A method, system, terminal, and storage medium for optimizing urban renewal planning schemes driven by a large language model.

Abstract

This invention belongs to the field of geographic information technology and discloses a method, system, terminal, and storage medium for optimizing urban renewal planning schemes driven by a large language model. The method includes: acquiring multi-source heterogeneous data and interpreting the site context based on the multi-source heterogeneous data to generate a site semantic representation; converting the semantic text input by the user into a planning parameter vector set using legal constraints and the site semantic representation; generating a spatial planning scheme that conforms to the urban renewal objectives based on the legal constraints, the site semantic representation, and the planning parameter vector set; and performing multi-dimensional indicator detection, semantic interpretation, and parameter feedback optimization on the spatial planning scheme based on a large language model that uses indicator evaluation and semantic feedback optimization to obtain an optimized urban renewal planning scheme. This invention effectively overcomes the problems of disconnect between scheme generation and optimization, and insufficient interactive interpretability in existing technologies.

Description

Technical Field

[0001] This invention relates to the field of geographic information technology, and in particular to a method, system, terminal, and storage medium for optimizing urban renewal planning schemes driven by a large language model. Background Technology

[0002] As urban development enters the stock renewal phase, urban renewal has become a key approach to improving spatial quality, optimizing functional structure, and promoting sustainable development. Renewal planning typically involves complex issues such as the transformation of old areas, land reuse, functional mixing, and spatial optimization, requiring comprehensive consideration of multiple dimensions including legal constraints, environmental carrying capacity, resident needs, and social benefits. In this process, planning and design personnel need to analyze the current site conditions, land potential, and spatial organization with the support of multi-source heterogeneous data, and formulate design schemes that balance policy regulations with human-centered needs. However, traditional urban renewal scheme formulation mainly relies on manual experience and two-dimensional spatial analysis, resulting in low data utilization, long scheme generation cycles, and imperfect feedback mechanisms, making it difficult to meet the needs of dynamic and complex renewal scenarios.

[0003] In recent years, artificial intelligence methods such as machine learning and deep generative models have been introduced into the field of urban planning, providing new ideas for scheme generation and optimization. For example, using graph convolutional networks, reinforcement learning, or generative adversarial networks for land use prediction, street morphology simulation, and building layout optimization can automate scheme exploration to a certain extent. However, these methods generally rely on a single type of structured input, lack the ability to understand and express semantic information such as regulatory texts and design intentions, and still have significant shortcomings in design logic interpretation, constraint compliance analysis, and human-computer interaction feedback.

[0004] The emergence of Large Language Models (LLMs) provides a new intelligent support path for the generation and optimization of urban renewal schemes. It possesses powerful semantic parsing, knowledge reasoning, and text generation capabilities, establishing a semantic mapping between natural language and planning parameters, achieving an intelligent transformation from understanding design intent to parameterized generation. By integrating site data, regulatory texts, and user needs, LLMs can achieve automated scheme generation, semantic evaluation, and feedback optimization at a higher level, significantly improving the intelligence and interpretability of the planning process.

[0005] However, existing methods largely rely on structured input and lack the ability to semantically understand and parameterize natural language descriptions, regulatory texts, and design intentions. This makes it difficult to achieve end-to-end collaboration from "human expression - machine computation - spatial generation" in the highly semantic and constrained scenario of urban renewal. Furthermore, the scheme evaluation and optimization stages of current intelligent generation models remain fragmented, lacking an adaptive improvement mechanism based on semantic feedback.

[0006] Existing technologies still suffer from problems such as insufficient semantic understanding, disconnect between generated results and regulatory constraints, lack of self-interpretability in indicator evaluation, and non-iterative optimization process. Therefore, existing technologies need further improvement. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to provide a method, system, terminal and storage medium for optimizing urban renewal planning schemes driven by a large language model, in order to solve the problems of insufficient semantic understanding, disconnect between generated results and legal constraints, lack of self-interpretation of indicator evaluation and non-iterative optimization process in the existing technology.

[0008] The technical solution adopted by this invention to solve the technical problem is as follows:

[0009] In a first aspect, this invention provides a method for optimizing urban renewal planning schemes driven by a large language model, including:

[0010] Acquire multi-source heterogeneous data, and interpret the site context based on the multi-source heterogeneous data to generate a site semantic representation;

[0011] Based on the semantic text input by the user, the semantic text is transformed into a set of planning parameter vectors using regulatory constraints and the semantic representation of the site.

[0012] Based on the aforementioned legal constraints, the aforementioned site semantic representation, and the aforementioned planning parameter vector set, a spatial planning scheme that conforms to the goals of urban renewal is generated.

[0013] Based on a large language model that combines indicator evaluation and semantic feedback optimization, the spatial planning scheme is subjected to multi-dimensional indicator detection, semantic interpretation, and parameter feedback optimization to obtain an optimized urban renewal planning scheme.

[0014] Secondly, this invention provides a large language model-driven urban renewal planning scheme optimization system, comprising:

[0015] The site semantic construction module is used to acquire multi-source heterogeneous data, interpret the site context based on the multi-source heterogeneous data, and generate a site semantic representation.

[0016] The planning parameter vector construction module is used to transform the semantic text input by the user into a planning parameter vector set by using regulatory constraints and the semantic representation of the site.

[0017] The spatial planning scheme generation module is used to generate a spatial planning scheme that conforms to the urban renewal goals based on the legal constraints, the site semantic representation, and the planning parameter vector set.

[0018] The indicator evaluation and semantic feedback optimization module is used to evaluate the spatial planning scheme in multiple dimensions, interpret it semantically, and optimize it with parameter feedback based on the large language model of indicator evaluation and semantic feedback optimization, so as to obtain the optimized urban renewal planning scheme.

[0019] Thirdly, the present invention provides a terminal, comprising: a processor and a memory, wherein the memory stores a large language model-driven urban renewal planning scheme optimization program, and the large language model-driven urban renewal planning scheme optimization program, when executed by the processor, is used to implement the operation of the large language model-driven urban renewal planning scheme optimization method as described in the first aspect.

[0020] Fourthly, the present invention also provides a computer-readable storage medium storing a large language model-driven urban renewal planning scheme optimization program, which, when executed by a processor, is used to implement the operation of the large language model-driven urban renewal planning scheme optimization method as described in the first aspect.

[0021] The present invention, by employing the above technical solution, has the following effects:

[0022] 1) This invention enables the planning system to have the semantic reasoning ability to understand regulations and user intentions, significantly improving the consistency between the generated results and the actual planning goals.

[0023] 2) This invention establishes a multi-dimensional indicator system and a semantic feedback mechanism, making the scheme evaluation process interpretable and interactive.

[0024] 3) This invention realizes adaptive adjustment and iterative optimization of planning parameters, and can dynamically generate the optimal solution under multi-objective constraints.

[0025] 4) This invention upgrades traditional rule-based or unidirectional generation-based planning assistance systems into intelligent planning optimization platforms with semantic understanding and intelligent feedback capabilities. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0027] Figure 1 This is a flowchart of the urban renewal planning scheme optimization method driven by a large language model in this invention.

[0028] Figure 2This is a schematic diagram of the framework of the urban renewal planning scheme optimization method driven by the large language model in this invention.

[0029] Figure 3 This is a schematic diagram of the data input and site context interpretation method in this invention.

[0030] Figure 4 This is a schematic diagram of the user semantic parsing and parameter mapping method in this invention.

[0031] Figure 5 This is a schematic diagram of the update planning scheme generation method in this invention.

[0032] Figure 6 This is a schematic diagram of the indicator evaluation and semantic feedback optimization method in this invention.

[0033] Figure 7 This is a functional schematic diagram of the terminal in one implementation of the present invention.

[0034] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0035] To make the objectives, technical solutions, and advantages of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0036] Exemplary methods

[0037] Existing methods for generating and optimizing urban renewal schemes using large language models largely rely on structured inputs and lack the ability to understand the semantics of natural language descriptions, legal texts, and design intentions. This makes it difficult to achieve end-to-end collaboration from "human expression - machine computation - spatial generation" in the highly semantic and constrained scenario of urban renewal. Furthermore, the scheme evaluation and optimization stages of current intelligent generation models remain fragmented, lacking an adaptive improvement mechanism based on semantic feedback.

[0038] Existing technologies still suffer from problems such as insufficient semantic understanding, disconnect between generated results and regulatory constraints, lack of self-interpretability in indicator evaluation, and non-iterative optimization process. Therefore, existing technologies need further improvement.

