A generative intelligent computing method and system for urban design space optimization

Abstract

The application relates to the technical field of urban space intelligent design, and discloses a generative intelligent computing method and system for urban design space optimization. The method acquires multi-source heterogeneous data of a built environment, constructs an urban space semantic graph after standardization processing and semantic alignment, codes conditions in combination with urban planning control conditions to form a design constraint vector containing hard constraints of planning rules and soft constraints of design preferences, establishes an urban design space generation model based on the urban space semantic graph and the design constraint vector, generates a candidate three-dimensional space optimization scheme, evaluates the candidate scheme in terms of multiple targets such as sunshine, wind environment, skyline visibility and space compactness, feeds back optimization in combination with user editing behavior, and outputs a target urban design space optimization scheme. The application can improve the scheme generation efficiency, planning compliance and design optimization pertinence, and is suitable for scenes such as urban renewal, plot development and block space form deduction.

Description

Technical Field

[0001] This invention relates to the field of intelligent urban space design technology, specifically to a generative intelligent computing method and system for optimizing urban space design. Background Technology

[0002] Urban design is gradually shifting from a traditional experience-driven approach to a data-driven and intelligently assisted model. In actual projects, designers typically need to comprehensively consider the existing built environment, site boundary conditions, planning control indicators, urban interface relationships, public space organization, and landscape guidance requirements, and complete the conceptualization, comparison, and optimization of multiple candidate schemes within a relatively short period. Due to the complexity of the relevant input information sources, including oblique photogrammetry 3D models, geographic information system vector data, and planning maps, as well as urban management regulations, design briefs, and human experience judgments, there are often significant differences between different data in terms of coordinate systems, expression methods, and semantic granularity, resulting in a large workload for data processing and rule translation in the early stages of design.

[0003] In existing technologies, the generation of urban spatial plans typically employs parametric modeling, rule-based modeling, or machine learning-based morphological generation methods for design assistance. For example, Chinese patent CN112099627B discloses an AI-based virtual reality real-time interactive platform for urban design. Its core is to first rasterize the site's geospatial information and combine it with a case library for convolutional neural network learning and adversarial generation, automatically generating multiple urban design schemes that meet planning indicators. Then, a holographic 3D sand table is constructed to display these schemes. Furthermore, this solution combines VR glasses, data gloves, voice recognition, and a command library to enable real-time adjustments to the urban design schemes via gestures or voice, and to verify planning indicators and update the 3D sand table display in real time after adjustments.

[0004] Therefore, existing technologies primarily focus on the 3D display and interactive modification of generated solutions. They lack a more systematic technical organization for unified and standardized processing of multi-source heterogeneous data in the built environment, deep expression of urban spatial semantic relationships, hierarchical coding of hard constraints from planning rules and soft constraints from design preferences, and continuous optimization mechanisms based on multi-objective performance evaluation and user preference feedback. Especially in the context of intelligent urban design management, existing technologies are insufficient in providing collaborative support between generative models, computational frameworks, evaluation feedback, and management decisions, making it difficult to better adapt to the application requirements of efficient, refined, and sustainable optimization in complex scenarios. Based on this, this application proposes a generative intelligent computing scheme for urban design spatial optimization. Summary of the Invention

[0005] To address the aforementioned shortcomings, this invention provides the following technical solution: a generative intelligent computing method for urban design spatial optimization, comprising the following steps: Acquire multi-source heterogeneous data of the built environment, and perform standardization and semantic alignment processing on the multi-source heterogeneous data to obtain a basic urban spatial dataset; Based on the aforementioned urban spatial basic dataset, spatial entity recognition and spatial relationship analysis are performed to construct an urban spatial semantic map. Based on the urban spatial semantic map and urban planning control conditions, condition coding is performed to obtain a design constraint vector. The design constraint vector includes hard constraint coding to characterize the statutory planning control content and soft constraint coding to characterize the design style and spatial tendency. An urban design space generation model is established based on the urban spatial semantic map and the design constraint vector, and candidate three-dimensional spatial optimization schemes are generated using the urban design space generation model. The candidate three-dimensional space optimization schemes are evaluated for multi-objective performance to obtain the evaluation results of the candidate schemes; Based on the evaluation results of the candidate schemes, the urban design space generation model is optimized by feedback, and the target urban design space optimization scheme is output.

[0006] Furthermore, the multi-source heterogeneous data includes at least oblique photogrammetry 3D reality model data, geographic information system vector data, urban land use classification planning map data, and urban management regulations text data; The standardization process includes data format unification, coordinate system unification, scale normalization, and attribute field normalization. The semantic alignment process includes unified label mapping for building objects, road objects, green space objects, open space objects, and planning control objects.

[0007] Furthermore, the construction of the urban spatial semantic map includes: Extract building entities, road entities, green space entities, open space entities, land parcel boundary entities, and public facility entities from the aforementioned urban spatial basic dataset; Based on the extracted spatial entities, calculate the adjacency relationships, line-of-sight relationships, environmental relationships, and functional relationships; Based on the spatial entities and their corresponding relationships, an urban spatial semantic map is constructed to represent the urban spatial organization structure.

[0008] Furthermore, the step of performing conditional encoding based on the urban spatial semantic map and urban planning control conditions to obtain the design constraint vector includes: The urban planning control conditions are classified by rules to obtain a set of hard constraints and a set of soft constraints; The set of hard constraints is mapped to a constraint mask matrix to represent the construction boundary, height limit, density limit, spacing requirements, and sunlight limit. The soft constraint set is mapped to a preference feature vector to characterize architectural style preference, interface continuity, open space preference, and skyline harmony. The constraint mask matrix and the preference feature vector are fused and encoded to obtain the design constraint vector.

