An industrial space adaptation optimization method and system based on a large model

Abstract

The application relates to an industry space adaptation optimization method and system based on a large model, which comprises the following steps: determining a target area, and obtaining a first data set and a second data set; constructing a digital twin base, and performing multidimensional analysis through a large language model; constructing an enterprise capability evaluation model and a migration risk early warning model based on the multidimensional analysis result, and obtaining enterprise capability information and migration risk grade information; obtaining interest point data and traffic network data, and determining geographic circle layer information; inputting the multidimensional analysis result, the enterprise capability information, the migration risk grade information and the geographic circle layer information into the large language model to generate an industry space optimization strategy; in summary, the application integrates multiple sources of data to construct a digital twin base, and uses a large language model to perform multidimensional analysis to generate an optimization strategy, solves the fragmentation problem in the traditional method, provides a decision basis, and has the effects of improving the scientific nature and operability of decision-making.

Description

Technical Field

[0001] This application relates to the technical field of industrial spatial planning, and in particular to a method and system for industrial spatial adaptation and optimization based on a large model. Background Technology

[0002] In the fields of industrial planning and regional economic development, traditional industrial spatial configuration analysis often relies heavily on human experience and fragmented data processing tools. Existing technologies typically employ geographic information systems (GIS) for spatial location visualization and combine them with statistical software for independent analysis of economic indicators. However, these existing technological models have significant limitations in handling multi-source heterogeneous data and in intelligent analysis.

[0003] First, at the data processing level, since industrial planning goals are usually in the form of unstructured natural language text, while the data reflecting the current situation is structured, the existing technology lacks effective natural language processing and semantic parsing technology. It is difficult to accurately transform the industrial development constraints and spatial configuration intentions in the planning text into quantitative indicators that can be understood and calculated by machines. As a result, the comparative analysis between the planning and the current situation can only rely on manual interpretation, making it difficult to achieve automated deviation identification. This is not only inefficient, but also prone to introducing subjective errors.

[0004] Secondly, at the analytical architecture level, existing technologies typically conduct isolated assessments of core analytical dimensions such as planning compliance, spatial economic efficiency, and enterprise innovation capabilities. This results in fragmented analytical processes and a lack of unified methods for multi-source data fusion and collaborative analysis. Such a fragmented technical architecture prevents systems and decision-makers from conducting comprehensive multi-dimensional assessments of industrial spatial configuration, thereby limiting the comprehensiveness and accuracy of analytical results. Summary of the Invention

[0005] To address the aforementioned shortcomings, this application provides a method and system for industrial space adaptation optimization based on a large model.

[0006] The above-mentioned objective of this application is achieved through the following technical solution: An industry space adaptation optimization method based on a large model includes the following steps: Identify the target area and obtain a first dataset and a second dataset for the target area. The first dataset includes industrial planning data and industrial layout data, and the second dataset includes park economic data and enterprise multidimensional data. A digital twin foundation is constructed based on the first and second datasets, and a multi-dimensional analysis of the digital twin foundation is conducted through a large language model. The multi-dimensional analysis includes planning compliance analysis and spatial efficiency analysis. Based on the results of multi-dimensional analysis, a capability assessment model and a migration risk early warning model for enterprises in the target region are constructed. By inputting multidimensional data of enterprises into the capability assessment model and the migration risk early warning model, enterprise capability information and migration risk level information can be obtained. Acquire point-of-interest data and transportation network data for the target area, and determine the geographical concentric circle information of the target area based on a preset range radius; By inputting multi-dimensional analysis results, enterprise capability information, migration risk level information, and geographical sphere information into the large language model, an industrial spatial optimization strategy is generated.

[0007] The second objective of this invention is achieved through the following technical solution: An industrial space adaptation and optimization system based on a large model includes: The data acquisition module is used to determine the target area and acquire a first dataset and a second dataset of the target area. The first dataset includes industrial planning data and industrial layout data, and the second dataset includes park economic data and enterprise multidimensional data. The multidimensional analysis module is used to construct a digital twin foundation based on the first dataset and the second dataset, and to perform multidimensional analysis on the digital twin foundation through a large language model. The multidimensional analysis includes planning compliance analysis and spatial efficiency analysis. The model building module is used to construct capability assessment models and migration risk early warning models for enterprises in the target region based on multi-dimensional analysis results. The enterprise information generation module is used to input multidimensional enterprise data into the capability assessment model and the migration risk early warning model to obtain enterprise capability information and migration risk level information. The geographic information determination module is used to acquire point of interest data and transportation network data of the target area, and determine the geographic concentric circle information of the target area based on a preset range radius; The optimization strategy generation module is used to input multi-dimensional analysis results, enterprise capability information, migration risk level information, and geographical layer information into the large language model to generate industrial spatial optimization strategies.

[0008] In summary, the industrial space adaptation and optimization method and system based on a large model provided in this application integrates multi-source data to construct a digital twin foundation and uses a large language model for multi-dimensional analysis to generate optimization strategies. This solves the fragmentation problem in traditional industrial space allocation methods and provides a systematic basis for decision-making, thereby improving the scientific nature and operability of decision-making. Attached Figure Description

[0009] Figure 1This is a flowchart of an embodiment of an industrial space adaptation and optimization method based on a large model according to this application; Figure 2 This is a schematic diagram of the interface for constructing a digital twin base and extracting multi-dimensional features of enterprises in an embodiment of an industrial space adaptation and optimization method based on a large model in this application. Figure 3 This is a schematic diagram of the interface for analyzing industrial spatial layout and agglomeration based on a digital twin platform in an embodiment of an industrial spatial adaptation and optimization method based on a large model according to this application. Detailed Implementation

[0010] The following is in conjunction with the appendix Figures 1-3 This application will be described in further detail.

[0011] In one embodiment, this application discloses an industry space adaptation optimization method based on a large model, such as... Figure 1 As shown, the specific steps include the following: S10: Determine the target area and obtain the first dataset and the second dataset of the target area. The first dataset includes industrial planning data and industrial layout data, and the second dataset includes park economic data and enterprise multidimensional data. In this embodiment, the target area refers to the specific geographical range targeted by the industrial spatial optimization analysis, such as cities, industrial parks, and economic development zones. The target area serves as the spatial carrier for data collection, model building, and strategy generation, and its boundaries are predefined by the analysis requirements. The first dataset includes industrial planning data and industrial layout data, encompassing strategic and spatial data related to industrial planning. The second dataset includes park economic data and multidimensional enterprise data, covering economic operation data and enterprise entity data. Industrial planning data refers to unstructured or semi-structured text data describing the industrial development goals, directions, and priorities of the target area, including industrial planning reports, regulations, and strategic outlines. Essentially, it is a natural language description of industrial development constraints and spatial configuration intentions. Industry layout data refers to data reflecting the actual spatial distribution of various industries within a target area. It is usually presented in the form of geographical coordinates, land use, and building area, and is used to characterize the current spatial base. Park economic data refers to economic operation indicators for specific parks within the target area, such as high-tech zones and economic development zones. These include tax revenue per unit area, rent levels, output value, vacancy rate, etc., and are used to quantitatively evaluate the economic efficiency of spatial units. Enterprise multidimensional data refers to a data set covering the multidimensional attributes of enterprises within the target area, including basic enterprise information, operating data, innovation capabilities, life cycle events, etc. Among them, basic enterprise information includes registered location and industry category, operating data includes revenue and tax payment, innovation capabilities include intellectual property output and R&D investment, and life cycle events include establishment, financing, and expansion.

[0012] Specifically, the target area is identified, and a first dataset and a second dataset for the target area are obtained. The first dataset includes industrial planning data and industrial layout data, while the second dataset includes park economic data and enterprise multidimensional data. In practice, the target area can be identified manually by selecting a box on the geographic information system interface, or by specifying it by entering the regional division code. The first and second datasets can be obtained in various ways. For example, industrial planning data can be extracted from planning texts downloaded from relevant official websites, industrial layout data can be obtained from enterprise registration information or satellite remote sensing imagery, park economic data can be collected from statistical yearbooks or reports from park management departments, and enterprise multidimensional data can be obtained from enterprise business registration information, recruitment platform data, or publicly available enterprise annual reports.

[0013] S20: Construct a digital twin foundation based on the first dataset and the second dataset, and conduct multi-dimensional analysis of the digital twin foundation through a large language model. The multi-dimensional analysis includes planning compliance analysis and spatial efficiency analysis. In this embodiment, the digital twin platform refers to a digital, visualized, and real-time interactive virtual mapping of a target area constructed by integrating multi-source heterogeneous data and utilizing technologies such as 3D modeling and virtual reality. The digital twin platform can dynamically reflect the current industrial status, spatial layout, and economic operation of the target area, providing a unified data view and operating platform for subsequent analysis and decision-making. The large language model refers to a computational model built based on deep learning technology, possessing massive parameters and natural language processing capabilities. The large language model can understand, generate, and analyze human language. In step S20, the large language model is used for semantic understanding and intent recognition of text information, and for correlation analysis and decision support of the digital twin platform constructed from multi-source heterogeneous data. Multi-dimensional analysis refers to a comprehensive analysis of the various types of data integrated in the digital twin platform. Multi-dimensional analysis includes planning compliance analysis and spatial efficiency analysis. Planning compliance analysis assesses the degree of matching and deviation between the actual industrial layout of the target area and the established industrial plan. Spatial efficiency analysis assesses the utilization efficiency, output benefits, and potential growth space of industrial spatial resources within the target area.

[0014] Before analysis using a large language model, spatial topology data and economic indicator data can be extracted from the digital twin base in advance, and converted into structured JSON text or descriptive prompts that can be read by the large language model through a serialization script, and then input into the large language model.

[0015] As an example, the large language model can be a general large model based on the Transformer architecture or a domain-specific fine-tuned large model. The input is structured text such as planning text, spatial indicators, enterprise data, and circle indicators. The output is semantic parsing results, planning evaluation conclusions, and industrial space optimization strategy text. The large language model uses the prompt project to limit the output format to structured JSON.

[0016] It should be noted that, for the large language model in the embodiments of this application, those skilled in the art can acquire, fine-tune, reproduce, and train the large language model based on the examples and content recorded in this application specification, combined with expert experience and historical conclusions.

[0017] Specifically, a digital twin foundation is constructed based on the first and second datasets, and a large language model is used to conduct multi-dimensional analysis of the digital twin foundation, including planning compliance analysis and spatial efficiency analysis. The construction of the digital twin foundation can be achieved by importing the acquired first and second datasets into a 3D geographic information platform, and forming a preliminary digital model of the target area through layer overlay and visualization techniques. When conducting multi-dimensional analysis of the digital twin foundation, the large language model can use preset keywords and rules to perform preliminary screening and classification of text in the industrial planning data to identify planning intentions related to spatial layout. By comparing the current industrial layout displayed in the digital twin foundation with the planning intentions, a planning compliance analysis is conducted. At the same time, based on the park's economic data, statistical methods can be used to calculate indicators such as output value per unit area and employment density to evaluate the utilization efficiency of spatial resources and achieve spatial efficiency analysis.