[0039] To address the above-mentioned technical problems, this invention provides a method for optimizing urban renewal planning schemes driven by a large language model. The method includes: acquiring multi-source heterogeneous data and interpreting the site context based on the multi-source heterogeneous data to generate a site semantic representation; converting the semantic text input by the user into a planning parameter vector set using legal constraints and the site semantic representation; generating a spatial planning scheme that conforms to the urban renewal goals based on the legal constraints, the site semantic representation, and the planning parameter vector set; and performing multi-dimensional indicator detection, semantic interpretation, and parameter feedback optimization on the spatial planning scheme using a large language model based on indicator evaluation and semantic feedback optimization to obtain an optimized urban renewal planning scheme. This invention effectively overcomes the problems of disconnect between scheme generation and optimization, and insufficient interactive interpretability in existing technologies.

[0040] like Figure 1 As shown, this embodiment of the invention provides a method for optimizing urban renewal planning schemes driven by a large language model, including the following steps:

[0041] Step S100: Obtain multi-source heterogeneous data, and interpret the site context based on the multi-source heterogeneous data to generate a site semantic representation.

[0042] This embodiment proposes a method for optimizing urban renewal planning schemes driven by a large language model. By constructing a legal semantic knowledge base and a user intent parsing mechanism, a full-process mapping of natural language, planning parameters, and spatial generation is achieved. Furthermore, by establishing a multi-dimensional indicator evaluation system and a semantic feedback mechanism, a dynamic closed loop of scheme generation, evaluation, and optimization is realized. An adaptive parameter adjustment algorithm is introduced, enabling the model to continuously optimize the scheme based on feedback results, thus achieving intelligent, self-explanatory, and iterative optimization of planning design.

[0043] like Figure 2 As shown in this embodiment, the first step of the urban renewal planning scheme optimization method driven by the large language model is data input and site context interpretation. This process requires input of multi-source data such as spatial vector data, raster and remote sensing data, three-dimensional model data, legal and normative texts, semantic text input by users, and spatiotemporal dynamic data. After semantic encoding and knowledge alignment, a unified site semantic vector, constraint matrix, and semantic report are formed, providing computable input for subsequent generation.

[0044] The first step aims to standardize, structure, and semantically process multi-source heterogeneous data, outputting three types of core products that are machine-readable and human-interpretable:

[0045] 1) Site semantic vector S (site_embedding) is used to represent the comprehensive semantic features of the updated site;

[0046] 2) The constraint matrix C is used to represent the computable mapping relationship between regulatory items and planning parameters;

[0047] 3) Site semantic report, used to provide readable explanations and traceable management of input data and regulatory constraints. The results of this step provide a structured input basis for subsequent "user semantic parsing and parameter mapping" and "updated planning scheme generation".

[0048] Specifically, in one implementation of this embodiment, step S100 includes the following steps:

[0049] Step S101: Acquire spatial vector data, raster and remote sensing data, three-dimensional model data, legal and regulatory text, semantic text input by the user, and spatiotemporal dynamic data to obtain the multi-source heterogeneous data.

[0050] like Figure 3 As shown, in this embodiment, the overall process of the first step includes:

[0051] 1.1 Multi-source data access and source labeling:

[0052] First, the multi-source heterogeneous data related to urban renewal projects are uniformly accessed and formatted, establishing a traceable data management mechanism. This data includes: spatial vector data, raster and remote sensing data, 3D model data, legal and regulatory texts, user-input semantic text, and spatiotemporal dynamic data, as shown in Table 1.

[0053] Table 1. Data Input Classification and Format Specifications:

[0054]

[0055] Spatial vector data is typically input in Shapefile (an open spatial data format) or GeoJSON (a geospatial data format) format to characterize basic information such as plot boundaries, road networks, green spaces, water bodies, and current land use. Raster and remote sensing data are provided in GeoTIFF (a geospatial image format) or DEM (Digital Elevation Model) format to describe topographic relief, slope, land cover type, and environmental indicators. 3D model data is input in CityGML (a geographic information standard for 3D city models), IFC (Industry Foundation Classes, a data exchange standard for building information models), or OBJ format (a file format for storing 3D geometric information), containing spatial structural attributes such as building volume, number of floors, orientation, and outline. Regulations and standards are input in PDF (Portable Document Format) or TXT (an unformatted text storage format), and may be scanned copies or electronic documents. The content covers provisions related to land use intensity control, setbacks, greening, and building spacing. User input is provided in natural language text or structured forms, describing planning objectives, design intent, preferences, and constraints. Spatiotemporal dynamic data includes traffic operation data, mobility signaling data, travel logs, and sensor event streams, used to characterize and update the dynamic features of pedestrian and traffic flow around the site.

[0056] During the data import process, this embodiment automatically identifies the format type and encoding system of each type of data, and completes coordinate system one, file integrity verification, and metadata extraction. Subsequently, a unique identifier (data_id), data source tag (source), version number, and timestamp are generated for each data entry, and its file verification hash value is recorded. Finally, a metadata index is formed to record the source and version information of all input data, ensuring the traceability and reproducibility of the subsequent semantic modeling process.

[0057] Step S102: Perform geometric, topological, and spatial feature processing on the spatial vector data and the raster and remote sensing data to obtain spatial structured information and environmental structured information, respectively.

[0058] In this embodiment, the overall process of the first step further includes:

[0059] 1.2 Spatial Data Preprocessing and Feature Derivation Calculation:

[0060] After data access is completed, this embodiment performs geometric, topological, and spatial feature processing on the input spatial vector data, raster data, and remote sensing data. For spatial vector data, geometric validity detection and topological repair are first performed to eliminate problems such as self-intersections, duplicate nodes, and hanging boundaries. Secondly, spatial alignment of data from different sources is achieved by unifying the projected coordinate system (such as the CGCS2000 coordinate system or the WGS84 coordinate system) and spatial resolution. Subsequently, an adjacency matrix is ​​established based on land parcel, road, and green space elements to support subsequent graph-based feature learning.

[0061] For both raster and remote sensing data, digital elevation models (DEMs) and multispectral bands were used to calculate relevant environmental characteristics, including slope, aspect, topographic relief, NDVI (ND Variable Visibility Index), land surface temperature (LST), land cover type (LULC), and sky visibility factor (SVF). These indicators are used to describe the site's topographic features, ecological status, and climate suitability.

[0062] Simultaneously, this embodiment performs spatial statistical operations at the plot scale, quantifying and summarizing elements such as building density, green space coverage, road connectivity, average building height, and accessibility to public facilities to form a plot-level spatial feature table. This table will serve as one of the inputs for subsequent semantic embedding and parameter mapping.

[0063] Step S103: Call the large language model fine-tuned by the urban planning corpus to parse the legal and regulatory texts and generate structured legal information for computer recognition and retrieval.

[0064] In this embodiment, the overall process of the first step further includes:

[0065] 1.3 Legal text analysis and itemized structured expression:

[0066] For legal and regulatory texts, this embodiment uses the semantic parsing capabilities of a large language model to automatically convert the content of the legal text into a structured expression that can be computed by a machine.

[0067] First, the input regulatory documents are transcribed and cleaned using OCR (Optical Character Recognition) to segment them into independent entries (by chapter, clause, or indicator). Then, a large language model, fine-tuned from the urban planning corpus, is invoked to identify core semantic elements in the regulatory text, including the planning indicator name (indicator), constraint type (operator), value range (value_range), applicable spatial scope (scope), and additional conditions (condition). Through semantic parsing and context disambiguation, these elements are organized into a uniform four-tuple record in this embodiment: (indicator, operator, value, scope). For example, for the clause "The plot ratio shall not exceed 2.5", the parsing result is: ("plot ratio", "<=", "2.5", "plot as a whole").

[0068] Each structured regulatory record is accompanied by a confidence score and a semantic evidence index (evidence_id) for subsequent credibility assessment and manual review. The parsing results are summarized into a regulation table, forming regulatory knowledge units that can be recognized and retrieved by computers.

[0069] Step S104: Based on the spatial structured information, the environmental structured information, and the regulatory structured information, a unified site semantic representation is constructed through a cross-modal feature fusion mechanism.

[0070] In this embodiment, the overall process of the first step further includes:

[0071] 1.4 Cross-modal fusion and site semantic representation generation:

[0072] After obtaining three types of structured information—space, environment, and regulations—this embodiment constructs a unified site semantic representation through a cross-modal feature fusion mechanism.