[0009] Furthermore, the establishment of the urban design space generation model includes: Based on the urban spatial semantic map, the current spatial features are extracted to obtain the spatial feature tensor; Based on the design constraint vector, conditional control features are extracted and injected into the generative model; Based on the spatial feature tensor and the conditional control features, conditional generation training is performed to obtain an initial generation model; The compliance of the scheme output by the initial generated model with the execution rules is checked, and the initial generated model is corrected based on the check results to obtain the urban design space generation model.

[0010] Furthermore, the urban design space generation model is a conditional generative adversarial network model or a conditional diffusion generation model; The output of the urban design space generation model includes at least the building block layout results, public space outline results, building height distribution results, and green space organization results.

[0011] Furthermore, the multi-objective performance evaluation of the candidate three-dimensional space optimization schemes to obtain the candidate scheme evaluation results includes: The candidate three-dimensional spatial optimization schemes are subjected to solar radiation analysis to obtain solar radiation evaluation values; Wind environment simulation was performed on the candidate three-dimensional spatial optimization schemes to obtain wind environment evaluation values; Visual field analysis is performed on the candidate 3D spatial optimization schemes to obtain the skyline evaluation value; Morphological statistical analysis was performed on the candidate three-dimensional spatial optimization schemes to obtain spatial compactness evaluation values; The candidate schemes are evaluated by comprehensively scoring the solar radiation evaluation value, wind environment evaluation value, skyline evaluation value, and space compactness evaluation value.

[0012] Furthermore, the step of optimizing the urban design space generation model based on the evaluation results of the candidate schemes includes: Based on the evaluation results of the candidate solutions, a set of preferred solutions is extracted; Collect user actions such as retaining, deleting, replacing, and partially editing the preferred scheme set to form a design preference sample set; Based on the design preference sample set, preference ranking learning is performed to obtain preference update parameters; The urban design space generation model is iteratively updated based on the preference update parameters and the evaluation results of the candidate schemes.

[0013] Furthermore, the method also includes: The training and inference tasks of the urban design space generation model are deployed on a distributed computing framework. Containerize and encapsulate training and inference tasks; Based on cloud computing resources, the packaged training and inference tasks are elastically scheduled and executed in parallel to accelerate model training and candidate solution generation.

[0014] This application also provides a generative intelligent computing system for urban design spatial optimization, including: The data construction unit is used to acquire multi-source heterogeneous data of the built environment and perform standardization processing, semantic alignment processing and spatial relationship parsing processing to construct an urban spatial semantic map; Conditional encoding unit, used to generate design constraint vectors based on urban planning control conditions; A computational unit is used to establish an urban design space generation model based on the urban spatial semantic map and the design constraint vector, and to generate candidate three-dimensional spatial optimization schemes. The evaluation and optimization unit is used to perform multi-objective performance evaluation on the candidate three-dimensional space optimization schemes and to perform feedback optimization on the urban design space generation model based on the evaluation results. The interactive display unit is used to receive user input parameters and output a 3D visualization of the target city's spatial design optimization scheme.

[0015] The beneficial effects of this invention are as follows: 1. This application standardizes and semantically aligns multi-source heterogeneous data of the built environment, unifying oblique photogrammetry 3D models, geographic information system vector data, planning map data, and regulatory text data into a computable data representation system. This reduces processing deviations caused by inconsistent formats, coordinates, and semantic granularity among different data sources, providing a more complete and consistent data foundation for subsequent spatial scheme generation.

[0016] 2. This application constructs an urban spatial semantic map, which incorporates building entities, road entities, green space entities, open space entities and their adjacency relationships, sight relationship relationships, environmental relationship relationships and functional relationship relationships into a unified expression framework. This enables the generated model to not only identify the control conditions within a single plot, but also to identify the relationship characteristics between the plot and the surrounding urban environment, thereby making the generated results more consistent with the real urban spatial organization.

[0017] 3. This application divides urban planning control conditions into hard constraints based on planning rules and soft constraints based on design preferences, encodes them separately, and then merges them to form a design constraint vector. This allows the statutory control content and the design preference content to be expressed hierarchically in the same generation process. This ensures that candidate schemes meet the minimum requirements for setbacks, height limits, plot ratio, building density, spacing, and sunlight, while also improving the controllability of the schemes in terms of style orientation, interface continuity, open space organization, and skyline coordination.

[0018] 4. This application establishes an urban design space generation model based on urban spatial semantic map and design constraint vector, and generates multiple candidate three-dimensional spatial optimization schemes. Compared with the scheme conception method that relies on repeated manual adjustments, it can output spatial results with diversity and comparability in a shorter time, thereby improving the efficiency of urban design scheme generation and shortening the early comparison and deduction cycle.

[0019] 5. This application conducts multi-objective performance evaluations of candidate three-dimensional spatial optimization schemes based on factors such as sunlight, wind environment, skyline visibility, and spatial compactness. It also performs feedback optimization on the generated model based on the evaluation results, thus linking the scheme generation process with the performance analysis process. This approach is no longer limited to the manual round-by-round correction method after generation, which helps to improve the pertinence and stability of scheme optimization. Attached Figure Description

[0020] The invention will now be described in more detail with reference to embodiments and the accompanying drawings. Figure 1 A schematic diagram of a generative intelligent computing method for urban design space optimization provided by the present invention; Figure 2 A schematic diagram of a generative intelligent computing system for urban design space optimization provided by the present invention; Figure 3 This shows the interim results of constructing the three-dimensional occupancy tensor for the candidate three-dimensional space optimization scheme in Embodiment 5 of this application after step S71; Figure 4 This shows the height tensor plane mapping result further formed by step S71 in Embodiment 5 of this application; Figure 5The multi-scale spatial evaluation map constructed by step 5S72 of embodiment 5 of this application is shown.