[0018] S30: Construct a capability assessment model and a migration risk early warning model for enterprises in the target region based on multi-dimensional analysis results; In this embodiment, the capability assessment model refers to a quantitative assessment system built based on multidimensional data of enterprises within the target area. The capability assessment model can objectively evaluate the comprehensive strength of enterprises and generate enterprise capability information, thereby providing enterprise-level support for industrial spatial allocation. The migration risk early warning model refers to predicting the possibility of enterprises migrating or leaving the target area in the future by analyzing the historical behavior data, operating status and external environmental factors of enterprises. The migration risk early warning model can identify enterprises with high migration risk and output migration risk level information, thereby providing early warning to regional management or decision-makers so that timely intervention measures can be taken.

[0019] Specifically, based on the results of multi-dimensional analysis, a capability assessment model and a migration risk early warning model for enterprises in the target region are constructed. The capability assessment model can be constructed by collecting qualitative evaluations of enterprises' innovation capabilities, market competitiveness, etc., through expert experience and transforming them into a scoring system. The migration risk early warning model can be constructed by analyzing enterprises' historical records of changes in registered address and abnormal business information, and setting various corresponding rules.

[0020] S40: Input the enterprise's multidimensional data into the capability assessment model and the migration risk early warning model respectively to obtain enterprise capability information and migration risk level information; In this embodiment, enterprise capability information refers to the output result after quantitatively assessing the comprehensive strength or core competitiveness of enterprises in the target area through a capability assessment model. It is usually a quantitative index or level used to characterize the enterprise's innovation capability, development potential, or comprehensive strength. It can help decision-makers identify leading enterprises, high-growth enterprises, or enterprises that need key support in the region, thereby providing a basis for industrial space optimization. For example, when formulating resource allocation, institutional support, or spatial layout strategies, enterprises with high enterprise capability information can be given priority. Migration risk level information refers to the risk assessment result output after predicting the possibility of enterprises relocating in the target area through a migration risk early warning model. It is usually a risk level classification or probability value. The risk early warning model makes judgments based on the enterprise's dynamic data.

[0021] Specifically, enterprise multidimensional data is input into both the capability assessment model and the migration risk early warning model to obtain enterprise capability information and migration risk level information. The enterprise multidimensional data can be in the form of structured tables, such as CSV files, and is imported in batches into the pre-built capability assessment model. The capability assessment model automatically generates enterprise capability information according to preset calculation logic. Similarly, enterprise multidimensional data can also be input into the migration risk early warning model in the same way. The migration risk early warning model outputs migration risk level information for each enterprise, such as low risk, medium risk, or high risk, according to its internal risk judgment rules.

[0022] S50: Acquire point of interest data and traffic network data of the target area, and determine the geographical concentric circle information of the target area based on the preset range radius; In this embodiment, point-of-interest (POI) data refers to location information within a target area that has specific functions or significance. It represents the geographical location of various facilities, services, or landmarks and is used to assess the service level of supporting facilities in the target area. By analyzing the distribution, density, and type of POIs, the convenience and resource accessibility of the project location can be determined. Furthermore, POI data includes commercial service points and their related data, public service points and their related data, transportation facilities and their related data, infrastructure and their related data, etc. Transportation network data refers to vector data describing the connectivity, attributes, and topological relationships of roads, rail lines, and other transportation routes within the target area, used to analyze accessibility. Furthermore, transportation network... Network data includes traffic road geometry information, traffic road attribute information, and traffic road topology; the preset radius refers to the distance value set in advance in spatial analysis to evaluate the surrounding environment of a certain point. The circular buffer zone obtained with this distance as the radius is the analysis range, which is usually used to quantitatively define the surrounding or service range; geographic concentric circle information refers to the spatial areas with different service functions and resource endowments, centered on industrial units within the target area, based on the preset radius and accessibility of supporting facilities. Among them, geographic concentric circle information can reflect the completeness of supporting services and transportation convenience around industrial units, providing a geographic location consideration for industrial spatial adaptation.

[0023] The preset radius can be calculated based on the historical traffic commuting data of the target area, using a road network isochronous circle algorithm to determine the commuting reachable boundary in 15 minutes or 30 minutes.

[0024] Specifically, the process involves acquiring point-of-interest (POI) data and transportation network data for the target area, and determining the geographic concentric circle information of the target area based on a preset radius. POI data can be obtained from publicly available online map service providers, for example, by acquiring location information of facilities such as restaurants, accommodations, education, and healthcare within a specific area via API interfaces. Transportation network data can be obtained from professional geographic data providers, including the geometry and hierarchy of roads. The geographic concentric circle information can be determined by setting a fixed buffer radius centered on each industry unit within the target area, and then counting the number of POIs within that buffer zone, thus forming the initial structure of the geographic concentric circle information.

[0025] S60: Input multi-dimensional analysis results, enterprise capability information, migration risk level information, and geographical layer information into the large language model to generate industrial spatial optimization strategies.

[0026] In this embodiment, the industrial space optimization strategy refers to a specific plan generated by a large language model based on multi-dimensional analysis results, enterprise capability information, migration risk level information, and geographical concentric circle information, which aims to improve the efficiency of industrial space allocation in the target area and promote industrial development. Specifically, the industrial space optimization strategy typically includes spatial layout suggestions, phased implementation paths, and expected results quantification. In step S60, the large language model is used to perform semantic understanding and intent recognition on the text information, and to perform correlation analysis and decision support based on multi-dimensional analysis results, enterprise capability information, migration risk level information, and geographical concentric circle information.

[0027] Specifically, the results of multi-dimensional analysis, enterprise capability information, migration risk level information, and geographical stratification information are input into the large language model to generate industrial spatial optimization strategies. These input information can be in the form of structured text summaries or key data points can be passed to the large language model via an API interface. After receiving the above input information, the large language model can generate a recommendation report on the industrial development of the target area based on its training knowledge and contextual understanding capabilities. This recommendation report may include a preliminary evaluation of the existing industrial layout and corresponding optimization directions.

[0028] For example, suppose a city planning department needs to optimize the industrial space at location A to promote the development of emerging industries and enhance regional economic vitality. In traditional methods, the planning department may face problems such as difficulty in quantifying planning documents, difficulty in integrating multi-source data, and insufficient assessment of enterprise development potential and risks, resulting in a lack of scientific rigor in the optimization strategy.

[0029] To this end, we first identified Location A as the target area and obtained the first and second datasets for Location A. The first dataset includes the industrial development plan text for Location A, such as "focusing on the development of strategic emerging industries such as artificial intelligence and biomedicine," as well as industrial layout data such as the geographical boundaries of existing industrial parks and the distribution map of resident enterprises. The second dataset includes the annual output value, tax revenue, and number of employees of each park in Location A, as well as multi-dimensional data of enterprises in Location A, such as registration information, intellectual property output, financing rounds, and changes in employee size.

[0030] Furthermore, a digital twin of location A is constructed based on the first and second datasets. This digital twin integrates geospatial information, industrial planning texts, economic indicators, and enterprise information to form a visualized virtual environment. A large language model is used to perform multi-dimensional analysis on this digital twin. Specifically, the large language model performs semantic understanding of the industrial planning text, identifies the intention to develop key industries such as "artificial intelligence" and "biomedicine," and compares this with existing industrial layout data in the digital twin to conduct planning compliance analysis, identifying which areas conform to the plan and which areas deviate. Simultaneously, the large language model combines park economic data to conduct spatial efficiency analysis, assessing the economic output efficiency and development potential of each industrial park. Through the above multi-dimensional analysis, this application can solve the problem that existing technologies' comparison analysis of planning and current status remains superficial and cannot deeply identify structural deviations.

[0031] Furthermore, based on the results of multi-dimensional analysis, a capability assessment model and a migration risk early warning model for enterprises in Location A are constructed. For example, the capability assessment model can evaluate the enterprise's innovation capabilities in the field of artificial intelligence based on indicators such as the enterprise's intellectual property output and technological coverage; the migration risk early warning model can predict the possibility of the enterprise's future migration based on time-series data such as the enterprise's financing rounds and fluctuations in employee size. By constructing the capability assessment model and the migration risk early warning model, the assessment at the enterprise level becomes more scientific, thereby making up for the shortcomings of traditional methods in enterprise assessment.

[0032] Furthermore, the acquired multidimensional enterprise data is input into the constructed capability assessment model and migration risk early warning model. After the capability assessment model and migration risk early warning model are run, they output specific capability information and migration risk level information for each enterprise. For example, enterprise B's innovation capability score in the field of artificial intelligence is 85 points, while the migration risk level information is, for example, enterprise C's migration risk is "medium". The enterprise-related information output by the capability assessment model and migration risk early warning model can provide support for subsequent optimization decisions.

[0033] Furthermore, data on points of interest for location A is obtained, such as the locations of surrounding schools, hospitals, and commercial centers, as well as transportation network data, such as road grades and bus routes. Based on a preset radius, the geographical concentric circle information of location A is determined by combining points of interest and transportation network data, with each industrial unit as the center. For example, an industrial park F may be assessed as having a geographical concentric circle of "convenient transportation and complete supporting facilities," while another park G may be assessed as having "inconvenient transportation and insufficient supporting facilities." Geographical concentric circle information can provide important locational considerations for industrial space adaptation.

[0034] Finally, the multi-dimensional analysis results, enterprise capability information, migration risk level information, and geographical layer information are input into a large language model, enabling the model to perform correlation analysis and decision reasoning to generate comprehensive industrial space optimization strategies. These strategies might include: prioritizing the location of artificial intelligence industries in areas with convenient transportation, complete supporting facilities, and high planning compliance; providing institutional support to biopharmaceutical companies with strong innovation capabilities but high migration risks, and guiding them to parks with less efficient spatial layouts; and upgrading infrastructure in areas not covered by planning to attract specific types of enterprises. Through this approach, the method achieves the fusion and collaborative analysis of multi-dimensional information, generating scientifically sound and actionable industrial space optimization strategies, thus effectively addressing the problems of isolated analysis models and insufficient decision-making basis in traditional methods.

[0035] Based on the above examples, compared to traditional methods that rely on human experience and decentralized data analysis tools, this method introduces a large language model to achieve deep semantic understanding and intent recognition of natural language text in industrial planning data. Specifically, in the example of location A, the large language model can identify the key development intentions of strategic emerging industries such as "artificial intelligence and biomedicine" in the plan, and quantitatively compare them with the actual industrial layout to generate planning compliance analysis results. This overcomes the semantic gap and data barriers between planning objectives and multi-source data in traditional methods, making the comparative analysis of planning and the current situation more automated and in-depth.