[0073] First, a spatial-topology graph is constructed using land parcels, roads, and buildings as nodes. Edge connections are established between nodes through road connectivity, functional proximity, and geographical adjacency. A graph neural network (GNN) is then used to extract spatial topological features, including semantics such as accessibility between land parcels, the influence of road hierarchy, and functional complementarity.

[0074] Secondly, two-dimensional spatial features are extracted from raster and remote sensing data using a convolutional neural network (CNN) or a Vision Transformer encoder; and three-dimensional model data are obtained from building volume and spatial structure features using neural representations based on point cloud or voxel encoding (such as the PointNet deep learning model and Occupancy Network).

[0075] Simultaneously, semantic constraint features are extracted from the structured regulatory table and user semantic text input using an attention-based language encoder (Transformer Encoder). These multi-source features are then fused using a cross-modal attention mechanism to form a unified-dimensional site semantic vector S, which is used to comprehensively represent and update the physical, environmental, and semantic attributes of the site.

[0076] Based on the structured results of regulations, this embodiment further constructs a constraint matrix C. The rows of the matrix represent regulations, the columns represent planning parameters or spatial units, and the matrix elements store the corresponding constraint type, value range, confidence level, and applicable scope.

[0077] The generated vector S and matrix C can serve as key inputs for subsequent semantic parsing, parameter mapping, and scheme generation optimization, while also supporting semantic retrieval, regulatory matching, and scheme constraint verification.

[0078] 1.5 Site Semantic Report Generation and Version Management:

[0079] After completing the semantic representation, this embodiment automatically generates a Site Semantic Report, which summarizes the input data, derived features, and regulatory constraints. The report includes:

[0080] 1) Summary of basic site statistics and derived indicators, such as plot area, building density, average building height, greening rate, road connectivity, traffic accessibility, terrain slope, ecological sensitivity, etc.

[0081] 2) Regulatory constraints and a structured list of items, outlining the constraints, numerical thresholds, and sources of semantic evidence for each planning indicator;

[0082] 3) Natural language summaries automatically generated by large language models provide semantic descriptions of the site’s main spatial features, regulatory restrictions, and potential development directions.

[0083] The report is output in JSON (a data exchange format) format, and the file includes a version number, generation timestamp, data source, and processing log. Each report update records changes and evidence indexes, enabling traceable version management. When missing fields, constraint conflicts, or entries with insufficient confidence are found, this embodiment automatically marks them as "Pending Manual Confirmation" for expert review and manual correction.

[0084] The first step of this embodiment realizes the complete processing chain from multi-source heterogeneous data to a unified semantic representation. Through data access and source annotation, spatial data preprocessing and feature derivation, legal text entry-based formatting, and cross-modal semantic fusion, three core results are ultimately output:

[0085] 1) The site semantic vector S represents the multidimensional semantic features of the updated site;

[0086] 2) Constraint matrix C quantifies the correspondence between regulatory constraints and planning parameters;

[0087] 3) Site semantic report, which integrates spatial features, regulatory constraints and semantic summary to achieve readability and auditability.

[0088] The above results provide a structured foundation for subsequent semantic parsing, parameter mapping and scheme optimization steps, and realize the bridge from "data perception" to "semantic cognition".

[0089] like Figure 1 As shown, this embodiment of the invention provides a method for optimizing urban renewal planning schemes driven by a large language model, including the following steps:

[0090] Step S200: Based on the semantic text input by the user, the semantic text is transformed into a planning parameter vector set using regulatory constraints and the semantic representation of the site.

[0091] In this embodiment, a large language model is used to parse semantic elements such as functional goals, spatial capacity, environmental preferences and experience demands in natural language, and combined with legal constraints to generate planning parameter vectors, thereby achieving accurate conversion from language description to planning parameters.

[0092] like Figure 2 As shown, in this embodiment, the second step of the large language model-driven urban renewal planning scheme optimization method is to transform the user's input natural language requirements, planning objectives, and design preferences, combined with legal constraints and site semantic information, into a computable planning parameter vector set P={p1,p2,...,p...} n These parameter vectors are used to describe constraints such as plot hierarchy, spatial layout, and functional configuration, achieving a two-way mapping from semantic requirements to quantitative parameters.

[0093] This embodiment employs a semantic parsing and retrieval-Augmented Generation (RAG) mechanism based on a large language model. It generates the value ranges, weight coefficients, and confidence levels of planning parameters through a two-layer mapping approach combining rule-based and model-based methods. This step provides the computational foundation for the subsequent automatic generation of planning schemes and constraint optimization.

[0094] Specifically, in one implementation of this embodiment, step S200 includes the following steps:

[0095] Step S201: Extract key semantic intent from the semantic text input by the user, and format the key semantic intent.

[0096] like Figure 4 As shown, in this embodiment, the overall process of the second step includes:

[0097] 2.1 User semantic input and formatting processing:

[0098] In urban renewal projects, stakeholders typically express their functional demands, spatial goals, and design preferences in natural language. This embodiment first receives natural language text or structured form information input by the user, extracts key semantic intent, and standardizes it into a standardized input format. Input methods include:

[0099] 1) Natural language description: such as "increase the greening rate and enhance pedestrian accessibility", "building height not exceeding 50 meters", "prioritize the continuity of public spaces", etc.

[0100] 2) Structured form input: Select the target type (functional land use, development intensity, environmental indicators, etc.) and target priority through a visual interface.

[0101] 3) Mixed input: Natural language and form descriptions are mixed, such as "We want the floor area ratio to be between 2.0 and 2.5, while keeping the sight corridor unobstructed".

[0102] After language normalization and encoding, the input content is unified into a standard structure:

[0103] ;

[0104] in, This indicates the content of the input statement. For input types (text / structured / hybrid). Indicates priority weight. For timestamps.

[0105] Before entering the large language model, the text portion undergoes word segmentation and stop word filtering, retaining noun and adjective keywords; the structured portion is parsed and converted into JSON format key-value pairs, then input into the model through a unified data channel. This standardization process ensures compatibility with different input types and traceability of subsequent semantic processing.

[0106] Step S202: The formatted key semantic intent is decomposed into semantic units using a pre-trained semantic analysis big language model, and the legal constraints, the site semantic representation, and the decomposed semantic units are associated based on the retrieval-enhanced big language model mechanism to obtain the semantic information after evidence retrieval.

[0107] In one implementation of this embodiment, step S202 includes the following steps:

[0108] Step S202a: The formatted key semantic intent is segmented, dependency relation parsed, and semantic unit extracted through the context attention layer and dependency parsing module.

[0109] In step S202b, the semantic importance of each word in the extracted semantic unit is calculated through a multi-head attention mechanism, and the extracted semantic units are filtered by confidence threshold and aggregated with context to obtain the semantic expression structure.

[0110] Step S202c: Based on the retrieval enhancement big language model mechanism, semantic vectorization and index retrieval processing are performed on the semantic units in the semantic expression structure using the legal constraints and the site semantic representation;

[0111] Step S202d: Semantic consistency and logical conflict detection are performed on the legal entries in the search results, and the legal entries with the highest confidence among those with overlapping constraints or directional contradictions are retained by constraint resolution algorithm.

[0112] Step S202e involves injecting the final retrieved regulatory entries and case texts into the pre-trained semantic analysis large language model with contextual hints, and combining the original semantic units to generate structured regulatory citation descriptions and initial parameter values ​​to obtain the semantic information after evidence retrieval.

[0113] like Figure 4 As shown, in this embodiment, the overall process of the second step also includes:

[0114] 2.2 Semantic Unit Decomposition and Formal Representation:

[0115] Based on the user's intent input, this embodiment utilizes a pre-trained semantic large language model (pLLM) for semantic understanding and entity recognition.

[0116] First, the semantic content is segmented, dependency relations are parsed, and semantic units are extracted using a context-aware attention layer and a dependency parsing module. The model then processes each natural language input... Decomposed into a set of several semantic units Each semantic unit includes the target topic, operational trend, semantic strength, and spatial scope. Its formal definition is as follows:

[0117] ;

[0118] in:

[0119] Semantic topic (Topic) represents planning or design objects, such as "green space ratio", "plot ratio", "building height", and "ventilation corridor";

[0120] Operational direction indicates semantic tendency, such as "increase", "decrease", "limit", "maintain" or "optimize";

[0121] : Semantic strength or confidence, obtained by normalizing the weights output by the model in the attention distribution, with a value range of [0,1];

[0122] Spatial scope is determined through location words and contextual analysis, such as "the entire plot", "along the road", and "local nodes".