[0021] Figure 6 The graph shows the feedback optimization iteration results of the generative model based on comprehensive evaluation values ​​provided in Embodiment 5 of this application. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Example 1 like Figure 1 As shown in this embodiment, a generative intelligent computing method for urban design space optimization is provided. This method is applicable to scenarios such as urban renewal, deepening of area control planning, urban design of key plots, and improvement of existing built-up environments. It is used to automatically generate and optimize three-dimensional spatial schemes under given plot boundaries, planning indicators, and design control conditions.

[0024] The method in this embodiment includes the following steps: S10: Obtain multi-source heterogeneous data of the built environment, perform standardization and semantic alignment processing on the multi-source heterogeneous data, and obtain the urban spatial basic dataset.

[0025] The multi-source heterogeneous data includes at least oblique photogrammetry 3D reality model data, geographic information system vector data, urban land use classification planning map data, and urban management regulations text data. In this embodiment, oblique photogrammetry 3D reality model data is used to characterize the spatial morphology and actual height distribution of the existing built environment; geographic information system vector data is used to characterize the planar boundary information of roads, plots, water systems, green spaces, and public facilities; urban land use classification planning map data is used to characterize planning control conditions such as land use nature, plot ratio, building density, height limit, and green space ratio; and urban management regulations text data is used to extract rule constraints such as setbacks, spacing, opening ratio, and sunlight baseline.

[0026] In this embodiment, the standardization process includes data format conversion, coordinate registration, scale unification, and field normalization. Specifically, model, vector, and text rule data from different sources are uniformly converted into a computable data structure. Spatial geometric information can be converted into a 3D voxel mesh representation with attributes or a vector area representation with semantic labels; text rule data is converted into structured rule fields after natural language parsing. Semantic alignment is used to uniformly map objects such as buildings, roads, green spaces, interfaces, public open spaces, and visual corridor nodes into the same semantic system, forming a data foundation that can directly participate in subsequent map construction and model training.

[0027] For example, for a specific site to be designed, its oblique photogrammetry model has a resolution of 0.2m, the GIS vector data uses the CGCS2000 coordinate system, and the planning map specifies a plot ratio of 4.5, a building density of 35%, and a building height limit of 80m. After standardization, the 3D model, site boundaries, road red lines, green space boundaries, and indicator text are uniformly mapped to the same coordinate system and the same attribute field system, forming a basic urban spatial dataset.

[0028] S20: Based on the aforementioned urban spatial basic dataset, perform spatial relationship parsing processing to obtain an urban spatial semantic map.

[0029] In this embodiment, the spatial relationship parsing process first performs spatial entity extraction processing on the urban spatial basic dataset to obtain an entity set. The entity set includes at least building entities, road entities, green space entities, open space entities, plot boundary entities, public service facility entities, and traffic node entities. Subsequently, adjacency relationship calculation processing, line-of-sight relationship calculation processing, ventilation corridor identification processing, and functional connection identification processing are performed on the entity set to obtain adjacency relationship sets, line-of-sight relationship sets, environmental association sets, and functional association sets, respectively. Finally, an urban spatial semantic graph is constructed based on the above-mentioned relationship sets.

[0030] Among them, adjacency relationship is used to characterize the boundary contact, proximity or interface continuity between buildings and roads, buildings and green spaces, and between buildings; sight relationship is used to characterize important landscape interfaces, urban skyline observation points, and visual corridor accessibility; environmental relationship is used to characterize the coupling relationship between prevailing wind direction, monsoon corridor, surface openness and spatial form; and functional relationship is used to characterize traffic accessibility, public facility service radius and plot function connection.

[0031] For example, in a commercial office site located along a main urban road, spatial relationships can be analyzed to reveal that the north side is the main road interface, the east side is adjacent to a subway entrance, the south side has a landscaped river, and the west side is a cluster of high-rise residential buildings. This spatial relationship is expressed in the form of a semantic graph, so that subsequent model generation can identify the site's location and design sensitivities within the overall urban landscape.

[0032] S30: Based on the urban spatial semantic map and urban planning control conditions, perform condition encoding processing to obtain the design constraint vector.

[0033] In this embodiment, the conditional encoding process includes rule classification processing, rule mapping processing, feature embedding processing, and fusion encoding processing.

[0034] First, based on the rules governing urban planning control conditions, a set of hard constraints and a set of soft constraints are obtained. The set of hard constraints includes at least the following: land setbacks, building height limits, floor area ratio, building density, fire safety distances, sunlight control conditions, minimum green space ratio, and vehicle access control conditions. The set of soft constraints includes at least the following: architectural style preference, interface continuity, open space preference, tower-to-slab ratio preference, street-front commercial activity, and skyline harmony.

[0035] Secondly, rule mapping is performed based on the set of hard constraints to obtain a constraint mask matrix. This constraint mask matrix indicates areas in the space where construction is permitted, restricted, or prohibited, and encodes volume height, layout density, and interface control range. Feature embedding is then performed based on the set of soft constraints to obtain a preference feature vector. This preference feature vector expresses the designer's or project brief's intention to control style and spatial tendencies. Finally, a fusion encoding process is performed on the constraint mask matrix and the preference feature vector to obtain the design constraint vector.

[0036] By expressing hard and soft constraints separately, control distortion caused by mixing legal rules with design preferences can be avoided. On the one hand, the generated results must satisfy the hard constraints; on the other hand, the generated results converge towards the design tendency expressed by the soft constraints after satisfying the hard constraints.

[0037] For example, for a mixed-use commercial and business site, hard constraints may include a 15m setback on the east side, an 80m height limit along the main road on the north side, an overall plot ratio of 4.5, and a building density of no more than 35%; soft constraints may include prioritizing open space along the river, maintaining a continuous skyline on the main facade, and favoring a twin-tower layout for the towers. After conditional encoding, a design constraint vector is formed that can be fed into the generative model.

[0038] S40: Based on the urban spatial semantic map and design constraint vector, perform generative modeling processing to obtain the urban design space generation model.