[0036] Furthermore, the digital twin platform constructed by this method effectively integrates the first and second datasets. During the optimization process of location A, all relevant information is uniformly integrated into the digital twin platform, and multi-dimensional analysis is performed through a large language model. This can solve the problem of fragmented evaluation and analysis dimensions in existing technologies. For example, during the generation of optimization strategies for location A, the large language model simultaneously considers the enterprise's innovation capabilities, migration risks, the spatial efficiency of the park, and the accessibility of surrounding supporting facilities, thereby generating a more comprehensive industrial space optimization strategy.

[0037] Therefore, this method can systematically assess the overall benefits and potential risks of industrial spatial allocation. Specifically, in the example of location A, high-potential and high-risk enterprises are identified through a capability assessment model and a migration risk warning model. Combined with geographic stratification information, corresponding spatial adaptation suggestions are provided for the identified enterprises. This method can improve the scientificity and basis of the final optimization decision and avoid the limitations of insufficient decision-making basis in the traditional fragmented analysis model.

[0038] In summary, this embodiment provides a technical solution to the problem of automating the computation and spatial mapping of multi-dimensional heterogeneous data by utilizing a three-dimensional geographic information system, a large language model, and a multi-source data fusion algorithm. It belongs to the category of technical solutions for data processing and spatial topology computation.

[0039] In one embodiment, step S20 includes: S21: Spatial registration and data fusion of the first and second datasets are performed and integrated into a 3D geographic information system to build a digital twin foundation; In this embodiment, step S21 aims to unify data from different sources and in different formats into a common spatial reference system and integrate it into a 3D geographic information system platform capable of 3D visualization and analysis. Spatial registration refers to transforming data from different coordinate systems or projections to align them in geographic space. Data fusion refers to integrating the aligned data to form a unified dataset. Integration into the 3D geographic information system enables the data after spatial registration and data fusion to be presented as a 3D model and supports complex 3D spatial analysis. For example, spatial registration can be performed using techniques such as projection transformation and geographic registration, and data fusion can be performed using methods such as attribute connection and spatial connection. The 3D geographic information system can be constructed using platforms such as ArcGIS Pro and QGIS3D.

[0040] Among them, such as Figure 2 The diagram illustrates the interface for constructing a digital twin foundation and extracting multidimensional features of enterprises. As shown, after spatial registration of the first and second datasets is completed, a physical building model of the target area is rendered on the front end of the 3D geographic information system, and the enterprise location mapping is completed through visual anchors with different labels. At the same time, the system interface displays a structured multidimensional data panel of the target enterprise and intuitively presents the microeconomic structure of the local area in the form of multidimensional charts. In particular, the system interface also displays time-series events in the enterprise's life cycle. This time-series data will be directly vectorized and input into the subsequent migration risk early warning model as a key underlying feature for predicting the enterprise's dynamic migration risk index. Through this interface-level feature aggregation, the accuracy and intuitiveness of multi-source heterogeneous data fusion can be improved.

[0041] S22: Using a large language model, semantic disambiguation and intent recognition are performed on natural language text in industrial planning data to extract industrial development constraints and spatial configuration intent; In this embodiment, step S22 aims to identify conditions that restrict industrial development and planning objectives that instruct spatial layout from unstructured text data. Semantic disambiguation refers to resolving the polysemy of words or phrases in the text, ensuring that the large language model can accurately understand their meaning in a specific context. Intent recognition refers to identifying the planning intent contained in the text, such as which industries are encouraged, which are restricted, and specific requirements for spatial layout. Furthermore, a large language model based on the Transformer architecture, such as the GPT series or BERT series, can be used for pre-training and fine-tuning to enable it to understand industrial planning texts, thereby achieving semantic disambiguation and intent recognition.

[0042] Among them, a large language model based on the Transformer architecture is used for pre-training and fine-tuning. The specific method of constructing the fine-tuning dataset includes: collecting historical industrial planning texts, which can be manually annotated by domain experts to form a large number of data pairs of input texts and structured constraint intentions. Through the data pairs obtained by the above process, the large language model is fine-tuned in a supervised manner, which enables the large language model to learn the mapping logic from natural language to spatial constraint rules.

[0043] Furthermore, during the inference phase, to ensure the accuracy of the large language model's output and avoid illusions, specific prompt word engineering templates can be used for semantic disambiguation and intent recognition. These prompt word templates include: system role settings, task instructions, few-sample examples, and the currently pending industry planning text. Through this structured input, the output boundaries of the large language model are limited, ensuring that the spatial configuration intent it extracts can be directly parsed and invoked by the 3D geographic information system.

[0044] S23: Perform spatial relationship calculation and consistency verification between the extracted industrial development constraints and spatial configuration intentions and the spatial elements in the digital twin base, and generate spatial conflict detection results and planning evaluation results; In this embodiment, step S23 aims to quantify and evaluate the degree of matching between planning intent and actual spatial elements through spatial analysis methods, and to identify potential conflicts. Spatial relationship calculation includes topological relationship analysis, distance analysis, buffer analysis, etc., to determine the spatial scope and impact involved in the planning intent. Consistency verification refers to comparing the spatial relationship calculation results with the extracted industrial development constraints and spatial configuration intent to determine whether there are any situations that do not meet the planning requirements. For example, GIS spatial analysis tools can be used to perform overlay analysis and proximity analysis on spatial elements such as land use type and infrastructure distribution in the digital twin base with the planning intent, thereby discovering planning coverage blind spots, layout deviation areas, or infrastructure gaps.

[0045] S24: Construct a multivariate spatial econometric function based on the economic data of the park, determine the weights of each indicator of the park's economic data using the entropy weight method, and conduct spatial autocorrelation analysis to generate spatial performance analysis results.

[0046] In this embodiment, step S24 aims to assess the spatial utilization efficiency and development potential of the industrial park from an economic perspective. A multivariate spatial econometric function is used to model the spatial dependencies between economic indicators of the park, considering the impact of geographical location on economic activities. The entropy weighting method refers to a weighting method that determines the weights of indicator data based on their dispersion, thereby avoiding subjective weighting. Spatial autocorrelation analysis is an analytical process used to detect the spatial clustering or dispersion patterns of economic indicators, thereby revealing the spatial heterogeneity of regional economic development. For example, a spatial lag model or spatial error model can be used as the multivariate spatial econometric function, and the entropy weighting method and spatial autocorrelation analysis can be implemented using Python's scipy library or R's spdep package.

[0047] Specifically, the proposed solution involves spatial registration and data fusion of the first and second datasets, integrating them into a 3D geographic information system to construct a multi-dimensional digital twin foundation, providing a spatial data basis for subsequent analysis. Based on this, a large language model is used to parse the natural language text in the industrial planning data, achieving semantic disambiguation and intent recognition. This accurately extracts industrial development constraints and spatial configuration intentions, overcoming the limitations of traditional manual interpretation of planning texts, which is inefficient and prone to errors. Spatial relationship calculations and consistency checks are performed between the extracted planning intents and the actual spatial elements in the digital twin foundation, effectively identifying spatial conflicts and evaluating the plan to ensure its scientific validity and feasibility. Simultaneously, by constructing a multivariate spatial econometric function and combining entropy weighting and spatial autocorrelation analysis to mine the park's economic data, a comprehensive assessment of the region's spatial efficiency can be achieved, revealing the spatial agglomeration and diffusion patterns of economic activities.

[0048] Through the above technical solutions, this application can overcome the limitations of traditional methods in processing heterogeneous data and unstructured planning texts and conducting multi-dimensional spatial analysis. Specifically, by constructing a digital twin foundation and combining the analytical capabilities of a large language model with spatial econometrics and autocorrelation analysis, this application can more accurately and comprehensively assess the planning compliance and spatial efficiency of the target area. This not only improves the efficiency and accuracy of the analysis but also lays the foundation for generating targeted and feasible industrial spatial optimization strategies, thereby achieving effective and scientific planning for regional industries.

[0049] In one embodiment, step S23 includes: S231: Based on spatial configuration intent extracted from a large language model, identify the priority of industrial spatial layout and the requirements for industrial agglomeration; In this embodiment, step S231 aims to clarify the priority industrial layout and the required spatial agglomeration level of these industries from the spatial configuration intent extracted from the large language model. The clarified industrial layout and its related information can directly reflect the strategic intent of industrial development. For example, by performing keyword matching, topic model analysis, or sentiment analysis on the text output by the large language model, the types of industries associated with words such as "key" and "core" can be identified, and the corresponding agglomeration requirements can be extracted by combining words such as "agglomeration," "cluster," and "scale." Alternatively, by constructing a rule-based expert system, the intent identified by the large language model can be compared with the predefined industrial development strategies to identify the spatial layout priority and agglomeration requirements of different industries.

[0050] S232: The priority of industrial spatial layout and the requirements for industrial agglomeration are quantified into spatial constraint rules, which include the minimum agglomeration unit area, the radiation radius of industrial correlation, and the infrastructure proximity threshold. In this embodiment, step S232 aims to transform abstract planning intentions into calculable and executable spatial constraints, which is key to achieving automated analysis and evaluation. The quantified spatial constraint rules can provide specific judgment criteria for subsequent spatial overlay analysis. For example, for the spatial constraint rule on the area of ​​the minimum agglomeration unit, the minimum land area or minimum enterprise density required for different industry types can be set; for the spatial constraint rule on the radiation radius of industry correlation, the maximum allowable spatial distance between different industries can be set according to the upstream and downstream relationships or synergistic effects of industries; for the spatial constraint rule on the infrastructure proximity threshold, the maximum acceptable distance between industrial units and infrastructure such as transportation hubs, energy supply points, and sewage treatment plants can be set; or, machine learning models can be used to combine historical success cases and expert experience to map the identified priorities and agglomeration requirements to the corresponding parameterized spatial constraint rules.

[0051] S233: Perform spatial topology analysis on the industrial layout data and industrial planning data in the digital twin base to generate spatial features of industrial layout characteristics and industrial planning characteristics. The spatial features include spatial overlap rate, proximity, and distribution dispersion. In this embodiment, step S233 aims to quantify the spatial matching degree and differences between industrial layout data and industrial planning data by analyzing the topological relationship between them, thereby providing basic data for subsequent consistency verification. Spatial topological relationship analysis is used to reveal the spatial correlation between industrial layout data and industrial planning data, that is, the spatial characteristics between industrial layout features and industrial planning features. Among them, industrial layout features are features extracted from industrial layout data, which are quantitative indicators used to describe the current spatial distribution of industries, such as spatial agglomeration, industrial proximity, land use efficiency, and spatial structure. Industrial planning features are features extracted from industrial planning data, which are used to describe the expected spatial distribution of industries in the future. Quantitative indicators for planning objectives include industrial zoning, development axes, and intensity requirements. Furthermore, spatial characteristics include spatial overlap, proximity, and dispersion. For example, spatial analysis tools in GIS software can be used to calculate the spatial overlap between industrial layout areas and planning areas to assess their spatial compatibility; the proximity between different industrial units can be calculated to reflect the degree of industrial agglomeration or dispersion; and the dispersion of industrial distribution can be calculated to reveal the trend of industrial concentration or dispersion. Alternatively, custom spatial algorithms, such as those based on quadtrees or R-tree indexes, can be used to query and calculate the topological relationships between large-scale spatial data and extract more complex spatial features, such as spatial connectivity and spatial isolation.

