updated after several consecutive iterations, thus obtaining the target location set. The candidate site with the lowest overall inadequacy score in the target site selection set is determined as the optimal site selection scheme.

9. A decision support system for site selection of construction projects in urban planning, characterized in that, The system includes: The acquisition module is used to acquire multi-source heterogeneous data on the city, construction project types, and site selection constraints. The modeling module is used to perform semantic modeling on the multi-source heterogeneous data of the city based on knowledge graph technology, and generate a knowledge graph of urban elements. The retrieval module is used to perform semantic retrieval in the urban element knowledge graph based on the construction project type and the site selection constraints to obtain a set of candidate sites that meet the site selection constraints. The evaluation module is used to evaluate each candidate site in the candidate site selection set using a dual-channel fusion scoring model of objective and subjective factors, and obtain a comprehensive fitness score for each candidate site. The dual-channel fusion scoring model of objective and subjective factors includes an objective data-driven channel and a subjective group preference channel. The objective data-driven channel and the subjective group preference channel dynamically adjust their output contribution ratio through a learnable gating network. The objective data-driven channel is used to output an objective fitness score, and the subjective group preference channel is used to output a group subjective preference score. The optimization module is used to iteratively optimize the candidate site selection set based on the comprehensive inadequacy score using a greedy take-away heuristic algorithm to obtain the optimal site selection scheme. The evaluation module is further configured to input each candidate site in the candidate site set into the subjective and objective dual-channel fusion scoring model; extract objective data features from each candidate site through the objective data-driven channel, and output an objective fitness score based on the extracted objective data features; collect group subjective preference information from each candidate site through the subjective group preference channel, and output a group subjective preference score based on the collected group subjective preference information; and dynamically fuse the objective fitness score and the group subjective preference score using a learnable gating network to obtain a comprehensive inadequacy score corresponding to each candidate site.

Images