[0039] In this embodiment, the urban design space generation model can adopt a conditional generative adversarial network architecture or a diffusion generation architecture. In a preferred embodiment, the generation model includes an encoder module, a conditional control module, and a decoder module. The encoder module is used to extract existing spatial features and map relationship features, the conditional control module is used to receive and inject design constraint vectors, and the decoder module is used to output the building block layout, public space outline, interface control form, and green space organization results.

[0040] Specifically, based on the urban spatial semantic map, the existing spatial features are extracted to obtain a spatial feature tensor; based on the design constraint vector, a condition injection process is performed to obtain condition control features; based on the spatial feature tensor and the condition control features, a condition generation training process is performed to obtain an initial generation model; based on the initial generation model, a compliance verification process is performed to obtain a rule compliance result; and based on the rule compliance result, a model correction process is performed to obtain an urban design spatial generation model.

[0041] In this embodiment, compliance verification can be performed by setting rule penalties, which will cause the generated schemes that violate rules such as back boundary, sunlight, height limit, and spacing to suffer higher losses during the training phase, thereby improving the compliance of the model output results.

[0042] Furthermore, in this embodiment, the model training task and the scheme inference task can also be deployed on a distributed computing framework based on a cloud-native architecture. Specifically, this includes: performing task splitting on the training and inference tasks to obtain a training task set and a generation task set; performing container encapsulation on the task set to obtain a computing service image; performing distributed scheduling on the computing service image to obtain a cloud computing instance; and performing parallel computing on the cloud computing instance to obtain the model training result and the scheme generation result. This significantly reduces training and inference time in scenarios involving concurrent design of multiple land parcels and parallel calculation of multiple parameter combinations.

[0043] S50: In response to user input conditions, generate candidate three-dimensional spatial optimization schemes using the urban design space generation model.

[0044] In this embodiment, user input conditions include plot boundaries, floor area ratio, building density, building height limits, setback control lines, architectural style preferences, street-facing interface control requirements, and public space preferences. The generated results are preferably output in an editable 3D vector format, including at least the building outline, tower location, podium boundary, building height, floor distribution, green space boundary, open space boundary, and relevant planning control indicator attribute fields.

[0045] For example, when three sets of different design style control parameters are input for the same plot of land, three candidate schemes can be output respectively: a more compact type, a more open type, and a more continuous interface type. The building block relationship and urban interface effect of each candidate scheme can be displayed in real time in the 3D interface.

[0046] S60: Based on the candidate three-dimensional spatial optimization scheme, perform multi-objective performance evaluation processing to obtain the evaluation results; and based on the evaluation results, perform feedback optimization processing to obtain the target city design spatial optimization scheme.

[0047] In this embodiment, the multi-objective performance evaluation process includes solar radiation analysis, wind environment simulation, visibility analysis, and morphological statistical processing. Specifically, solar radiation analysis is performed on candidate schemes to obtain solar radiation evaluation values; wind environment simulation is performed to obtain wind environment evaluation values; visibility analysis is performed to obtain skyline evaluation values; and morphological statistical processing is performed to obtain spatial compactness evaluation values. Then, a comprehensive scoring process is performed based on each evaluation value to obtain a comprehensive evaluation sequence, which is then sorted and filtered to obtain the evaluation results.

[0048] In this embodiment, the feedback optimization process is based not only on simulation evaluation results but also on preference learning based on designers' manual modification behavior of candidate schemes. Specifically, the process involves: extracting preferred schemes based on the evaluation results to obtain a set of preferred schemes; performing preference sampling based on user actions such as retaining, deleting, replacing, and partially editing the preferred schemes to obtain a design preference sample set; performing preference ranking learning based on the design preference sample set to obtain preference update parameters; performing model iteration based on the preference update parameters and evaluation results to obtain an optimized urban design space generation model; and finally, performing a regeneration process on the optimized model to obtain the optimized urban design space scheme for the target city.

[0049] Through the above methods, this embodiment can not only optimize the scheme based on objective indicators such as sunlight, wind environment, and skyline, but also incorporate the designer's subjective judgment, enabling the model to gradually develop design preference adaptability for specific cities, regions, and project types. This application further combines user actions such as retaining, deleting, replacing, and partially editing candidate schemes to form design preference samples, and iteratively updates the generated model based on these samples. This allows the model to gradually learn the designers' judgment tendencies in different project scenarios, thereby improving the matching degree between subsequent output results and actual design goals and reducing repetitive manual adjustments.

[0050] This application can also combine cloud-based distributed computing frameworks to containerize, elastically schedule, and execute training and inference tasks in parallel, thereby improving the computational efficiency of model training and scheme generation. This makes the application applicable not only to single-plot scheme design, but also to application scenarios such as urban renewal, area development, and parallel simulation of multiple projects.

[0051] Example 2 like Figure 2 As shown, this embodiment provides a generative intelligent computing system for urban design space optimization. This system is used to implement the method described in Embodiment 1, and includes a data construction module, a conditional coding module, a generative computing module, an evaluation and optimization module, a cloud scheduling module, and a visualization and interaction module.

[0052] The system comprises the following modules: a data construction module acquires multi-source heterogeneous data of the built environment and performs standardization, semantic alignment, and spatial relationship parsing to obtain an urban spatial semantic map; a conditional coding module performs conditional coding based on urban planning control conditions to obtain design constraint vectors; a generative calculation module performs generative modeling based on the urban spatial semantic map and design constraint vectors, and outputs candidate 3D spatial optimization schemes; an evaluation and optimization module performs multi-objective performance evaluation and feedback optimization on candidate schemes to obtain the target urban design spatial optimization scheme; a cloud scheduling module handles resource allocation, elastic scaling, and parallel scheduling for training and inference tasks; and a visualization and interaction module receives user input parameters and provides 3D display, local adjustments, and scheme comparisons of the generated results.