[0052] Among them, such as Figure 3 The diagram illustrates an interface for industrial spatial layout and agglomeration analysis based on a digital twin platform. During spatial topology analysis and basic layer information construction, a GIS spatial computing engine automatically identifies the physical boundaries of each industrial unit within the target area and generates polygonal agglomeration blocks of different industrial types on a 2D / 3D base map, such as traditional manufacturing and pharmaceutical manufacturing. Based on this, macroscopic indicators of a specific spatial base, such as land area, plot ratio, and dominant industry type, are extracted using data probes to calculate a multi-dimensional industrial agglomeration distribution matrix. The quantitative values ​​of current layout area, distance, and agglomeration ratio automatically extracted by the spatial engine are directly used as spatial feature vectors. These vectors are then input into a large language model for overlay comparison and consistency verification with industrial planning intentions, providing quantifiable computing power and data support for identifying planning coverage blind spots and layout deviation areas.

[0053] S234: Through spatial overlay analysis, the consistency between spatial constraint rules and spatial characteristics is verified, and blind spots in planning coverage, areas of layout deviation, and areas with infrastructure gaps are identified. In this embodiment, spatial overlay analysis refers to a method of overlaying multiple geographic layers to analyze the spatial relationships between them. In step S234, spatial overlay analysis is used to compare quantified spatial constraint rules with actual or planned spatial characteristics to identify areas that do not meet planning requirements or have resource shortcomings. For example, spatial constraint rules can be transformed into a series of spatial buffer zones or regions, and then overlaid with industrial layout feature and planning feature layers for analysis. If a planning area does not reach the minimum agglomeration unit area, it can be identified as a planning coverage blind spot. If an industrial layout area exceeds its correlation radiation radius, it can be identified as a layout deviation area. If an industrial area is more than the proximity threshold of infrastructure, it can be identified as an infrastructure gap area. Alternatively, a multi-criteria decision analysis method can be used to use spatial constraint rules and spatial features as different evaluation criteria, and generate a comprehensive compliance layer through weighted overlay. Then, based on a preset threshold, the above three types of areas are identified from the compliance layer.

[0054] S235: Generate a conflict detection report based on planning coverage blind spots, layout deviation areas, and infrastructure gap areas, including spatial conflict type, spatial location coordinates, and conflict intensity level, and use the conflict detection report as the spatial conflict detection result; In this embodiment, step S235 aims to structurally and quantitatively describe the identified problem areas and generate a conflict detection report, thereby providing an operational basis for subsequent decision support and optimization. The introduction of conflict intensity levels allows for the differentiation of the severity of different conflicts. For example, for each identified problem area, its geographical coordinate range, the type of industry involved, and the specific rules violated can be recorded, and different conflict intensity levels can be assigned according to the degree of violation. Alternatively, a pre-built conflict scoring model can be used to comprehensively quantify multiple conflict factors into a corresponding conflict intensity score and generate a detailed report.

[0055] Furthermore, as a preferred implementation, the pre-built conflict scoring model can be implemented using a multi-factor weighted penalty algorithm; specifically, for any identified conflict region... The formula for calculating its conflict intensity score includes: ,in, The number of types of conflict factors identified (such as insufficient area, excessive distance, and facility gaps). For the first Normalized values ​​for the degree of excess of various conflict factors, such as the proportion of actual area missing from the planned minimum cluster area; For the preset first The weighting coefficients for various conflict factors can be pre-set using the analytic hierarchy process (AHP) or expert scoring; specifically, As a veto penalty coefficient, when the conflict area involves inviolable bottom-line constraints such as hard ecological protection zones or basic farmland, The value should be either a maximum value or infinity, otherwise... The value can be 1; in this case, the conflict intensity score is calculated. Then, the conflict area can be mapped to different conflict intensity levels according to the preset discrete interval threshold.

[0056] S236: By combining the results of spatial conflict detection with a large language model, a land use development direction recommendation including a priority development sequence and a spatial optimization path is generated, and the land use development direction recommendation is used as the planning evaluation result.

[0057] In this embodiment, the large language model acts as an intelligent decision support in step S236. It can understand complex conflict reports and, combined with its inherent knowledge and reasoning capabilities, propose strategic and actionable land development suggestions. This transforms conflict detection results into specific planning actions. For example, the conflict detection report can be used as input, combined with preset planning principles, industrial systems, and regional development goals. Through the text generation capabilities of the large language model, detailed land development direction suggestions, including priority development order for specific plots, suggestions for industrial type adjustment, and infrastructure improvement plans, can be output. Alternatively, the conflict detection results can be associated with knowledge bases such as historical planning adjustment cases and successful industrial transformation experiences. The knowledge graph reasoning capabilities of the large language model can be used to generate feasible spatial optimization path suggestions.

[0058] In step S236, the process of generating land development direction suggestions by combining the spatial conflict detection results with the large language model can be achieved using contextual retrieval-enhanced generation technology. Specifically, the conflict detection report generated in step S235 is converted into a serialized text description and combined with preset planning principles and regional development goals to form contextual prompts. Simultaneously, before generating suggestions, the large language model vectorizes the current spatial conflict features and retrieves the top K most similar historical resolution cases from a pre-established local historical planning adjustment case knowledge base, such as historical industrial land replacement schemes or cross-regional infrastructure sharing schemes. Subsequently, the large language model integrates the input prompts and retrieved historical cases, and outputs suggestions according to a preset logical framework through autoregressive decoding. This preset logical framework includes: a priority development sequence (specific plot IDs arranged in reverse order of conflict intensity), spatial optimization paths (including industrial relocation instructions, industrial relocation instructions, and land use change suggestions), and expected performance assessment. Through the aforementioned contextual retrieval-enhanced generation technology, the planning assessment results generated by the large language model can be made more technically traceable.

[0059] Specifically, the solution proposed in this application combines the understanding of planning intent through a large language model with spatial analysis techniques to achieve a complete process from abstract intent to quantitative rules, from current situation analysis to conflict identification, and then to decision-making recommendations. This method not only accurately identifies inconsistencies between planning and actual layout, but also further quantifies the severity of conflicts. Furthermore, it utilizes the reasoning capabilities of the large language model to provide decision-makers with forward-looking and actionable land development directions, thereby effectively addressing the shortcomings of traditional planning assessments in delving into the essence of conflicts and providing intelligent optimization suggestions.

[0060] Through the above technical solution, this application can transform abstract industrial planning intentions into specific and quantifiable spatial constraint rules. By combining the actual spatial data in the digital twin platform for topological relationship analysis and overlay verification, it can accurately identify conflicts between the planning and the current situation and present them in the form of a structured conflict detection report. This report specifically includes the conflict type, location, and intensity level. More importantly, by analyzing the conflict detection results through a large language model, it can generate strategic and actionable land development direction suggestions, including priority development sequences and spatial optimization paths. This not only solves the problem that traditional planning assessments are unable to delve into the essence of conflicts and provide intelligent optimization suggestions, but also enhances the scientific nature and feasibility of industrial spatial planning.

[0061] In one embodiment, step S24 includes: S241: Construct an economic indicator matrix from the park's economic data, and perform spatial interpolation to complete the missing data in the economic indicator matrix; In this embodiment, step S241 aims to organize the scattered economic data of the park into a structured matrix form and solve the problem of data incompleteness through spatial interpolation. The economic indicator matrix can include data from multiple dimensions such as output value, tax revenue, number of employees, R&D investment, and output per unit of land. For spatial interpolation of missing data, various methods can be used. For example, the inverse distance weighting method can be used to estimate the value of unknown points based on the values ​​of known points and distances, ensuring the spatial continuity of the data. Alternatively, Kriging interpolation can be used to consider the spatial autocorrelation of the data, thereby providing a more accurate estimate of missing values.

[0062] S242: The entropy weight method is used to calculate the weight coefficients of each economic indicator in the economic indicator matrix, and a weighted comprehensive evaluation function is generated. In this embodiment, step S242 aims to objectively determine the importance of each economic indicator in the comprehensive evaluation. Among them, the entropy weight method refers to an objective weighting method based on information entropy, which determines the weight according to the degree of variation of the indicator value. The greater the degree of variation of the indicator, the greater the amount of information it provides, and the higher the weight, thereby reducing the bias that may be caused by subjective weighting. The weighted comprehensive evaluation function refers to multiplying the standardized value of each indicator with its corresponding weight coefficient and summing them to obtain the comprehensive evaluation score.

[0063] S243: Global spatial autocorrelation analysis is performed using a predefined first autocorrelation algorithm to examine the spatial clustering of economic indicators in the economic indicator matrix, and local spatial autocorrelation analysis is performed using a predefined second autocorrelation algorithm to identify spatial clustering patterns. In this embodiment, step S243 aims to comprehensively reveal the spatial distribution characteristics of economic indicators. The predefined first autocorrelation algorithm can use the global Moran index to assess whether there are significant spatial clustering or dispersion phenomena of economic indicators in the entire target area. The predefined second autocorrelation algorithm can use the local Moran index to identify specific spatial clustering patterns, such as high-value clustering, low-value clustering, and spatial outliers, thereby providing a more refined spatial distribution insight.

[0064] S244: The weighted comprehensive evaluation function is coupled with the spatial clustering model, a continuous spatial efficiency surface is generated by Kriging interpolation, and the efficiency level of the continuous spatial efficiency surface is divided to generate spatial efficiency analysis results.

[0065] In this embodiment, step S244 aims to combine the comprehensive evaluation results with spatial distribution characteristics and present them in an intuitive and continuous manner. The coupling process uses the score of the weighted comprehensive evaluation function as input for Kriging interpolation, while considering spatial autocorrelation patterns as covariates for Kriging interpolation, to generate a surface reflecting the continuous changes in regional economic efficiency. Furthermore, the Kriging interpolation method can predict the attribute values ​​of unknown points based on the spatial distribution and attribute values ​​of known points, generating a smooth and statistically significant continuous surface. The efficiency level classification of the continuous spatial efficiency surface can be achieved using the natural breakpoint method or the equal interval method, specifically by dividing the continuous efficiency values ​​into several discrete levels to generate spatial efficiency analysis results for easy understanding and application.