[0123] The model calculates the semantic importance of each word using a multi-head attention mechanism:

[0124] ;

[0125] in, and These are the query and key vectors, respectively. For the attention dimension.

[0126] Finally, the semantic unit set is filtered by confidence threshold and aggregated with context to obtain the semantic expression structure:

[0127] ;

[0128] For example, given the input statement "Increase the green space ratio in residential areas and limit building height to no more than 50 meters", the model will parse it as follows:

[0129] ;

[0130] The parsing results are output in the form of a structured table for subsequent regulatory matching and parameter mapping.

[0131] 2.3 Legal knowledge retrieval and semantic evidence matching:

[0132] To ensure consistency between user semantic needs and regulatory constraints, this embodiment introduces a retrieval-enhanced large language model mechanism during the semantic parsing process. This mechanism associates semantic units with regulatory and case knowledge bases through two stages: embedded vector retrieval and context generation. The specific process is as follows:

[0133] 1) Semantic vectorization and index retrieval: Each semantic unit Embedded as a semantic vector by a text encoder The legal and case knowledge base is pre-coded as a set of vectors. Each vector corresponds to a regulatory entry or case summary. Semantic similarity is calculated using cosine similarity:

[0134] ;

[0135] And retrieve the most similar ones using a vector database (such as the FAISS vector database or the Milvus vector database). Count the records to obtain the matching result set. .

[0136] 2) Semantic matching and conflict detection: Semantic consistency and logical conflict detection are performed on the legal entries in the search results. If multiple entries have overlapping constraints or directional contradictions, the entry with the highest confidence is retained through the constraint resolution algorithm, and the reason for the conflict is marked.

[0137] 3) Context Integration and Generation Enhancement: The retrieved regulations and case texts are injected into the large language model as context prompts (ContextPrompt), and combined with the original semantic units to generate structured regulatory reference descriptions and initial parameter values. For example, the user semantic "increase the greening rate" is matched by the regulation "the greening rate of residential land shall not be less than 30%" (confidence level 0.95) and the case mean of 35%. The model generates a recommendation interval [0.30, 0.38] and attaches the evidence index {REG2024-GR-015, CASE2023-A02}.

[0138] 4) Result structuring and evidence storage: The generated results are stored in a triplet structure:

[0139] ;

[0140] The knowledge database records the source of evidence, confidence level, and matching algorithm version number to enable traceability and updates.

[0141] Step S203: Through a two-layer mechanism combining rule-based mapping and model reasoning mapping, the semantic information retrieved from the evidence is transformed into the planning parameter vector set.

[0142] like Figure 4 As shown, in this embodiment, the overall process of the second step also includes:

[0143] 2.4 Two-layer mapping generates computable parameter vectors:

[0144] After completing semantic unit parsing and evidence retrieval, a two-layer mechanism combining rule-based mapping and model inference mapping is used to transform semantic information into a computable planning parameter vector set P={p1,p2,...,p n}

[0145] 1) Rule-based Mapping: Based on a predefined semantic-indicator mapping rule library (Mapping Dictionary), this layer establishes a correspondence between semantic tags and specific planning indicators. For each matching entry, if the regulations define a parameter range... According to semantic trends With strength Calculation parameter adjustment range:

[0146] ;

[0147] in, This is an empirical adjustment coefficient (generally taken as 0.1–0.3), reflecting the impact of semantic strength on interval expansion.

[0148] 2) Model-based Mapping: Utilizing a large language model fine-tuned from the domain corpus, this layer performs vectorized reasoning on semantic descriptions, outputting the parameter value range, target preference weights, and confidence scores. The parameter vector is defined as follows:

[0149] ;

[0150] in, The weights represent user preference, determined by input priority. With semantic confidence Joint decision, The confidence level of the inference.

[0151] 3) Parameter set output: The final parameter set is as follows:

[0152] ;

[0153] Example output: [

[0155] {"parameter":"GreenRatio","range":[0.30,0.38],"weight": 0.8, "confidence": 0.92},

[0156] {"parameter":"BuildingHeight","range":[0, 45], "weight": 0.7, "confidence": 0.89}

[0157] ].

[0158] For abstract descriptions whose numerical range cannot be determined (such as "enhancing vitality" or "optimizing form"), the model inference layer generates alternative parameter indicators (such as "mixed land use ratio" and "public space density") through semantic embedding and case retrieval, and attaches low confidence labels to prompt human intervention.

[0159] 2.5 Parameter Consistency Verification and Feasible Domain Generation:

[0160] After obtaining the initial parameter vector set, this embodiment performs parameter consistency verification and feasible domain generation.

[0161] 1) Regulatory compliance check: for each parameter If its upper or lower limit exceeds the scope permitted by regulations. Then execute the correction:

[0162] .

[0163] 2) Parameter Coupling and Conflict Resolution: For correlated parameter pairs (such as floor area ratio and greening rate, building height and density), a multi-objective constraint solver (such as mixed integer programming) is used to solve the feasible parameter domain.

[0164] ;

[0165] in, This indicates that all regulatory constraints are met.

[0166] 3) Versioned Output and Logging: The calculated feasible domain is output as a JSON file, recording the parameter range, confidence level, conflict correction log, and evidence index. Each output includes a version number and processing timestamp to support backtracking and review. Example:

[0167] {

[0168] "site_id": "SZ2025025",

[0169] "parameters": [

[0170] {"name":"GreenRatio","feasible_range":[0.30,0.38],"confidence":0.92},

[0171] {"name":"BuildingHeight", "feasible_range":[0, 45], "confidence":0.95}

[0172] ],

[0173] "conflict_resolution": ["height adjusted per REG2024-BLD-015"],

[0174] "version": "v2.0"

[0175] }

[0176] The second step of this embodiment achieves a closed-loop transformation from natural language requirements to computable parameters through formal description of semantic units, retrieval-enhanced generation mechanisms, and a two-layer mapping strategy combining rule-based and model-based approaches. This method not only ensures regulatory consistency and semantic interpretability but also improves the reliability and executability of the parsing results through confidence propagation and parameter correction mechanisms. The resulting parameter feasible domain file provides quantifiable, legal, and traceable input boundaries for the automatic generation of subsequent planning schemes, realizing a complete semantic parsing framework from "semantic understanding - regulatory verification - parameter constraints."

[0177] like Figure 1 As shown, this embodiment of the invention provides a method for optimizing urban renewal planning schemes driven by a large language model, including the following steps:

[0178] Step S300: Based on the legal constraints, the site semantic representation, and the planning parameter vector set, generate a spatial planning scheme that meets the urban renewal goals.

[0179] In this embodiment, semantic data and parameter vectors are integrated, and a spatial optimization algorithm based on reinforcement learning and generative modeling is adopted to automatically generate an urban renewal plan that complies with regulations and objectives, covering two levels of structure: macro land use layout and micro building volume.

[0180] like Figure 2As shown in this embodiment, the third step of the urban renewal planning scheme optimization method driven by the large language model is to use the structured semantic input results obtained in the first two steps (i.e., site semantic vector S, constraint matrix C, planning parameter set P and feasible region Ω) to automatically generate a spatial planning scheme that meets the urban renewal goals under the dual constraints of regulations and semantics.

[0181] The scheme generation adopts a top-down hierarchical process, divided into two parts: macro-level land use layout generation and micro-level building layout generation, corresponding to the two levels of urban spatial organization and building form composition, respectively. Through the synergy of deep reinforcement learning and parametric generation models, a logical closed loop from "semantics-parameters-space-performance" is achieved.

[0182] Specifically, in one implementation of this embodiment, step S300 includes the following steps:

[0183] Step S301: Based on the legal constraints, the site semantic representation, and the planning parameter vector set, generate a functional layout scheme for the urban renewal plot at the macro level, and generate a building volume and spatial layout scheme that conforms to the legal constraints and design objectives at the micro level.

[0184] like Figure 5 As shown, in this embodiment, the overall process of the third step includes:

[0185] 3.1 Input Data and Target Definition:

[0186] Before proceeding to the planning scheme generation phase, the input data and task objectives are first uniformly organized and defined. The input data includes the following categories:

[0187] 1) Site semantic vector S: generated in the first step, used to comprehensively represent the natural environmental characteristics, spatial structural relationships, ecological sensitivity and geographical semantic context of the updated site, and is the basic semantic input for spatial reasoning and generation.