[0053] In one implementation, the data construction module can be deployed on a data platform server, the generation and calculation module and the evaluation and optimization module can be deployed on a cloud GPU computing cluster, and the visualization and interaction module can be deployed on a browser or desktop design terminal. This front-end and back-end decoupling enables asynchronous generation and real-time feedback under high-concurrency task requests.

[0054] Example 3 This embodiment provides a computer-readable storage medium on which a computer program is stored. When executed by a processor, the computer program is used to implement the generative intelligent computing method for urban design spatial optimization described in Embodiment 1.

[0055] The computer-readable storage medium can be a disk, optical disk, solid-state drive, read-only memory, random access memory, or other medium capable of storing program instructions. The program instructions include at least data standardization processing instructions, semantic graph construction instructions, conditional encoding instructions, generative model training instructions, candidate solution generation instructions, multi-objective evaluation instructions, and feedback optimization instructions. After the processor invokes the above program instructions, it can complete functions such as built environment data parsing, urban design scheme generation, performance evaluation, and feedback iterative optimization.

[0056] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit the scope of protection of this application. For those skilled in the art, any equivalent substitutions or modifications made to the order of steps, model architecture, composition of evaluation indicators, deployment methods, and human-computer interaction forms without departing from the technical concept of this application shall fall within the scope of protection of this application.

[0057] Example 4 To further illustrate the role of the generative intelligent computing method for urban design space optimization described in this application in terms of scheme generation efficiency, planning compliance, spatial performance balancing ability, and feedback optimization effectiveness, this embodiment selects three different types of built environment plots as test objects and sets up comparative schemes for verification.

[0058] In this embodiment, the test objects include commercial office-dominated land parcels, mixed residential land parcels, and waterfront mixed-use development land parcels. Among them, test parcel A has an area of ​​12.6 hectares. 2 Located at the intersection of a main urban road and a secondary road, the surrounding area is currently dominated by commercial office buildings and public service facilities; the test site B has an area of ​​18.4 hectares. 2 Located in a residential area redevelopment zone, the area is surrounded by numerous schools, community green spaces, and a network of secondary roads; test plot C has an area of ​​21.3 hectares. 2 Located in a riverside mixed development area, it is bordered by the waterfront to the east and adjacent to an urban expressway to the west.

[0059] In this embodiment, the multi-source heterogeneous data of the built environment includes oblique photogrammetry 3D reality model data, geographic information system vector data, urban land use classification planning map data, and urban management regulations text data. Relevant data from the three test plots and their surrounding 500m range are input into the system of this application and uniformly converted into a spatial dataset under the same coordinate reference. The 3D model has a spatial resolution of 0.5m, and the vector layer includes roads, green spaces, water systems, plot boundaries, existing building outlines, and public service facility locations. The regulations text, after structured parsing, forms rule entries such as setback control, building height limits, plot ratio, building density, green space ratio, fire safety distance, and sunlight baseline.

[0060] Furthermore, a total of 2400 sets of historical urban design case data were selected, including 1920 sets for training, 240 sets for validation, and 240 sets for testing. Each case set includes site boundaries, surrounding road network, distribution of green space and water system, existing building conditions, control indicators, and corresponding implemented or approved urban design schemes. Based on the above data, the urban design space generation model of this application was trained, with 180 training epochs, a batch size of 8, an initial learning rate of 0.0002, and loss functions including morphological reconstruction loss, rule violation penalty loss, spatial relationship consistency loss, and preference ranking loss.

[0061] To compare the technical effects of the solution in this application, the following three comparative examples are provided in this embodiment: Comparative Example 1: Using the traditional parametric rule generation method, blocks are generated only based on plot boundaries and indicators such as plot ratio, building density, and height limit, without constructing an urban spatial semantic map or performing generative learning; Comparative Example 2: A common condition generation model without introducing a hard constraint coding and soft constraint coding separation mechanism is adopted, and only the land parcel boundary and some indicators are used as unified condition inputs; Comparative Example 3: The same urban spatial semantic map and generative model structure as this application are used, but user preference samples are not introduced and feedback optimization iterations are not performed.

[0062] In this embodiment, the candidate schemes output by each method are statistically analyzed based on six indicators: rule compliance rate, sunshine compliance rate, proportion of wind environment comfort zone, skyline coordination score, scheme generation time, and number of manual adjustment rounds. The rule compliance rate represents the proportion of candidate schemes that simultaneously meet the requirements for setbacks, height limits, plot ratio, building density, fire safety spacing, and green space ratio. The sunshine compliance rate represents the proportion of individual buildings that meet the prescribed sunshine minimum. The proportion of wind environment comfort zone represents the area where pedestrian wind speeds are within the comfortable range. The skyline coordination score uses a 5-point expert scoring method. The scheme generation time is the time required for the system to output 10 sets of candidate schemes. The number of manual adjustment rounds is the average number of modification rounds required for designers to adjust the initial candidate schemes to a version ready for submission for discussion.

[0063] The average test results of each method on the three test plots are shown in Table 1.

[0064] Table 1 Comparison of test results for different methods

[0065] As shown in Table 1, the proposed scheme outperforms the other schemes in terms of rule compliance rate, sunshine compliance rate, proportion of wind comfort zone, and skyline coordination score. It also demonstrates superior results in terms of scheme generation time and number of manual adjustment rounds. These results indicate that the proposed scheme does not simply increase generation speed, but rather achieves higher usability and better overall spatial performance in a shorter time.

[0066] To further illustrate the roles of "hard constraint coding for planning rules" and "soft constraint coding for design preferences" in this application, this embodiment conducts an ablation test on whether dual-channel conditional coding is used. During the test, the urban spatial semantic graph structure, training set, number of network layers, and optimization parameters were fixed; only the form of the conditional input was changed. The test results are shown in Table 2.