[0066] Specifically, the proposed solution ensures data integrity and analytical reliability by spatially interpolating and completing missing data in the park's economic data; it objectively determines the weights of each economic indicator using the entropy weight method, avoiding biases in subjective judgment, thus enabling the weighted comprehensive evaluation function to more accurately reflect the park's overall economic efficiency; furthermore, through global and local spatial autocorrelation analysis, it can not only perceive the overall spatial clustering of economic indicators but also identify specific spatial hotspots and cold spots, providing a foundation for accurate spatial efficiency assessment; finally, it couples the weighted comprehensive evaluation results with spatial clustering patterns and uses Kriging interpolation to generate a continuous spatial efficiency surface, followed by efficiency level classification, making the spatial efficiency analysis results more intuitive and continuous.

[0067] Through the above technical solutions, this application can effectively solve the problems of missing economic data in industrial parks, overly subjective determination of indicator weights, and insufficient precision in identifying spatial agglomeration patterns. Specifically, data interpolation can improve the completeness and accuracy of the analysis, while the entropy weight method can enhance the objectivity and scientific nature of the evaluation results. The combination of global and local spatial autocorrelation analysis can comprehensively reveal the spatial distribution characteristics of economic efficiency. Furthermore, by coupling and Kriging interpolation, the discrete evaluation results are transformed into a continuous and intuitive efficiency surface, and a grade classification is performed, making the spatial efficiency analysis results more refined, reliable, and easy to understand.

[0068] In one embodiment, step S30 includes: S31: Based on the industrial layout characteristics and spatial efficiency characteristics in the multi-dimensional analysis results, select target enterprises within the target area; In this embodiment, step S31 aims to identify enterprises that have a key impact on regional industrial development or have specific optimization potential based on the industrial layout characteristics and spatial efficiency characteristics obtained from multi-dimensional analysis of the digital twin platform. For example, enterprises that are highly compatible with the region's leading industries or those that show high growth potential but low space utilization efficiency within a specific spatial area can be selected. Among them, industrial layout characteristics refer to quantitative indicators extracted from the current status data of the target area to describe the actual spatial distribution of various industries, such as the degree of agglomeration and spatial distribution patterns of different industries. Spatial efficiency characteristics refer to the distribution characteristics of economic output efficiency calculated spatially by integrating economic indicators such as per-unit tax revenue, rent, output value, and vacancy rate. Furthermore, high-efficiency areas and low-efficiency areas can be their core characteristics to identify enterprise clusters that need to be focused on from the perspective of economic benefits.

[0069] S32: Filter out the multidimensional data of the target company, and build enterprise capability assessment indicators and enterprise risk characteristic library based on the multidimensional data of the target company; In this embodiment, step S32 aims to collect comprehensive multidimensional data on the selected target enterprises and extract key indicator features for assessing their capabilities and risks. The enterprise capability assessment indicator system can be constructed based on the target enterprise's intellectual property output, specifically including the number of patents, patent citation frequency, technical field coverage, and patent growth indicators. It can also further include R&D investment, the number of R&D personnel, and high-tech enterprise certification. The enterprise migration risk feature database can be constructed based on the target enterprise's full lifecycle ledger data, specifically including the enterprise's establishment time, financing rounds, fluctuations in employment scale, penalty records, and a sequence of events related to capacity expansion. It can also include changes in market share, major customer attrition, and supply chain stability.

[0070] S33: Calculate the comprehensive innovation capability score of each target enterprise using the entropy weight method, and construct an enterprise capability assessment model; In this embodiment, step S33 aims to utilize the entropy weight method, an objective weighting method, to determine the weights of various innovation capability indicators based on the data dispersion, thereby avoiding the influence of subjective factors on the evaluation results. The entropy weight method reflects the amount of information in the indicators; the greater the information content, the higher the weight. Through weighted summation, the comprehensive innovation capability score of each target enterprise can be obtained, thus forming a quantitative enterprise capability evaluation model. Furthermore, in addition to the entropy weight method, other objective weighting methods such as principal component analysis and factor analysis can also be used.

[0071] S34: Establish a full lifecycle ledger record based on the enterprise risk characteristic database, wherein the full lifecycle ledger record includes the key event sequence of the target enterprise; In this embodiment, step S34 aims to organize various key time-series data collected from the enterprise risk feature database into a detailed full lifecycle ledger record in chronological order; wherein, the collected key time-series data may include enterprise registration, financing, major investment, senior management changes, product launch, market expansion, institutional subsidies, environmental penalties, labor disputes, etc., which together constitute the complete trajectory of enterprise development.

[0072] S35: Based on the full lifecycle ledger records, a migration risk early warning model is constructed using the random forest algorithm.

[0073] In this embodiment, step S35 aims to use the random forest algorithm to learn and analyze the established full lifecycle ledger records. The random forest algorithm can effectively handle high-dimensional data and nonlinear relationships by constructing multiple decision trees and combining their prediction results, thereby identifying key event patterns and temporal characteristics that lead to enterprise migration. The migration risk warning model can predict the probability or risk level of future migration based on the current event sequence of the enterprise. Furthermore, in addition to the random forest algorithm, machine learning algorithms such as gradient boosting trees or recurrent neural networks can also be used.

[0074] Specifically, the proposed solution first identifies key target companies based on multi-dimensional analysis results to ensure the relevance of subsequent analyses; then, it constructs enterprise capability assessment indicators and enterprise risk characteristic databases for these target companies, providing a data foundation for comprehensive evaluation; further, it objectively quantifies innovation capabilities using the entropy weight method to avoid subjective bias and make enterprise capability assessment more scientific and reliable; simultaneously, it transforms enterprise risk characteristics into key event sequences in the full lifecycle ledger, providing time-series data for capturing dynamic changes in enterprises; finally, it utilizes the random forest algorithm to construct a migration risk early warning model, which can effectively identify potential risks of enterprise migration.

[0075] Through the above technical solution, this application can achieve a high-precision assessment of the capabilities and migration risks of enterprises within the target area. This not only solves the problems of strong subjectivity and insufficient data utilization in traditional assessment methods, but also makes the formulation of industrial space optimization strategies more forward-looking and targeted by constructing a dynamic migration risk early warning model.

[0076] In one embodiment, step S35 includes: S351: Extract time-series feature vectors associated with key event nodes of a target enterprise based on key event sequences; In this embodiment, step S351 aims to transform the discrete key event sequence in the enterprise's full lifecycle ledger records into numerical features that can be processed by machine learning models. The key event sequence records important events and behaviors of the enterprise from its inception to its current state, such as time-series events corresponding to financing rounds, fluctuations in workforce size, and capacity expansion. Furthermore, the extraction of the time-series feature vector can be achieved through a sliding time window method, statistically analyzing and encoding events within a specific time period before and after each key event node. For example, the frequency of events, event type distribution, and event intensity can be statistically analyzed to form the time-series feature vector corresponding to that key event node. Alternatively, models such as Long Short-Term Memory networks in deep learning can be used to directly encode the key event sequence, learn and extract time-dependent feature representations, and use them as time-series feature vectors.

[0077] S352: Obtain historical migration enterprise samples as the training set, use time-series feature vectors as input features, and use enterprise migration status as the prediction label. Train the classification model through the random forest algorithm to obtain the initial model. In this embodiment, step S352 aims to train a classification model using known historical data through supervised learning. The historical migration enterprise samples refer to enterprise data that has or has not migrated in the past, with their migration status serving as the prediction target of the classification model. The random forest algorithm is an ensemble learning method that improves the accuracy and generalization ability of the model by constructing multiple decision trees and combining their prediction results. Using the historical migration enterprise samples as the training set, historical enterprise migration status information can be obtained from public or authorized data sources such as official enterprise registration change records, annual business reports, and tax relocation records. This information is then combined with the enterprise's full lifecycle ledger records to construct a training dataset containing time-series feature vectors and migration status labels. Furthermore, during training, techniques such as cross-validation can be used to evaluate the model's performance, and the parameters of the random forest algorithm, such as the number of decision trees, the maximum depth of each decision tree, and the feature sampling ratio, can be adjusted based on the evaluation results to obtain a better-performing initial model.

[0078] S353: During the training of the classification model, nodes are split according to the principle of minimizing the Gini coefficient to generate several decision trees, which constitute an ensemble learning model; In this embodiment, step S353 describes the decision tree construction mechanism in the random forest algorithm. The Gini coefficient is an indicator of dataset purity; a smaller value indicates higher purity. When splitting at each node of the decision tree, the training algorithm selects a feature and its split point to minimize the sum of the Gini coefficients of the child nodes after splitting, thereby minimizing uncertainty and improving classification accuracy. By repeating this process, multiple decision trees with different structures and decision rules can be generated, collectively forming an ensemble learning model. When constructing each decision tree, a bootstrap sampling method can be used to randomly extract samples from the training set and randomly select a subset of features for node splitting, increasing the diversity of the decision trees and reducing the risk of model overfitting. Furthermore, the decision tree generation process can be terminated under certain conditions, such as stopping splitting when the number of samples in a node is below a preset threshold, the Gini coefficient decreases less than a preset tolerance, or a preset maximum depth is reached, to control model complexity and computational cost.

[0079] S354: The hyperparameters of the initial model are optimized based on the ensemble learning model and the preset parameter optimization strategy, and the output of the initial model is calibrated by the preset confidence threshold to obtain the migration risk warning model. In this embodiment, step S354 aims to further improve the predictive performance and reliability of the initial model. Hyperparameter optimization refers to finding the optimal parameter combination that makes the initial model perform best on the validation set by searching and adjusting the external parameters of the initial model. Confidence calibration refers to adjusting the original probability values ​​output by the initial model to more accurately reflect the true probability and improve the credibility of the initial model's prediction results. Furthermore, parameter optimization strategies can employ methods such as grid search, random search, or Bayesian optimization to find the optimal solution in a predefined hyperparameter space. Confidence calibration can be achieved through post-processing methods such as Platt scaling and isotonic regression to map the original output probability of the initial model to a range closer to the true probability.

[0080] S355: Input the time-series feature vector of the target enterprise into the migration risk early warning model, output the enterprise migration risk probability, and assign risk level based on the preset risk threshold.

[0081] In this embodiment, step S355 describes the actual application process of the migration risk warning model; wherein, the migration risk warning model calculates the probability of the enterprise's migration behavior based on the input enterprise time series feature vector; to facilitate understanding and decision-making, the output enterprise migration risk probability is further converted into discrete risk levels; wherein, the risk threshold can be set according to actual business needs and risk tolerance, and similarly, the allocation of risk levels can also be combined with expert experience or historical data analysis to dynamically adjust the risk threshold to adapt to the migration characteristics of enterprises in different industries, different regions or different economic cycles.

[0082] Specifically, the system extracts time-series feature vectors associated with each event node from the key event sequence of the target enterprise, thereby capturing the dynamic changes and potential trends in the enterprise's development process. Using these time-series feature vectors, combined with historical migration enterprise samples as the training set and their migration status as prediction labels, a classification model is trained using a random forest algorithm. During training, node splitting is performed based on the Gini coefficient minimization principle, generating multiple decision trees and forming an ensemble learning model, thus effectively handling nonlinear relationships and improving prediction robustness. On this basis, the hyperparameters of the ensemble learning model are optimized, and the model output is calibrated using a confidence threshold to ensure the accuracy and reliability of the model's prediction results, ultimately yielding a migration risk warning model. The time-series feature vectors of the target enterprise are input into the migration risk warning model, which outputs the enterprise migration risk probability and converts it into an intuitive risk level based on a preset risk threshold.