[0188] 2) Constraint matrix C: Contains multi-dimensional constraint relationships from planning regulations, design specifications and control indicators. Each element records the constraint type, numerical range and logical relationship, which is used to limit the solution space during the generation process.

[0189] 3) Planning parameter set P: Output from step two, containing planning objectives and indicator ranges obtained from natural language semantic parsing, such as plot ratio, building density, green space ratio, building height, etc., along with priority weights. With confidence level .

[0190] 4) Feasible region Ω: Defines the boundary region of all legal solutions in the parameter space, describes the coupling constraints and mutual exclusion relationships between different indices, and is used to determine the legality and constraint consistency of the generated results.

[0191] 5) Basic spatial data G: including study area boundaries, plot division, road centerlines, terrain grids, green space and water body elements, and existing building models, etc., to provide spatial carriers and topological constraints for generating models.

[0192] The formal definition of a planning scheme is:

[0193] ;

[0194] in, The set of parameters representing a plot of land or a building unit must meet the following conditions:

[0195] ;

[0196] The generated solution must meet all regulatory constraints and be within the feasible parameter domain.

[0197] To guide the model in adaptive optimization under multi-objective constraints, this step defines the planning synthesis objective function:

[0198] ;

[0199] in, It represents various performance indicators (such as land use efficiency, accessibility, environmental comfort, visual continuity, etc.). These are the weighting coefficients determined in the second step.

[0200] The ultimate goal of generating planning schemes is to find a solution:

[0201] .

[0202] Within this framework, the process of generating a solution is modeled as a multi-objective optimization problem within a constrained space. Through reinforcement learning and generative modeling, the solution is iteratively optimized to gradually form a macroscopic and microscopic spatial layout.

[0203] 3.2 Macro-level: Land use layout generation:

[0204] The macro-level task is to automatically generate functional layout schemes for urban renewal plots based on site semantics, regulatory constraints, and parameter ranges. The generation process combines reinforcement learning and graph space optimization, enabling the model to form a spatially coordinated and functionally complementary land use structure under complex terrain and regulatory conditions.

[0205] 1) Spatial state modeling and feature representation:

[0206] First, the study area is discretized into several minimal analysis units:

[0207] ;

[0208] Each unit Each is associated with a comprehensive feature vector:

[0209] ;

[0210] in, This represents spatial features such as topography, adjacency, ventilation, and ecology extracted from the site semantic vector; For planning parameter constraints related to this unit; These are the regulatory constraints for this unit.

[0211] Construct a spatial adjacency matrix using these units as nodes:

[0212] ;

[0213] This results in a graph structure representation that includes topological relationships and semantic attributes:

[0214] ;

[0215] This structure serves as the state input for the reinforcement learning model, enabling the agent to consider both local semantics and global spatial constraints when making decisions.

[0216] 2) Action space and decision-making strategies:

[0217] At each time step The model agent is based on the current state. Select a spatial unit and assign it a function type The action space is defined as:

[0218] ;

[0219] LandUseSet represents the land set by the user, which typically includes types such as "residential land", "commercial land", "public service facilities", "green space", and "roads".

[0220] The proxy strategy function has the following form:

[0221] ;

[0222] in, The parameters of the policy network are continuously optimized through deep learning, enabling the model to gradually learn to make optimal allocation decisions under complex constraints.

[0223] 3) Reward function design and optimization mechanism:

[0224] To ensure that the generated land use layout achieves a balance between regulatory compliance, spatial coordination, and planning objectives, several composite reward functions are defined:

[0225] ;

[0226] in:

[0227] This reflects the degree to which the layout conforms to the regulatory constraint matrix C; violations will result in negative rewards.

[0228] Assess the spatial connectivity and overall continuity of the functional layout;

[0229] : Measures whether the land use ratio, plot ratio, and development intensity are in balance with the target parameter set P;

[0230] Used to encourage spatial aggregation and zoning integrity of similar functions.

[0231] Reinforcement learning models aim to maximize the cumulative expected reward.

[0232] ;

[0233] in, This is a discount factor used to balance short-term and long-term benefits.

[0234] The policy network employs a multi-layer graph convolutional network (GCN) combined with an attention mechanism, enabling the model to propagate policies across the spatial adjacency graph, achieving semantic-level spatial cognition and functional transfer. During training, the model gradually learns optimal land use layout strategies under different site contexts and regulatory constraints through continuous trial, evaluation, and feedback.

[0235] 4) Post-layout processing and scheme generation:

[0236] After the reinforcement learning agent generates the initial layout results, further rule-based post-processing corrections are performed to ensure the integrity and compliance of the solution. This mainly includes:

[0237] Regulatory correction: Perform local replacement and parameter scaling on plots that violate regulatory constraint matrix C (such as exceeding the plot ratio, insufficient green space ratio, or exceeding the functional mixing limit);

[0238] Spatial restructuring: Clustering and merging fragmented and disconnected land use units of the same type to improve spatial integrity;

[0239] Smoothing and boundary optimization: Geometric smoothing and topological reconstruction are performed at road nodes, boundary plots and other locations to ensure the natural continuity of spatial form;

[0240] Indicator verification: Recalculate all planning indicators (land use ratio, average plot ratio, accessibility index, etc.) to confirm that the plan is still within the feasible region Ω.

[0241] The final macro layout scheme is denoted as: ;

[0242] It also includes an indicator calculation table and a regulatory compliance report, which serve as the basic inputs for generating the building layout in the next stage.

[0243] Step S302: Based on the scheme integration and structural mapping rules, the regulatory consistency verification rules, and the parameter consistency and performance verification rules, the functional layout scheme and the building volume and spatial layout scheme of the urban renewal plot are constrained and verified to obtain the spatial planning scheme that meets the urban renewal goals.

[0244] like Figure 5 As shown, in this embodiment, the overall process of the third step also includes:

[0245] 3.3 Micro-level: Building layout generation:

[0246] Macro-level land use layout Once determined, the task of this layer is to generate building volume and spatial layout schemes that comply with regulatory constraints and design objectives at the site level. This process comprehensively considers the site's functional attributes, environmental semantic characteristics, regulatory restrictions, and parameter objectives to form the spatial forms of individual buildings and building clusters. The generation mechanism adopts a two-stage approach of conditional generation + parameter optimization, combining site semantics and regulatory constraints to achieve automatic generation of high-dimensional spaces and multi-objective performance balance.

[0247] 1) Architectural generation input and feature modeling:

[0248] For each plot of land in the macro layout First, extract the corresponding input feature set: ;

[0249] in:

[0250] The land use type (e.g., residential, commercial, public service, green space, etc.).

[0251] It is a semantic sub-vector of the site, containing spatial contextual information such as topography, orientation, slope, ventilation channels, and landscape view.

[0252] This refers to the regulatory constraints on the land parcel (including setback lines, maximum height, building density, spacing requirements, etc.).

[0253] Planning parameters related to the plot (plot ratio, building coverage ratio, greening rate, building height limit, etc.).

[0254] Each plot is defined as a generation unit, the goal of which is to satisfy... and Generate a set of building layouts under constraints: ;

[0255] Each of them A vector representing the attributes of a single building (base location, number of floors, height, orientation, volume shape, etc.).

[0256] 2) Generation of building foundation and layout:

[0257] When generating the initial building layout, a conditional generation model is used to jointly encode the geometric and semantic features of the land parcel to generate a probability distribution of the building's base layout. Input features After passing through the Transformer encoder, the semantic vector embedding is obtained. : ;

[0258] The model outputs a two-dimensional probability field of the building's base layout: ;

[0259] in, This represents the probability that each location in the plot coordinate system is assigned as the base of a building.

[0260] Through multiple sampling and morphological optimization, several candidate bottom surface configurations are generated:

[0261] .

[0262] Subsequently, candidate base surfaces are screened based on regulatory constraints. For example, it is checked whether the building base surfaces exceed the setback range, whether the base surface spacing meets the fire protection and lighting specifications, and whether the base surface is in line with the distance conditions between the base surface and roads, green spaces, and water bodies. The qualified base surface combinations are used as inputs for the massing stage.

[0263] 3) Generation of building volume and form:

[0264] After the base layout is determined, the implicit volume function (IVF) is used to generate and adjust the three-dimensional form of the building. This method uses spatial coordinates. Input, output volume occupancy state:

[0265] ;

[0266] in, The point indicates that it is located inside the building volume.