[0067] Table 2 Ablation test results of dual-channel condition coding

[0068] As shown in Table 2, when only a unified hybrid coding is used, although the model can perceive the plot boundaries and some control indicators, it is prone to problems such as volume exceeding the boundaries, insufficient local spacing, and large fluctuations in the continuity of the street interface. When only hard constraint coding is used, the compliance of the scheme is significantly improved, but the skyline coordination and overall spatial style control are still insufficient. When a method of separating hard constraint coding and soft constraint coding is further adopted, on the one hand, the generated space that violates the bottom line rules can be stably compressed, and on the other hand, the guiding role of design intentions such as architectural style, open space tendency, and interface control can be preserved. Therefore, a better balance is achieved in terms of rule compliance rate and spatial performance quality.

[0069] To further illustrate the role of the "urban spatial semantic map" in this application, this embodiment compares the impact of introducing a spatial relationship map on the generated results. The test results are shown in Table 3.

[0070] Table 3 Ablation Test Results of Urban Spatial Semantic Map

[0071] Understandably, without the introduction of urban spatial semantic mapping, the generative model primarily relies on local geometric features for mass inference. While it can output several feasible solutions, its ability to identify the relationships between plots and main roads, waterfront interfaces, public space nodes, and surrounding building clusters is weak, easily leading to a disconnect between building layout and the organization of the surrounding interface. With the introduction of urban spatial semantic mapping, the model can simultaneously utilize adjacency relationships, sightline relationships, and environmental connections in mass generation, thus improving performance in areas such as the proportion of wind-friendly comfort zones, skyline harmony scores, and neighborhood interface continuity scores.

[0072] To illustrate the role of "feedback optimization processing" in this application, this embodiment records the designer's actions of retaining, deleting, replacing, and partially editing the initial candidate solutions, transforming them into preference samples, and performing three rounds of feedback learning. The results of each iteration are shown in Table 4.

[0073] Table 4 Feedback optimization iteration test results

[0074] As shown in Table 4, with the continuous accumulation of preference samples and their participation in feedback learning, the matching degree between the model output and the designer's goals continuously improves. This is manifested in the user's preferred hit rate increasing round by round, while the number of manual adjustment rounds decreasing round by round. This indicates that this application does not treat planners merely as backend map editors, but rather transforms their decision preferences into effective inputs for subsequent model iterations, thereby achieving a continuous optimization process of "generation—evaluation—feedback—regeneration".

[0075] Furthermore, to verify the feasibility of this application at the computing framework level, this embodiment deploys the model training task and inference task in a cloud-based distributed computing environment, with one management node and four computing nodes, each configured with a graphics processor with 24GB of video memory. The concurrent generation tasks for the 10 plots were executed in both a single-machine environment and a distributed environment, and the test results are shown in Table 5.

[0076] Table 5 Performance test results of the computing framework

[0077] As shown in Table 5, after adopting the cloud-based distributed computing framework, the task response latency is significantly reduced and the solution generation time is significantly shortened under multi-task concurrency conditions. Resource utilization and task completion success rate are both improved. Therefore, this application not only demonstrates good feasibility at the algorithm level but also possesses a solid foundation for engineering deployment in the context of intelligent urban design management.

[0078] Based on the test results above, it can be seen that this application, by standardizing and semantically aligning multi-source heterogeneous data of the built environment, further constructs an urban spatial semantic map, separates and encodes hard constraints of planning rules and soft constraints of design preferences, and combines multi-objective performance evaluation and user preference feedback learning, so that the generated model shows better results in terms of scheme compliance, spatial performance, landscape coordination and human-machine collaboration efficiency. The urban spatial semantic map constructed in this application enables the generated model to no longer rely solely on local geometric information for volume inference, but to simultaneously identify the correlation features between the plot and surrounding roads, green spaces, water systems, open spaces, and interface relationships, thereby improving the adaptability of the generated results to the real built environment. At the same time, by separating and encoding the hard constraints of planning rules and the soft constraints of design preferences, legal control requirements and spatial form orientation play constraining and guiding roles respectively in the generation process, thereby reducing the output of invalid schemes that violate the bottom-line rules and improving the controllability of the scheme's style expression and spatial organization. Furthermore, through feedback optimization processing, the planner's retention, deletion, replacement, and local editing behavior of candidate schemes are transformed into iteratively learnable preference samples, so that the model gradually approaches the actual design goals in subsequent generation, thereby reducing the frequency of repeated manual adjustments.

[0079] Example 5 This embodiment illustrates a specific implementation method for performing multi-objective performance evaluation on candidate three-dimensional spatial optimization schemes in this application. In this embodiment, the multi-objective performance evaluation does not merely perform static calculations on planning indicators such as plot ratio, building density, or building height of the candidate schemes. Instead, it converts the candidate three-dimensional spatial optimization schemes into multi-scale spatial evaluation maps, and performs evaluations on solar shading, physical field proxy wind environment, skyline continuity, and three-dimensional morphological compactness based on these multi-scale spatial evaluation maps, thereby obtaining a comprehensive evaluation result that can be used for ranking, screening, and feedback optimization. In this embodiment, let the k-th candidate three-dimensional spatial optimization scheme output by the urban design spatial generation model be: ; in, This represents the k-th candidate 3D space optimization scheme; This represents the i-th building block in the k-th candidate 3D spatial optimization scheme; This represents the number of building blocks in the k-th candidate 3D spatial optimization scheme; Represents the set of road spaces; It represents a collection of green spaces and open spaces; This represents the set of spaces associated with water systems or ventilation corridors. In this embodiment, the following evaluation process is performed on the candidate 3D spatial optimization schemes. 571: Spatial rasterization processing is performed on the candidate 3D spatial optimization schemes to obtain the 3D occupancy tensor and height tensor. Specifically, the candidate 3D spatial optimization schemes are mapped to a unified spatial grid to obtain the 3D occupancy tensor: ; in, Indicates the spatial location of the k-th candidate 3D spatial optimization scheme. The 3D occupancy status of the location; the value is 1 when the location is inside the building block, and 1 when the location is not inside the building block. These represent the coordinates of the spatial grid in the east-west, north-south, and vertical directions, respectively.