[0083] Through the above technical solutions, this application can effectively solve the problems of insufficient utilization of time-series information and low prediction accuracy of models in enterprise migration risk assessment. Specifically, by extracting the time-series feature vectors of key event sequences of enterprises, subtle changes and potential risk signals in the dynamic development process of enterprises can be captured. By introducing the random forest algorithm combined with the principle of minimizing the Gini coefficient for node splitting, as well as subsequent hyperparameter optimization and confidence calibration, the prediction accuracy and generalization ability of the migration risk early warning model can be improved, enabling it to more reliably identify enterprises with migration tendencies. In addition, the enterprise migration risk probability output by the migration risk early warning model is converted into a risk level, which facilitates decision-makers' understanding and application of enterprise migration risk. Based on the construction of enterprise capability assessment models and migration risk early warning models, more accurate and dynamic enterprise migration risk assessment results can be provided, thereby effectively improving the scientificity and foresight of industrial space adaptation.

[0084] In one embodiment, step S353 includes: (a) Sample the temporal feature vectors in the training set to generate several training subsets; In this embodiment, sampling refers to the process of selecting a portion of data samples from the original training set to construct a new training dataset. Its purpose is to introduce data diversity, enabling the subsequently generated decision trees to learn different data patterns, thereby reducing the risk of model overfitting and improving the generalization ability of the ensemble model. Furthermore, the sampling methods include sampling with replacement, where samples are randomly drawn from the original training set and replaced after each draw, allowing the same sample to be drawn multiple times; or sampling without replacement, where a portion of samples is randomly drawn from the original training set, and each sample is drawn only once. Through these sampling methods, a slightly different training view can be provided for each decision tree, thereby enhancing the robustness of the ensemble model.

[0085] (b) Construct the corresponding decision tree based on each training subset, and calculate the information gain of all features when splitting nodes. Select the feature with the largest decrease in Gini coefficient and its corresponding threshold as the optimal split point. In this embodiment, a decision tree refers to a tree-structured model used for classification or regression tasks. Its construction process involves recursively partitioning the data based on features, starting from the root node. Node splitting refers to the process of dividing the dataset into two or more subsets at each node of the decision tree based on the value of a certain feature. Information gain is an indicator that measures the effectiveness of a feature in partitioning the dataset, representing the degree to which the uncertainty of the dataset is reduced after a given feature is selected. The Gini coefficient is an indicator that measures the purity of the dataset; the smaller the value, the higher the purity of the dataset. The feature with the largest decrease in the Gini coefficient and its corresponding threshold are selected as the optimal split point, that is, the feature and split point that maximize the improvement in the purity of the subsets after partitioning are selected, so that the decision tree can minimize uncertainty and improve classification accuracy at each split.

[0086] (c) Generate a decision tree by recursively splitting until a preset termination condition is reached. The termination condition includes the number of node samples being lower than a preset sample number threshold or the decrease in the Gini coefficient being less than a preset tolerance. In this embodiment, recursive splitting refers to the iterative and top-down process of building the decision tree. Specifically, it involves continuously splitting nodes starting from the root node until a specific termination condition is met. The termination condition is set to prevent the decision tree from overgrowing, thereby avoiding overfitting. The termination condition includes the number of samples in a node falling below a preset threshold. For example, when a node contains too few samples, continuing to split may cause the model to become overly sensitive to noise, so splitting is stopped. Additionally, the termination condition also includes the decrease in the Gini coefficient being less than a preset tolerance. For example, when splitting cannot significantly improve the purity of the subset, i.e., the decrease in the Gini coefficient is not significant, splitting is stopped, and the node is considered sufficiently pure or can no longer be effectively split.

[0087] (d) Repeat steps (a), (b), and (c) in sequence to generate a preset number of decision trees, and integrate the prediction results of each decision tree to form an ensemble learning model.

[0088] In this embodiment, repeating the above sampling, decision tree construction, and recursive splitting processes can generate multiple independent decision trees. The preset number of decision trees refers to the number determined based on actual needs and computing resources. Generally, a larger number of decision trees results in more stable performance of the ensemble model, but also higher computational costs. Integrating the prediction results of each decision tree means summarizing the prediction results of multiple decision trees to obtain the final prediction output. Furthermore, for classification problems, a voting mechanism is typically used, such as majority voting, where the category with the most votes is selected as the final prediction result. For regression problems, an averaging method can be used.

[0089] Specifically, the proposed solution optimizes the training process of the migration risk early warning model by constructing an ensemble learning model. Specifically, it samples the temporal feature vectors from the original training set to generate multiple different training subsets, ensuring that each decision tree encounters slightly different data views during training. This increases the diversity among the decision trees and prevents all decision trees from learning the same pattern, thus avoiding insufficient model generalization. Based on these diverse training subsets, separate decision trees are constructed. During the construction of each decision tree, the principle of minimizing the Gini coefficient is used for node splitting; that is, at each splitting step, the node with the highest Gini coefficient is selected. To improve data purity, features and thresholds are set to enable each decision tree to efficiently learn classification rules from the data. Simultaneously, to prevent individual decision trees from overfitting the training data, a threshold for the number of node samples and a tolerance for the decrease in the Gini coefficient are set as termination conditions. This ensures that each decision tree maintains a certain level of complexity while avoiding overlearning from noise. Finally, by repeating the above process to generate a predetermined number of decision trees, and integrating the prediction results of these decision trees, the bias and variance that may exist in a single decision tree can be effectively reduced, thereby significantly improving the overall prediction accuracy and robustness of the migration risk early warning model.

[0090] By employing the aforementioned technical solutions, when constructing a migration risk early warning model, sampling the training data and building multiple decision trees based on the principle of minimizing the Gini coefficient, and then integrating the prediction results of these multiple decision trees, the generalization ability of the migration risk early warning model can be significantly enhanced. The decision tree construction method described above can effectively avoid the overfitting problem that may occur with a single decision tree, enabling the migration risk early warning model to better adapt to unseen data. Simultaneously, by integrating the judgments of multiple decision trees, the variance of the migration risk early warning model's predictions can be effectively reduced, improving its stability and accuracy.

[0091] In one embodiment, step S50 includes: S51: Obtain point of interest data for the target area, wherein the point of interest data includes commercial service information, public service information, transportation facility information, and infrastructure information; In this embodiment, step S51 aims to provide detailed location and attribute information of various supporting facilities within the target area, providing basic data support for assessing the service environment surrounding the industrial unit. Among them, point of interest data can be obtained through publicly available geographic information service platforms, by using web crawling technology or API interface calls to download POI data in batches within a specified area; or, it can be purchased or obtained through official departments or professional data service providers. This point of interest data is usually cleaned and classified, and contains more detailed attribute fields.

[0092] S52: Obtain traffic network data for the target area and construct a traffic network topology structure based on the traffic network data, including road grades, traffic capacity, and topological relationships; In this embodiment, step S52 aims to organize discrete traffic data into a graph structure that can be used for path analysis and accessibility calculation. Its function is to accurately simulate traffic flow and provide a basis for evaluating the traffic convenience of industrial units. The traffic network data can be obtained from open street maps or basic geographic information centers, and processed using GIS software to extract information such as road grade, number of lanes, and speed limit to construct topological relationships. Alternatively, detailed real-time or historical data such as traffic flow and congestion index can be obtained from traffic management departments, and combined with road network data, a dynamic traffic network topology structure can be constructed through graph databases or specialized traffic simulation software.

[0093] S53: Identify all industrial units within the target area, and based on the preset range radius, generate equidistant buffer zones centered on each industrial unit to form basic concentric circle information; In this embodiment, step S53 aims to initially define the direct impact range of each industrial unit as the starting point for subsequent analysis. The industrial unit can be identified by using GIS software in conjunction with industrial layout data to identify the geometric center point or boundary of the industrial unit. Then, a buffer analysis tool is used to generate a circular or polygonal area with a specified radius based on each center point or boundary. Alternatively, image recognition technology can be used to analyze high-resolution satellite imagery to automatically identify industrial building clusters or factory areas and use them as industrial units to generate buffer zones.

[0094] S54: Based on point-of-interest data, identify several types of supporting service information and their corresponding supporting facility locations, and calculate the spatial coverage of various supporting service information. In this embodiment, step S54 aims to quantify the distribution density and accessibility of different types of services within the target area, reflecting the service level of the area. The supporting service information can be obtained by classifying and filtering the acquired point-of-interest (POI) data. For example, POIs categorized as "hospitals," "schools," or "banks" can be identified as supporting facilities. Then, spatial analysis methods such as kernel density analysis or Thiessen polygon analysis can be used to calculate the spatial distribution density of these supporting facilities within the area, thereby obtaining spatial coverage. Alternatively, population density data can be combined to weight the service radii of different supporting facilities and calculate their effective coverage area ratio for surrounding industrial units.

[0095] S55: Calculate the optimal path impedance from each industrial unit to the location of supporting facilities based on the transportation network topology, and evaluate the facility accessibility index of the basic concentric circle information based on the optimal path impedance. In this embodiment, step S55 aims to overcome the limitation that equidistant buffer zones cannot reflect actual traffic conditions, so as to more accurately measure the convenience of industrial units in obtaining supporting services. The optimal path impedance can be calculated using the network analysis module in GIS software, taking the industrial unit as the starting point and the location of supporting facilities as the ending point, based on attributes such as road grade and traffic capacity in the traffic network topology, and the path length or time is used as the optimal path impedance. Alternatively, real-time traffic data can be introduced, and the dynamic optimal path impedance can be calculated through a traffic simulation model, combined with time window constraints, to evaluate the accessibility of facilities in different time periods.

[0096] S56: Optimize basic concentric circle information based on spatial coverage and facility accessibility indicators to generate geographic concentric circle information.

[0097] In this embodiment, step S56 aims to refine the simple equidistant buffer zone into a geographical range that reflects actual service capabilities and transportation accessibility, providing accurate geographical dimension information for subsequent industrial spatial adaptation. Optimizing the basic concentric circle information can be achieved by weighting and superimposing spatial coverage and facility accessibility indicators to form a comprehensive geographic service index. Based on this geographic service index, the initial buffer zone for each industrial unit can be adjusted. For example, the buffer zone for high service index areas can be appropriately expanded, while the buffer zone for low service index areas can be appropriately reduced or its shape adjusted to reflect its true service range. Alternatively, cluster analysis or machine learning algorithms can be used to classify industrial units using spatial coverage and facility accessibility indicators as features, and different geographic concentric circle boundaries or service levels can be defined for different categories of industrial units, thereby generating refined geographic concentric circle information.