[0267] The volume function is parameterized using a multilayer perceptron (MLP) and trained or optimized under constraints.

[0268] The model ensures the legality of the volume ratio and density through volume integral constraints:

[0269] ;

[0270] Meanwhile, multiple performance constraints were introduced during the morphological optimization process, including building height difference constraints. To ensure a smooth skyline; to optimize the angle between building orientation and the prevailing wind direction to improve ventilation efficiency; and to increase the shading ratio between buildings. Limiting sunlight exposure to within threshold limits ensures compliance with regulations.

[0271] Volume function generated in each iteration If any violations are found after compliance verification, the bottom position, number of layers, or height gradient will be automatically corrected through the parameter feedback mechanism until the result meets all constraints.

[0272] 4) Building group coordination and cluster optimization:

[0273] For building groups on the same or adjacent plots, cluster-level optimization is performed to coordinate spatial form and environmental performance. The optimization objective function is defined as follows:

[0274] ;

[0275] in:

[0276] : Ventilation field optimization index, calculated based on a simplified CFD (Computational Fluid Dynamics) model to determine wind speed field uniformity;

[0277] Indicators of sunshine coverage and shading rate;

[0278] Indicators of pedestrian accessibility and network connectivity between buildings;

[0279] Indicators for visual corridor continuity and skyline harmony.

[0280] This optimization problem is solved iteratively using a hybrid approach combining genetic algorithms and gradient descent. After each round of optimization, the building height, orientation, and location parameters are fine-tuned until the objective function is achieved. The convergence yields the final architectural scheme. .

[0281] 5) Output format and parameterized representation:

[0282] The generated building layout results are output in both parametric and geometric forms. The parametric output includes the mass parameters of each building (height, number of floors, land area, orientation, floor area ratio, etc.), while the geometric output is stored in CityGML or OBJ format to provide input for subsequent performance analysis and visualization. At the same time, a semantic index table is output to record the building mass and the corresponding legal entries, parameter sources and confidence levels.

[0283] Thus, a micro-level solution that meets the constraints is obtained, which serves as the input for subsequent global integration and optimization.

[0284] 3.4 Constraint Verification and Synthesis Scheme Output:

[0285] When macro land use layout With micro-architectural layout After both levels of results are generated, the next stage is scheme integration and constraint verification. The goal of this stage is to integrate the two-level results into a complete planning scheme. And ensure that it is fully legal and optimized convergent in terms of regulations, parameters and performance.

[0286] 1) Solution integration and structure mapping:

[0287] First, the macro layout and architectural layout are hierarchically mapped and integrated: ;

[0288] Macro-level indicators such as land use zoning, plot ratio, and greening rate correspond to micro-level indicators such as building density and volume form.

[0289] Through data mapping functions: This combines land parcels and building units into a unified spatial structural entity and updates the global parameter set. This ensures that all indicators remain feasible after integration.

[0290] 2) Regulatory consistency verification:

[0291] Regarding the plan All components undergo regulatory compliance verification. For each constraint item... Perform the following verification:

[0292] ;

[0293] like Exceeding the legal scope Then execute the correction:

[0294] .

[0295] Correction records are stored in log form, and the corresponding entries and correction amounts are marked in the output report to ensure that the solution is fully compliant with regulations.

[0296] 3) Parameter consistency and performance verification:

[0297] To verify whether the solution is still within the feasible region Inside, recalculate the parameter vector set:

[0298] ;

[0299] And determine:

[0300] ;

[0301] If a conflict is found (e.g., inconsistency between floor area ratio and building height, or between green space ratio and density), the local optimizer is invoked to apply linear constraints to the affected parameters, bringing the proposed solution back into the feasible region. Subsequently, multi-objective performance indicators are calculated and output, including development intensity, ecological performance, accessibility, public space connectivity, and landscape continuity.

[0302] 4) Comprehensive output and version control:

[0303] The final planning solution output includes:

[0304] Macro-level layout data: GeoJSON formatted land parcel functional zoning and indicator table;

[0305] Building massing model: a 3D geometry file in CityGML or OBJ format;

[0306] Performance metrics report: Multi-objective evaluation results in CSV or JSON format;

[0307] Regulatory compliance log: Records the verification and correction status of all constraints;

[0308] Metadata file: contains version number, generation time, model parameters and random seed information.

[0309] All outputs are entered into the version control system to support tracking and comparison during subsequent semantic evaluation and feedback optimization phases.

[0310] This step realizes the core technology transformation process from semantic parameter input to spatial planning scheme generation. Through reinforcement learning-driven land use layout generation and implicit modeling-driven building volume generation, a multi-level, structured, and interpretable planning scheme can be automatically formed within the regulatory constraint matrix and parameter feasible domain. The final result achieves a complete generation from two-dimensional functional layout to three-dimensional building form in form, and logically forms a closed-loop structure of semantics-parameters-space-performance, providing a complete input foundation for subsequent semantic evaluation and feedback optimization mechanisms.

[0311] like Figure 1 As shown, this embodiment of the invention provides a method for optimizing urban renewal planning schemes driven by a large language model, including the following steps:

[0312] Step S400: Based on the large language model of indicator evaluation and semantic feedback optimization, the spatial planning scheme is subjected to multi-dimensional indicator detection, semantic interpretation and parameter feedback optimization to obtain the optimized urban renewal planning scheme.

[0313] In this embodiment, a multi-dimensional index system covering spatial structure, environmental performance, functional efficiency, and social experience is constructed to comprehensively evaluate the solution; a large language model generates semantic interpretations and optimization suggestions, transforming feedback signals into parameter corrections to achieve adaptive optimization and iterative updates of the solution.

[0314] like Figure 2 As shown in this embodiment, the fourth step of the urban renewal planning scheme optimization method driven by the large language model is to evaluate the scheme using multi-dimensional indicators, semantic interpretation, and parameter feedback optimization based on the obtained planning scheme, thereby achieving intelligent improvement and closed-loop optimization of the planning scheme based on the large language model. This method combines quantitative calculation with semantic understanding, establishing a mapping relationship between numerical indicators and natural language, enabling the generation, evaluation, and optimization of urban renewal schemes to achieve adaptive iteration. The entire process includes three core components: construction and calculation of a multi-dimensional indicator system, semantic judgment and interpretation driven by the large language model, and semantic feedback and parameter remapping optimization.

[0315] Specifically, in one implementation of this embodiment, step S400 includes the following steps:

[0316] Step S401: Based on the spatial structure, building volume and environmental characteristics of the spatial planning scheme, construct a quantifiable multi-dimensional indicator system, and perform indicator calculations based on the multi-dimensional indicator system to obtain multi-dimensional indicator calculation results.

[0317] like Figure 6 As shown, in this embodiment, the overall process of the fourth step includes:

[0318] 4.1 Construction and Evaluation Calculation of Multidimensional Indicator System:

[0319] This embodiment is first based on the scheme. Based on the spatial structure, building volume, and environmental characteristics, a quantifiable multi-dimensional indicator system is constructed. This indicator system covers four dimensions: spatial efficiency, environmental performance, functional accessibility, and social experience, providing quantitative support for subsequent semantic evaluation.

[0320] 1) Composition of the indicator system:

[0321] Define the indicator set as follows: .

[0322] Each of them This represents a computable metric. Typical metrics include:

[0323] Table 2 Evaluation Dimensions of Updated Planning Schemes:

[0324]

[0325] The indicator results have been normalized.

[0326] ;

[0327] The overall performance score is defined as follows:

[0328] ;

[0329] in, These are the indicator weights obtained from the second step of analysis.

[0330] The results are output as structured data, example:

[0331] {

[0332] FloorAreaRatio: 0.32,

[0333] "GreenCoverage": 0.85,

[0334] "Connectivity": 0.74,

[0335] "CompositeScore": 0.83

[0336] }

[0337] 2) Semantic labeling of indicators:

[0338] To achieve semantic analysis, this embodiment adds semantic tags to each indicator: ;

[0339] Indicator theme;

[0340] Target trend (e.g., increase, decrease);

[0341] Semantic weight or confidence level;

[0342] : Corresponding regulatory entries or parameter index.

[0343] The labeled results and indicator values ​​are input together into the large language model to form contextual information for semantic reasoning.