[0080] Furthermore, a height tensor is generated based on the height of the top of the building block: ; in, Indicates the planar position of the k-th candidate 3D spatial optimization scheme. The building height at that location; when there is no building block at the corresponding location. The value is 0.

[0081] Figure 3 The figure shows the interim results of constructing a 3D occupancy tensor for the candidate 3D spatial optimization scheme after step S71. Each colored cuboid in the figure represents an identified and encoded building block unit. Different colors are only used to distinguish different building blocks and do not indicate differences in materials, functions, or ownership. The position of each colored block in the horizontal and vertical axes corresponds to the spatial position of the building block in the site plan, and the column height in the Height direction corresponds to the 3D height or floor height encoding value of the building block. As can be seen from the figure, the building blocks in the candidate scheme are not simply recorded as two-dimensional outlines, but are converted into 3D occupancy expressions with spatial location and height attributes. Taller blocks are concentrated in the middle and rear of the site, while low and medium-height blocks are distributed in the front and lateral areas. This interim result demonstrates that this application has transformed the generated scheme into a 3D tensor structure that can be directly read by the computational model before proceeding to subsequent evaluation, providing a unified data foundation for subsequent height tensor extraction, multi-scale relationship mapping, solar shading judgment, and wind environment proxy evaluation.

[0082] Figure 4The figure shows the height tensor plane mapping results further formed by step S71. In the figure, the x and y axes represent the x and y coordinates of the candidate schemes within the plot plane, respectively. The height tensor values ​​on the right, indicated by color, represent the height tensor values ​​at the corresponding planar grid locations. Dark blue areas indicate locations with no building blocks or heights close to 0, while green, yellow, and orange areas indicate locations with taller building blocks. As can be seen from the figure, the highest height response is formed in the region with x approximately 3.0 to 4.0 and y approximately 3.6 to 4.4, with height tensor values ​​approaching 35 or higher. Medium to high value responses are formed in the regions with x approximately 4.5 to 5.5 and y approximately 2.0 to 3.0, and x approximately 1.0 to 2.0 and y approximately 2.2 to 3.0. Large areas of low values ​​around these areas correspond to roads, green spaces, open spaces, or areas without buildings. This result demonstrates that this application does not rely solely on static indicators such as total building area, floor area ratio, and building density to describe candidate schemes. Instead, it compresses the three-dimensional building height distribution into a calculable continuous height field, enabling subsequent evaluations of solar shading, skyline continuity, wind comfort zone, and spatial compactness to be calculated uniformly based on the same height tensor, thereby enhancing the spatial sensitivity of the evaluation results.

[0083] S72: Perform multi-scale spatial evaluation map construction processing based on the 3D occupancy tensor and urban spatial semantic map to obtain a multi-scale spatial evaluation map. Specifically, spatial evaluation maps are constructed at the block scale, plot scale, and building cluster scale respectively: ; in, This represents the spatial evaluation diagram of the k-th candidate 3D spatial optimization scheme at the r-th scale; r represents the scale level, which can include the block scale, plot scale, and building cluster scale. It represents a set of spatial nodes, which can correspond to building blocks, open space units, road interface units, or green space units. This represents the set of edges that represent the relationships between nodes in space. Represents the adjacency matrix; This represents the node feature matrix. The adjacency matrix can be determined as follows: ; in, This represents the adjacency relationship between the i-th spatial node and the j-th spatial node at the r-th scale; This represents the i-th spatial node at the r-th scale; This represents the planar distance between two spatial nodes; This represents the adjacency distance threshold at the r-th scale; This indicates whether there is a semantic relationship between two spatial nodes. The value is 1 when the two nodes are connected by roads, sight lines, ventilation corridors, or open spaces, and 0 otherwise.

[0084] The node feature matrix can be represented as: ; in, Indicates the first The first scale Node characteristics of a spatial node; This indicates the average building height of the spatial unit corresponding to the node; This indicates the land area of ​​the spatial unit corresponding to the node; This indicates the building coverage intensity of the spatial unit corresponding to the node; This indicates the percentage of green space or open space in the spatial unit corresponding to the node; Indicates the visual openness of the spatial unit corresponding to the node; It indicates the degree of continuity between the spatial unit corresponding to the node and the surrounding urban interface.

[0085] Figure 5 The diagram shows a multi-scale spatial evaluation map constructed in step S72. Blue dots represent building nodes, which should be labeled B1 to B10 in the final drawing, corresponding to the 10 building blocks or building cluster nodes in the candidate schemes. Orange squares represent road interface nodes, which should be labeled R1 to R6 in the final drawing, corresponding to road interface sampling nodes on the site boundary or road system. Green triangles represent green space or open space nodes, with G1, G2, and G3 corresponding to open space or green space nodes on the north, northeast, and southeast sides of the site, respectively. Solid lines represent the adjacency relationships between building nodes, characterizing the spatial proximity and interface connections of building blocks at the block or cluster scale. Dashed lines represent visual relationships or landscape sightlines, characterizing the visual accessibility between building nodes and green space / open space nodes. Dotted lines represent ventilation corridors or continuous open space relationships, characterizing the wind environment or open space connections between road interface nodes and the internal space of the site. As shown in the figure, B1 to B10 are not evaluated in isolation, but rather form a spatial relationship network with R1 to R6 and G1 to G3 through adjacent edges, visible edges, and ventilation edges. This interim result demonstrates the creativity of this application: candidate schemes are transformed into a multi-scale graphical structure that includes buildings, road interfaces, green open spaces, and their semantic relationships before evaluation. This allows subsequent evaluation to move beyond being limited to a single building block or a single planning indicator, and to simultaneously consider the organizational relationships between building groups, the continuity of road interfaces, the connection of open spaces, and ventilation corridor conditions, thereby improving the reliability of scheme selection and feedback optimization.