[0098] Specifically, the method acquires various points of interest (POIs) data within the target area, including commercial service information, public service information, transportation facility information, and infrastructure information, providing a foundation for subsequent assessment of service support levels. Simultaneously, it acquires transportation network data and constructs a transportation network topology structure for path calculation based on road grades, traffic capacity, and topological relationships, laying the foundation for assessing transportation convenience. Building upon this, it identifies all industrial units within the target area and generates preliminary equidistant buffer zones centered on these identified industrial units, based on a preset radius, serving as basic concentric circle information. Furthermore, to overcome the limitations of simple equidistant buffer zones, this solution utilizes POI data to identify specific supporting service information and their corresponding facility locations, calculating the spatial coverage of these services. Simultaneously, it calculates the optimal path impedance for each industrial unit to reach these facilities, based on the constructed transportation network topology structure, thereby assessing facility accessibility indicators. Finally, it comprehensively considers spatial coverage and facility accessibility indicators, optimizing and adjusting the initially generated basic concentric circle information to generate geographical concentric circle information that more accurately reflects the actual service capabilities and transportation convenience of industrial units.

[0099] Through the above technical solution, this application can overcome the limitations of traditional methods that rely solely on static methods such as preset range radii to determine geographic concentric circle information. Specifically, by comprehensively considering the spatial coverage of point of interest data and the facility accessibility index calculated from transportation network data, geographic concentric circle information that is closer to the actual situation can be generated. This geographic concentric circle information not only reflects the physical distance between industrial units and surrounding supporting facilities, but also incorporates actual traffic impedance and facility service capabilities. This allows for a more accurate assessment of the actual accessibility of industrial units to various supporting services, providing high-quality geographic dimension input for generating industrial space optimization strategies using large language models. This enhances the scientificity and practicality of industrial space adaptation optimization strategies, and reduces or even avoids unreasonable resource allocation or decision-making biases caused by inaccurate geographic information.

[0100] In one embodiment, step S60 includes: S61: Generate a decision feature matrix based on multi-dimensional analysis results, enterprise capability information, migration risk level information, and geographical sphere information; In this embodiment, step S61 aims to structurally integrate different types of data and analysis results to form a unified input format. The decision feature matrix can be understood as a high-dimensional data table, where each row may represent a potential industrial spatial unit or decision-making unit, and each column represents a related feature or indicator. This can be achieved by uniformly encoding and filling the corresponding positions of the matrix with multi-dimensional analysis results, enterprise capability information, migration risk level information, and geographical sphere information through data cleaning, standardization, and feature engineering. For example, different types of numerical data can be directly used as columns of the matrix, and categorical data can be one-hot encoded or embedded.

[0101] S62: Perform nonlinear correlation analysis and constraint conflict analysis on the decision feature matrix, and generate a preliminary optimization strategy set based on the results of nonlinear correlation analysis and constraint conflict analysis; In this embodiment, step S62 aims to identify the complex interaction relationships between the features in the decision feature matrix. These interactions may not be captured by a simple linear model. For example, the agglomeration effect of certain industries may exhibit non-linear growth at a specific spatial scale. This can be achieved using machine learning algorithms, such as neural networks, support vector machines, and decision tree ensemble models, to uncover deep correlations between features. Constraint conflict analysis refers to identifying various constraints and contradictory demands that may exist during the industrial space optimization process. For example, a region may simultaneously face conflicts between industrial development land use demands and ecological protection restrictions. This can be achieved by defining pre-defined hard constraints. The system includes constraints such as land use and environmental standards, as well as soft constraints such as industrial relevance and infrastructure carrying capacity. Then, through logical reasoning, rule engines, or optimization models, the feature combinations in the decision feature matrix are verified to identify potential parties that do not meet the constraints or have conflicts. The preliminary optimization strategy set refers to the candidate optimization schemes that meet the basic requirements and have a certain degree of feasibility and are initially screened or generated. Its implementation can be based on the association mode revealed by nonlinear association analysis, combined with the feasible space identified by constraint conflict analysis, and through heuristic algorithms, rule matching, or knowledge graph-based recommendation systems, to generate a variety of possible industrial spatial layouts, functional zoning, or enterprise relocation suggestions.

[0102] S63: Construct a multi-objective optimization problem based on a preliminary optimization strategy set, and generate a Pareto front by solving the problem; In this embodiment, step S63 aims to simultaneously consider multiple conflicting or mutually influential optimization objectives, such as maximizing economic benefits, minimizing environmental impact, maximizing employment opportunities, and minimizing migration risks. Constructing a multi-objective optimization problem requires clearly defining these objective functions and related decision variables and constraints. This can be achieved using mathematical programming methods, such as multi-objective linear programming or multi-objective nonlinear programming. Typically, solving a multi-objective optimization problem yields a Pareto front, which consists of several non-dominated solutions. Furthermore, commonly used solution algorithms include evolutionary algorithms such as NSGA-II.

[0103] S64: Sort the non-dominated solutions in the Pareto front according to a predefined priority sorting strategy to generate an optimization strategy sequence; In this embodiment, step S64 aims to rank multiple non-dominated solutions according to their overall merits to obtain an optimized strategy sequence. There are typically multiple non-dominated solutions on the Pareto front, and these solutions make different trade-offs among different objectives. To select the solution that best meets the actual needs from the non-dominated solutions, a priority ranking strategy needs to be introduced. This priority ranking strategy reflects the decision-maker's preference for different optimization objectives. Its implementation can include expert scoring, where domain experts score the importance of different objectives, and then the solutions on the Pareto front are weighted and summed according to their weights to select the solution with the highest score; or, using the Analytic Hierarchy Process (AHP), a hierarchical structure of objectives is constructed, the relative weights of each objective are determined through pairwise comparisons, and then a comprehensive evaluation is performed.

[0104] S65: Map the optimization strategy sequence to the digital twin platform to verify spatial feasibility, and generate industrial spatial optimization strategies that include spatial layout schemes, phased implementation paths, and quantification of expected results.

[0105] In this embodiment, step S65 aims to concretely implement the abstract optimization strategy in actual geographic space to ensure that the strategy is physically feasible and to visually demonstrate its effects. This can be achieved through the spatial analysis and visualization capabilities of the digital twin platform. For example, the spatial layout schemes in the optimization strategy sequence can be simulated and displayed on the 3D geographic information system of the digital twin platform to check for conflicts with existing land use planning, infrastructure carrying capacity, environmental capacity, and other actual spatial elements. This generates an industrial space optimization strategy that includes spatial layout schemes, phased implementation paths, and quantified expected results. This industrial space optimization strategy is the final output optimization strategy for industrial space, which not only includes specific spatial configuration suggestions but also provides implementation details. The plan includes a timeline and measurable expected results. The spatial layout plan specifies which areas are suitable for developing which industries, which enterprises should relocate or move out, and how infrastructure should be configured. The phased implementation path addresses the complexity of actual operations, as optimization strategies typically require phased implementation, including clearly defining the goals, tasks, resource requirements, and timelines for each phase. Quantifying expected results involves quantitatively assessing and predicting the potential economic, social, and environmental benefits of the strategy, providing decision-makers with clear decision-making support. This can be achieved by leveraging the spatial feasibility verification results and the text generation capabilities of a large language model to output the verified optimization plan in a structured report format, including detailed illustrations, data analysis, and forecast results.

[0106] Specifically, this application's solution, based on obtaining multi-dimensional analysis results, enterprise capability information, migration risk level information, and geographical layer information of the target region, integrates these heterogeneous data into a unified decision feature matrix to comprehensively reflect the target region's industrial status, enterprise characteristics, spatial potential, and limitations. Through nonlinear correlation analysis, it delves into the complex internal relationships between various features, identifying potential industrial development patterns and interaction mechanisms, achieving a deeper understanding of patterns beyond simple linear relationships. Simultaneously, it performs constraint conflict analysis on the decision feature matrix, proactively identifying and eliminating potential conflicts that do not conform to planning, regulations, or resource constraints, ensuring the compliance and feasibility of subsequent strategies. Based on the above analysis results, a preliminary set of optimization strategies is generated, further aiming to find the optimal balance among multiple competing objectives. The initial optimization strategy set is transformed into a multi-objective optimization problem. By solving this multi-objective optimization problem, a Pareto front is generated. Each non-dominated solution on the Pareto front represents a strategy that achieves an optimal trade-off between different optimization objectives, i.e., no other objective can be improved without sacrificing any objective. To select the solution that best meets the actual needs from the non-dominated solutions, a predefined priority ranking strategy is introduced to rank the non-dominated solutions on the Pareto front, thereby obtaining an optimized strategy sequence. This sequence not only provides multiple alternative solutions but also clearly demonstrates the performance of the alternative solutions on different objectives. Finally, the optimized strategy sequence is mapped onto a digital twin platform for spatial feasibility verification. Through simulation and analysis, the feasibility of the optimization strategy in physical space is ensured, and its expected effects can be intuitively displayed.

[0107] Through the aforementioned technical solutions, this application can integrate and analyze multi-source heterogeneous industrial space-related data, overcoming the limitations of traditional decision-making methods in comprehensively balancing multiple objectives and complex constraints. Specifically, through nonlinear correlation analysis, it can reveal the underlying patterns and interaction mechanisms of the data, making the generated optimization strategies more scientific and forward-looking. Meanwhile, constraint conflict analysis ensures the compliance and feasibility of the optimization strategies, avoiding resource waste and potential risks caused by blind decision-making. Furthermore, by constructing a multi-objective optimization problem and generating a Pareto front, it achieves optimal trade-offs among multiple dimensions such as economic benefits, environmental impact, and social equity, avoiding the negative effects that may result from single-objective optimization. The introduction of a priority ranking strategy further ensures that the final optimization strategy sequence accurately matches actual institutional guidance and development needs. Finally, through spatial feasibility verification using a digital twin platform, it ensures the physical feasibility and effectiveness of the optimization strategies, providing specific and quantifiable implementation paths and expected results, thus enhancing the scientific rigor, accuracy, and operability of industrial space optimization decisions.

[0108] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0109] In one embodiment, a large-model-based industrial space adaptation and optimization system is provided, which corresponds one-to-one with the large-model-based industrial space adaptation and optimization method described in the previous embodiment. The large-model-based industrial space adaptation and optimization system includes: The data acquisition module is used to determine the target area and acquire a first dataset and a second dataset of the target area. The first dataset includes industrial planning data and industrial layout data, and the second dataset includes park economic data and enterprise multidimensional data. The multidimensional analysis module is used to construct a digital twin foundation based on the first dataset and the second dataset, and to perform multidimensional analysis on the digital twin foundation through a large language model. The multidimensional analysis includes planning compliance analysis and spatial efficiency analysis. The model building module is used to construct capability assessment models and migration risk early warning models for enterprises in the target region based on multi-dimensional analysis results. The enterprise information generation module is used to input multidimensional enterprise data into the capability assessment model and the migration risk early warning model to obtain enterprise capability information and migration risk level information. The geographic information determination module is used to acquire point of interest data and transportation network data of the target area, and determine the geographic concentric circle information of the target area based on a preset range radius; The optimization strategy generation module is used to input multi-dimensional analysis results, enterprise capability information, migration risk level information, and geographical layer information into the large language model to generate industrial spatial optimization strategies.