[0344] Step S402: Based on the calculation results of the multi-dimensional indicators, the spatial planning scheme is semantically evaluated and interpreted, and the generated semantic suggestions are transformed into adjustment information in the parameter space. Based on the adjustment information, parameter feedback optimization is performed to obtain the optimized urban renewal planning scheme.

[0345] In one implementation of this embodiment, step S402 includes the following steps:

[0346] Step S402a: Based on the calculation results of the multi-dimensional indicators, the spatial planning scheme is subjected to compliance reasoning through the context attention mechanism to determine whether each indicator meets the regulatory constraints and target parameter range, and potential improvement directions are identified according to the target trend and weight to generate the semantic suggestions.

[0347] Step S402b: Simultaneously generate deviation information in the evaluation output, integrate all semantic deviation information into a feedback vector, and map it to the parameter set to obtain the corrected parameter set;

[0348] Step S402c: Based on the modified parameter set, iteratively generate a new spatial planning scheme that conforms to the urban renewal goals, and obtain the optimized urban renewal planning scheme.

[0349] like Figure 6 As shown, in this embodiment, the overall process of the fourth step also includes:

[0350] 4.2 Semantic evaluation and interpretation driven by large language models:

[0351] After the indicators are calculated, this embodiment utilizes the contextual understanding and reasoning capabilities of a large language model to perform semantic evaluation and natural language interpretation of the plan. The model establishes a logical mapping between quantitative indicators and regulatory semantics, thereby generating an interpretable planning evaluation.

[0352] 1) Input construction:

[0353] Input includes: ;

[0354] That is, site semantic vector constraint matrix Parameter set With indicator results .

[0355] This information is organized into a structured prompt template, which includes: a site semantic summary (from step one), parameter objectives and weight information (from step two), the correspondence between indicator results and regulations, and a description of the scheme space and environment (from step three).

[0356] 2) Semantic reasoning and logical evaluation:

[0357] The model uses a context attention mechanism for two-layer reasoning: compliance reasoning to determine whether each indicator meets regulatory constraints. and target parameter range Optimization reasoning based on target trends With weight Identify potential areas for improvement.

[0358] The semantic results of the output are structured as follows:

[0359] ;

[0360] in:

[0361] Indicator status (satisfied / deviation / violation);

[0362] Analysis of the reasons for deviation;

[0363] Natural language optimization suggestions.

[0364] For example, the model might generate the following output:

[0365] "The overall layout of the plan meets the requirements for plot ratio control, but the green space ratio is slightly low. It is recommended to add roof greening near plot G2 to improve ventilation performance."

[0366] 3) Semantic interpretation and traceability:

[0367] This embodiment also records semantic source information during the model generation process, including indicator references, legal basis, and confidence scores. All generated text is bound to corresponding indicators, forming a structured semantic log to ensure the transparency and auditability of the evaluation results.

[0368] 4.3 Semantic Feedback and Parameter Remapping Optimization:

[0369] After evaluation, this embodiment transforms the semantic suggestions generated by the large language model into parameter space adjustment information, thereby achieving further optimization of the solution. This process establishes a mapping function from the natural language evaluation results to parameter vector correction, enabling the semantic feedback to be quantitatively calculated and driving a new round of solution generation.

[0370] 1) Semantic feedback quantification:

[0371] The model simultaneously generates bias information in the evaluation output: ;

[0372] in, The adjusted indicator themes To adjust the direction (e.g., increase the green space ratio, reduce the building height). The feedback strength is determined by integrating all semantic bias information into a feedback vector.

[0373] ;

[0374] And map to the parameter set: ;

[0375] in, This is the semantic feedback adjustment coefficient, which controls the magnitude of the semantic signal's correction of the parameters.

[0376] 2) Parameter remapping and optimization updates:

[0377] Corrected parameter set The new solutions are re-input into the third step of the planning scheme generation process, driving the iteration of new scheme generation. After generation, this embodiment calculates the performance improvement:

[0378] ;

[0379] like If the optimization is successful, the new solution is retained and archived; if the improvement is insufficient, the feedback weights are adjusted and the iteration is repeated until convergence.

[0380] 3) Human-machine co-adaptation and semantic learning:

[0381] After each round of semantic evaluation, planning experts can manually revise the model output. The revised content is re-encoded into training samples through natural language parsing, which are used to fine-tune the model's semantic mapping function. As the number of iterations increases, this embodiment gradually develops an adaptive learning capability for specific urban renewal contexts, enabling continuous optimization of semantic understanding, parameter adjustment, and spatial generation processes.

[0382] 4.4 Output Results and Optimization Closed Loop Formation:

[0383] After the semantic feedback iteration is completed, this embodiment outputs the final optimized solution. And the corresponding semantic evaluation report. Output results include:

[0384] Optimization plan documents: Updated land use and building layout;

[0385] Indicator Evaluation Table: Final values ​​and rates of change for each indicator;

[0386] Semantic evaluation summary: Explanatory text generated by the model;

[0387] Feedback log: Records semantic deviations and parameter adjustments;

[0388] Version information: includes optimization rounds, parameter changes, and model configuration.

[0389] Thus, the method completes the entire closed loop from data input, semantic parsing, parameter mapping, spatial generation, indicator evaluation, and semantic feedback, achieving intelligent optimization of urban renewal planning schemes.

[0390] This embodiment achieves semantic evaluation and adaptive parameter optimization of planning schemes through multi-dimensional index assessment and a large language model semantic reasoning mechanism. By semantic labeling and feedback remapping, this embodiment establishes a computable connection between numerical parameters and natural language, enabling the planning process to possess interpretability, traceability, and self-learning capabilities. Ultimately, this forms a closed-loop optimization method based on a large language model, capable of continuously improving the overall quality of urban renewal planning schemes under multi-objective constraints.

[0391] This embodiment achieves the following technical effects through the above technical solution:

[0392] 1) This embodiment enables the planning system to have the semantic reasoning ability to understand regulations and user intentions, which significantly improves the consistency between the generated results and the actual planning goals.

[0393] 2) This embodiment establishes a multi-dimensional indicator system and a semantic feedback mechanism to make the scheme evaluation process interpretable and interactive.

[0394] 3) This embodiment realizes adaptive adjustment and iterative optimization of planning parameters, and can dynamically generate the optimal solution under multi-objective constraints.

[0395] 4) This embodiment upgrades the traditional rule-based or one-way generation-based planning assistance system into an intelligent planning optimization platform with semantic understanding and intelligent feedback capabilities.

[0396] Exemplary device

[0397] Based on the above embodiments, the present invention also provides a large language model-driven urban renewal planning scheme optimization system, comprising:

[0398] The site semantic construction module is used to acquire multi-source heterogeneous data, interpret the site context based on the multi-source heterogeneous data, and generate a site semantic representation.

[0399] The planning parameter vector construction module is used to transform the semantic text input by the user into a planning parameter vector set by using regulatory constraints and the semantic representation of the site.

[0400] The spatial planning scheme generation module is used to generate a spatial planning scheme that conforms to the urban renewal goals based on the legal constraints, the site semantic representation, and the planning parameter vector set.

[0401] The indicator evaluation and semantic feedback optimization module is used to evaluate the spatial planning scheme in multiple dimensions, interpret it semantically, and optimize it with parameter feedback based on the large language model of indicator evaluation and semantic feedback optimization, so as to obtain the optimized urban renewal planning scheme.

[0402] Based on the above embodiments, the present invention also provides a terminal, the principle block diagram of which can be as follows: Figure 7 As shown.

[0403] The terminal includes: a processor, a memory, an interface, a display screen, and a communication module connected via a system bus; wherein, the processor of the terminal provides computing and control capabilities; the memory of the terminal includes a computer-readable storage medium and internal memory; the computer-readable storage medium stores an operating system and computer programs; the internal memory provides an environment for the operation of the operating system and computer programs in the computer-readable storage medium; the interface is used to connect to external devices; the display screen is used to display relevant information; and the communication module is used to communicate with a cloud server or other devices.

[0404] When executed by the processor, this computer program is used to implement a method for optimizing urban renewal planning schemes driven by a large language model.

[0405] It will be understood by those skilled in the art that Figure 7 The schematic diagram shown is only a partial structural diagram related to the present invention and does not constitute a limitation on the terminal to which the present invention is applied. A specific terminal may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements.

[0406] In one embodiment, a terminal is provided, comprising: a processor and a memory, the memory storing a large language model-driven urban renewal planning scheme optimization program, which, when executed by the processor, is used to implement the operation of the large language model-driven urban renewal planning scheme optimization method described above.