[0110] For specific limitations regarding the industrial space adaptation and optimization system based on a large model, please refer to the limitations of the industrial space adaptation and optimization method based on a large model mentioned above, which will not be repeated here. Each module in the aforementioned industrial space adaptation and optimization system based on a large model can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0111] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for industrial spatial adaptation and optimization based on a large model, characterized in that, Including the following steps: Identify the target area and obtain a first dataset and a second dataset for the target area. The first dataset includes industrial planning data and industrial layout data, and the second dataset includes park economic data and enterprise multidimensional data. A digital twin foundation is constructed based on the first and second datasets, and a multi-dimensional analysis of the digital twin foundation is conducted through a large language model. The multi-dimensional analysis includes planning compliance analysis and spatial efficiency analysis. Based on the results of multi-dimensional analysis, a capability assessment model and a migration risk early warning model for enterprises in the target region are constructed. By inputting multidimensional data of enterprises into the capability assessment model and the migration risk early warning model, enterprise capability information and migration risk level information can be obtained. Acquire point-of-interest data and transportation network data for the target area, and determine the geographical concentric circle information of the target area based on a preset range radius; By inputting multi-dimensional analysis results, enterprise capability information, migration risk level information, and geographical sphere information into the large language model, an industrial spatial optimization strategy is generated.

2. The industrial space adaptation optimization method based on a large model according to claim 1, characterized in that: The digital twin foundation is constructed based on the first and second datasets, and a multi-dimensional analysis of the digital twin foundation is performed using a large language model. This multi-dimensional analysis includes steps such as planning compliance analysis and spatial performance analysis. Spatial registration and data fusion were performed on the first and second datasets, and then integrated into a 3D geographic information system to build a digital twin foundation. By using a large language model, semantic disambiguation and intent recognition are performed on natural language text in industrial planning data to extract industrial development constraints and spatial configuration intent. The extracted industrial development constraints and spatial configuration intentions are compared with the spatial elements in the digital twin base to calculate spatial relationships and verify consistency, generating spatial conflict detection results and planning evaluation results. Based on the economic data of the park, a multivariate spatial econometric function is constructed. The weights of each indicator of the park's economic data are determined by the entropy weight method, and spatial autocorrelation analysis is performed to generate spatial performance analysis results.

3. The industrial space adaptation optimization method based on a large model according to claim 2, characterized in that: The steps of calculating the spatial relationship and verifying the consistency between the extracted industrial development constraints and spatial configuration intentions and the spatial elements in the digital twin base, and generating spatial conflict detection results and planning evaluation results, include the following steps: Based on spatial configuration intent extracted from a large language model, we can identify the priority of industrial spatial layout and the requirements for industrial agglomeration. The priority of industrial spatial layout and the requirements for industrial agglomeration are quantified into spatial constraint rules, which include the minimum agglomeration unit area, the radiation radius of industrial correlation, and the infrastructure proximity threshold. Spatial topology analysis is performed on the industrial layout data and industrial planning data in the digital twin base to generate spatial features of industrial layout characteristics and industrial planning characteristics. The spatial features include spatial overlap rate, proximity, and distribution dispersion. Spatial overlay analysis is used to verify the consistency between spatial constraint rules and spatial characteristics, and to identify blind spots in planning coverage, areas where layout deviates, and areas with infrastructure gaps. Based on the blind spots in planning coverage, areas of layout deviation, and areas with infrastructure gaps, a conflict detection report is generated, which includes the type of spatial conflict, spatial location coordinates, and conflict intensity level. The conflict detection report is used as the spatial conflict detection result. By combining large language models with spatial conflict detection results, we generate land use development direction suggestions, including priority development sequences and spatial optimization paths, and use these suggestions as the planning evaluation results.

4. The industrial space adaptation optimization method based on a large model according to claim 2, characterized in that: The steps of constructing a multivariate spatial econometric function based on park economic data, determining the weights of each indicator in the park economic data using the entropy weight method, performing spatial autocorrelation analysis, and generating spatial performance analysis results include the following steps: The economic data of the park is used to construct an economic indicator matrix, and spatial interpolation is used to complete the missing data in the economic indicator matrix. The entropy weight method is used to calculate the weight coefficients of each economic indicator in the economic indicator matrix, and a weighted comprehensive evaluation function is generated. Global spatial autocorrelation analysis is performed using a predefined first autocorrelation algorithm to examine the spatial clustering of economic indicators in the economic indicator matrix, and local spatial autocorrelation analysis is performed using a predefined second autocorrelation algorithm to identify spatial clustering patterns. The weighted comprehensive evaluation function is coupled with the spatial clustering model, and a continuous spatial efficiency surface is generated by Kriging interpolation. The efficiency level of the continuous spatial efficiency surface is then divided to generate spatial efficiency analysis results.

5. The industrial space adaptation optimization method based on a large model according to claim 1, characterized in that: The steps for constructing a capability assessment model and a migration risk early warning model for enterprises in the target region based on multi-dimensional analysis results include the following steps: Based on the industrial layout characteristics and spatial efficiency characteristics in the multi-dimensional analysis results, target enterprises within the target area are selected. Select multidimensional data of target companies and construct enterprise capability assessment indicators and enterprise risk characteristic database based on the multidimensional data of target companies; The comprehensive innovation capability score of each target enterprise is calculated using the entropy weight method, and an enterprise capability assessment model is constructed. A full lifecycle ledger record is established based on the enterprise risk characteristic database, and the full lifecycle ledger record includes the key event sequence of the target enterprise; Based on the full lifecycle ledger records, a migration risk early warning model is constructed using the random forest algorithm.

6. The industrial space adaptation optimization method based on a large model according to claim 5, characterized in that: The steps for constructing a migration risk early warning model based on full lifecycle ledger records and using the random forest algorithm include the following steps: Extracting temporal feature vectors associated with key event nodes of a target enterprise based on key event sequences; Historical migration enterprise samples are obtained as the training set, time-series feature vectors are used as input features, and enterprise migration status is used as the prediction label. A classification model is trained using the random forest algorithm to obtain the initial model. During the training of the classification model, nodes are split according to the principle of minimizing the Gini coefficient to generate several decision trees, which constitute an ensemble learning model. The hyperparameters of the initial model are optimized based on the ensemble learning model and the preset parameter optimization strategy, and the output of the initial model is calibrated by the preset confidence threshold to obtain the migration risk warning model. The time-series feature vector of the target enterprise is input into the migration risk early warning model, which outputs the enterprise migration risk probability and assigns a risk level based on a preset risk threshold.

7. The industrial space adaptation optimization method based on a large model according to claim 6, characterized in that: The step of splitting nodes to generate several decision trees and forming an ensemble learning model during the training process of the classification model includes the following steps: (a) Sample the temporal feature vectors in the training set to generate several training subsets; (b) Construct the corresponding decision tree based on each training subset, and calculate the information gain of all features when splitting nodes. Select the feature with the largest decrease in Gini coefficient and its corresponding threshold as the optimal split point. (c) Generate a decision tree by recursively splitting until a preset termination condition is reached. The termination condition includes the number of node samples being lower than a preset sample number threshold or the decrease in the Gini coefficient being less than a preset tolerance. (d) Repeat steps (a), (b), and (c) in sequence to generate a preset number of decision trees, and integrate the prediction results of each decision tree to form an ensemble learning model.

8. The industrial space adaptation optimization method based on a large model according to claim 1, characterized in that: The step of acquiring point-of-interest data and traffic network data of the target area, and determining the geographic concentric circle information of the target area based on a preset radius, includes the following steps: Acquire point-of-interest (POI) data for the target area, including commercial service information, public service information, transportation facility information, and infrastructure information; Acquire traffic network data for the target area and construct a traffic network topology structure based on the traffic network data, including road class, traffic capacity, and topological relationships; Identify all industrial units within the target area, and based on a preset radius, generate equidistant buffer zones centered on each industrial unit to form basic concentric circle information; Based on point-of-interest data, several types of supporting service information and their corresponding supporting facility locations are identified, and the spatial coverage of various supporting service information is calculated. Calculate the optimal path impedance from each industrial unit to the location of supporting facilities based on the transportation network topology, and evaluate the facility accessibility index of the basic concentric circle information based on the optimal path impedance. The basic concentric circle information is optimized based on spatial coverage and facility accessibility indicators to generate geographic concentric circle information.

9. The industrial space adaptation optimization method based on a large model according to claim 1, characterized in that: The steps for inputting multi-dimensional analysis results, enterprise capability information, migration risk level information, and geographic layer information into a large language model to generate industrial spatial optimization strategies include the following steps: A decision feature matrix is ​​generated based on multi-dimensional analysis results, enterprise capability information, migration risk level information, and geographical sphere information. Nonlinear correlation analysis and constraint conflict analysis are performed on the decision feature matrix, and a preliminary optimization strategy set is generated based on the results of nonlinear correlation analysis and constraint conflict analysis. A multi-objective optimization problem is constructed based on a preliminary set of optimization strategies, and a Pareto front is generated by solving the problem. The non-dominated solutions in the Pareto front are sorted according to a predefined priority sorting strategy to generate an optimization strategy sequence; The optimization strategy sequence is mapped to a digital twin platform for spatial feasibility verification, generating industrial spatial optimization strategies that include spatial layout plans, phased implementation paths, and quantified expected results.

10. An industrial spatial adaptation and optimization system based on a large model, characterized in that, include: The data acquisition module is used to determine the target area and acquire a first dataset and a second dataset of the target area. The first dataset includes industrial planning data and industrial layout data, and the second dataset includes park economic data and enterprise multidimensional data. The multidimensional analysis module is used to construct a digital twin foundation based on the first dataset and the second dataset, and to perform multidimensional analysis on the digital twin foundation through a large language model. The multidimensional analysis includes planning compliance analysis and spatial efficiency analysis. The model building module is used to construct capability assessment models and migration risk early warning models for enterprises in the target region based on multi-dimensional analysis results. The enterprise information generation module is used to input multidimensional enterprise data into the capability assessment model and the migration risk early warning model to obtain enterprise capability information and migration risk level information. The geographic information determination module is used to acquire point of interest data and transportation network data of the target area, and determine the geographic concentric circle information of the target area based on a preset range radius; The optimization strategy generation module is used to input multi-dimensional analysis results, enterprise capability information, migration risk level information, and geographical layer information into the large language model to generate industrial spatial optimization strategies.

Images