[0407] In one embodiment, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores a large language model-driven urban renewal planning scheme optimization program, which, when executed by a processor, is used to implement the operation of the large language model-driven urban renewal planning scheme optimization method described above.

[0408] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, database, or other media used in the embodiments provided by this invention can include both non-volatile and volatile memory.

[0409] In summary, this invention provides a method, system, terminal, and storage medium for optimizing urban renewal planning schemes driven by a large language model. The method includes: acquiring multi-source heterogeneous data and interpreting the site context based on the multi-source heterogeneous data to generate a site semantic representation; converting the semantic text input by the user into a planning parameter vector set using legal constraints and the site semantic representation; generating a spatial planning scheme that conforms to the urban renewal objectives based on the legal constraints, the site semantic representation, and the planning parameter vector set; and performing multi-dimensional indicator detection, semantic interpretation, and parameter feedback optimization on the spatial planning scheme using a large language model based on indicator evaluation and semantic feedback optimization to obtain an optimized urban renewal planning scheme. This invention effectively overcomes the problems of disconnect between scheme generation and optimization, and insufficient interactive interpretability in existing technologies.

[0410] It should be understood that the application of the present invention is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A large language model driven urban renewal planning scheme optimization method, characterized in that, The method comprises the following steps: acquiring multi-source heterogeneous data, and generating a site semantic representation according to site context interpretation based on the multi-source heterogeneous data; translating semantic text input by a user into a planning parameter vector set by using regulation constraints and the site semantic representation; generating a spatial planning scheme conforming to an urban renewal target based on the regulation constraints, the site semantic representation, and the planning parameter vector set; detecting the spatial planning scheme in multiple dimensions, interpreting the spatial planning scheme semantically, and optimizing parameters of the spatial planning scheme based on a large language model for index evaluation and semantic feedback optimization, to obtain an optimized urban renewal planning scheme; the step of translating the semantic text input by the user into the planning parameter vector set by using the regulation constraints and the site semantic representation comprises the following steps: extracting key semantic intents from the semantic text input by the user, and performing format processing on the key semantic intents; performing semantic unit decomposition on the formatted key semantic intents by using a pre-trained semantic analysis large language model, and associating the regulation constraints, the site semantic representation, and the decomposed semantic units based on a retrieval-enhanced large language model mechanism to obtain evidence-retrieved semantic information; translating the evidence-retrieved semantic information into the planning parameter vector set by using a double-layer mechanism combining rule mapping and model reasoning mapping; the step of generating the spatial planning scheme conforming to the urban renewal target based on the regulation constraints, the site semantic representation, and the planning parameter vector set comprises the following steps: generating a functional layout scheme of an urban renewal plot at a macro level and generating an architectural volume and spatial layout scheme conforming to regulation constraints and design targets at a micro level based on the regulation constraints, the site semantic representation, and the planning parameter vector set; performing constraint verification on the functional layout scheme of the urban renewal plot and the architectural volume and spatial layout scheme based on scheme integration and structure mapping rules, regulation consistency verification rules, and parameter consistency and performance review rules, to obtain the spatial planning scheme conforming to the urban renewal target; the step of detecting the spatial planning scheme in multiple dimensions, interpreting the spatial planning scheme semantically, and optimizing parameters of the spatial planning scheme based on the large language model for index evaluation and semantic feedback optimization, to obtain the optimized urban renewal planning scheme comprises the following steps: constructing a quantifiable multi-dimensional index system based on spatial structures, architectural volumes, and environmental characteristics of the spatial planning scheme, and performing index calculation based on the multi-dimensional index system to obtain multi-dimensional index calculation results; performing semantic evaluation and interpretation on the spatial planning scheme based on the multi-dimensional index calculation results, converting generated semantic suggestions into adjustment information of a parameter space, and performing parameter feedback optimization based on the adjustment information to obtain the optimized urban renewal planning scheme.

2. The large language model driven urban renewal planning scheme optimization method according to claim 1, characterized in that, the step of acquiring multi-source heterogeneous data and generating a site semantic representation according to site context interpretation based on the multi-source heterogeneous data comprises the following steps: Obtain the multi-source heterogeneous data by acquiring spatial vector data, raster and remote sensing data, three-dimensional model data, regulation and specification texts, semantic texts input by the user, and spatio-temporal dynamic data; Perform geometric, topological, and spatial feature processing on the spatial vector data and the raster and remote sensing data to obtain spatial structured information and environmental structured information, respectively; Call the large language model fine-tuned by the urban planning corpus to analyze the regulation and specification texts and generate regulation structured information for computer recognition and retrieval; Based on the spatial structured information, the environmental structured information, and the regulation structured information, construct a unified site semantic representation through a cross-modal feature fusion mechanism; wherein the site semantic representation is used to comprehensively represent the physical, environmental, and semantic properties of the updated site.

3. The large language model driven urban renewal planning scheme optimization method of claim 1, wherein, The pre-trained semantic analysis large language model is used to perform semantic unit decomposition on the formatted key semantic intent, and based on the retrieval-enhanced large language model mechanism, the regulation constraints, the site semantic representation, and the decomposed semantic units are associated to obtain evidence-retrieved semantic information, including: The formatted key semantic intent is segmented, dependency relationship is analyzed, and semantic units are extracted through a context attention layer and a dependency syntax analysis module; The semantic importance of each word in the extracted semantic units is calculated through a multi-head attention mechanism, and the extracted semantic units are filtered and aggregated through a confidence threshold to obtain a semantic expression structure; Based on the retrieval-enhanced large language model mechanism, the regulation constraints and the site semantic representation are used to perform semantic vectorization and index retrieval processing on the semantic units in the semantic expression structure; The regulation items in the retrieval results are detected for semantic consistency and logical conflict, and the regulation item with the highest confidence is retained in the regulation items with overlapping constraints or directional contradictions through a constraint resolution algorithm; The finally retrieved regulation items and the case texts are injected into the pre-trained semantic analysis large language model in a context prompt manner to generate structured regulation reference descriptions and parameter initial values in combination with the original semantic units, and the evidence-retrieved semantic information is obtained.

4. The large language model driven urban renewal planning scheme optimization method of claim 1, wherein, Based on the multi-dimensional index calculation results, the spatial planning scheme is semantically judged and explained, and the generated semantic suggestions are converted into parameter space adjustment information, and based on the adjustment information, parameter feedback optimization is performed to obtain the optimized urban renewal planning scheme, including: Based on the multi-dimensional index calculation results, the spatial planning scheme is compliance-inferred through a context attention mechanism to determine whether each index meets the regulation constraints and target parameter interval, and according to the target trend and weight, the potential improvement direction is identified to generate the semantic suggestions; Bias information is generated synchronously in the judgment output, all semantic bias information is integrated into a feedback vector, and is mapped to a parameter set to obtain a revised parameter set; Based on the revised parameter set, a new spatial planning scheme that meets the urban renewal target is iteratively generated to obtain the optimized urban renewal planning scheme.

5. A large language model driven urban renewal planning scheme optimization system for implementing the large language model driven urban renewal planning scheme optimization method according to any one of claims 1-4, characterized in that, including: A site semantic construction module is configured to acquire multi-source heterogeneous data and interpret a site context according to the multi-source heterogeneous data to generate a site semantic representation; A planning parameter vector construction module is configured to convert semantic text input by a user into a planning parameter vector set by using regulation constraints and the site semantic representation; A spatial planning scheme generation module is configured to generate a spatial planning scheme in line with a city renewal target based on the regulation constraints, the site semantic representation, and the planning parameter vector set; An index evaluation and semantic feedback optimization module is configured to perform multi-dimensional index evaluation, semantic interpretation, and parameter feedback optimization on the spatial planning scheme based on a large language model for index evaluation and semantic feedback optimization to obtain an optimized city renewal planning scheme.

6. A terminal, characterized by comprising: The method comprises the following steps: A processor and a memory are provided, and the memory stores a large language model driven city renewal planning scheme optimization program, which is used to implement the operations of the large language model driven city renewal planning scheme optimization method according to any one of claims 1-4 when executed by the processor.

7. A computer readable storage medium characterized in that, The computer readable storage medium stores a large language model driven city renewal planning scheme optimization program, which is used to implement the operations of the large language model driven city renewal planning scheme optimization method according to any one of claims 1-4 when executed by the processor.

Images