[0086] S73: Perform solar shading evaluation processing based on the 3D occupancy tensor to obtain solar shading evaluation values. Specifically, select a set of evaluation points. and sunshine hours By assessing the direction of sunlight at each evaluation point during different time periods, a function to determine whether sunlight is permissible is obtained. ; in, Indicates the first Evaluation points among candidate three-dimensional space optimization schemes During the period Whether or not it receives direct sunlight; Indicates from the evaluation point Along the direction of the sun Emitted rays; Indicates time period The corresponding solar direction vector; when the ray does not intersect with the building block, the value is... When the ray intersects with the building block, the value is taken as... Evaluation points The cumulative sunshine duration is: ; in, Indicates evaluation points In the Cumulative sunshine duration under candidate three-dimensional spatial optimization schemes; This indicates the time interval between two adjacent sunshine calculation periods.

[0087] Sunshine rating: ; in, This represents the solar radiation evaluation value of the k-th candidate three-dimensional spatial optimization scheme; Indicates the number of evaluation points; This indicates an indicator function that takes the value 1 when the condition inside the parentheses is true, and 0 otherwise. This represents the minimum sunshine duration requirement corresponding to evaluation point p. S74: Perform physical field proxy environment evaluation processing based on the 3D occupancy tensor and height tensor to obtain the wind environment evaluation value. In this embodiment, a physical field neural operator proxy model is used to quickly predict the pedestrian height wind environment of candidate schemes. The 3D occupancy tensor, height tensor, incoming wind speed, incoming wind direction, and surface roughness are encoded as wind environment input tensors: ; in, This represents the wind environment input tensor of the k-th candidate three-dimensional space optimization scheme under the m-th wind direction condition; This represents the reference incoming wind speed under the m-th wind direction condition; This indicates the incoming wind direction under the m-th wind direction condition; This represents the surface roughness parameter. The calculation process of the l-th layer of the physical field neural operator surrogate model is as follows: ; in, This represents the feature tensor of the l-th layer of the neural operator surrogate model; This represents the feature tensor of the (l+1)th layer; Represents the local linear transformation parameters of the l-th layer; Indicates Fourier transform; Indicates the inverse Fourier transform; Represents frequency domain variables; This represents the learnable kernel function of the l-th layer in the frequency domain; This represents a non-linear activation function.

[0088] The surrogate model outputs the predicted wind field at pedestrian height: ; in, Indicates the first The candidate three-dimensional space optimization schemes are in the... Predicted wind field under various wind direction conditions; The parameter is The physical field neural operator surrogate model is used. To reduce the impact of the surrogate model's prediction error on the evaluation results, this embodiment further calculates the wind speed prediction information penalty term: ; in, Indicates the first The wind environment uncertainty penalty term for each candidate three-dimensional spatial optimization scheme; Indicates areas highly rated by pedestrians; This represents a set of wind direction conditions; Indicates the proxy model at location The first The predicted standard deviation of the output for each wind direction condition; Indicates position The predicted wind speed vector at the location; This represents the smallest positive number to prevent the denominator from being zero. The wind environment assessment value is: ; in, Indicates the first Wind environment evaluation values ​​of candidate three-dimensional spatial optimization schemes; This indicates the lower limit of the wind speed for pedestrian comfort. This indicates the upper limit of wind speed for pedestrian height comfort; This represents the penalty weight for uncertainty.

[0089] S75: Perform skyline continuity evaluation processing based on the multi-scale spatial evaluation map and height tensor to obtain the skyline evaluation value.

[0090] Specifically, a viewpoint set Q is selected, and the skyline height sequence of the candidate scheme is obtained by sampling along the observation direction corresponding to each viewpoint q: ; in, This represents the skyline height sequence of the k-th candidate 3D spatial optimization scheme at viewpoint q; Let L represent the skyline elevation angle in the l-th sampling direction; L represents the number of sampling directions. The skyline elevation angle in the l-th sampling direction is: ; in, This indicates the position where the l-th sampling direction intersects with the candidate scheme contour; Indicates the position from viewpoint q. Horizontal distance; This indicates an extremely small positive number to prevent the denominator from being zero. A reference skyline sequence is generated based on the surrounding built environment. The skyline rating is: ; in, This represents the skyline evaluation value of the k-th candidate 3D spatial optimization scheme; This represents the reference skyline sequence corresponding to viewpoint q; Represents the L2 norm; Indicates the scale factor for the height deviation of the skyline; The second difference represents the sequence of skyline heights; Represents the variance function; The scale coefficient representing the continuity of skyline undulations.

[0091] S76: Perform three-dimensional morphological compactness evaluation processing based on the candidate three-dimensional spatial optimization scheme to obtain the spatial compactness evaluation value.

[0092] Specifically, the ratio between the total volume of the candidate building blocks and their envelope volume is calculated, and the spatial compactness evaluation value is obtained by combining this with the degree of fragmentation of the building clusters. ; in, Indicates the first The spatial compactness evaluation value of each candidate three-dimensional spatial optimization scheme; Indicates the first The volume of each building block; Indicates the envelope number The minimum three-dimensional envelope of all building blocks in the candidate three-dimensional space optimization scheme; This represents the volume of the minimum three-dimensional envelope. Indicates the first The number of building cluster connectivity components in each candidate 3D spatial optimization scheme; Indicates the number of building blocks; This indicates a very small positive number to prevent the denominator from being zero. S77: A comprehensive evaluation sequence is obtained by performing a comprehensive evaluation process based on each individual evaluation value and the penalty for violating hard constraints. First, the evaluation values ​​for solar radiation, wind environment, skyline, and spatial compactness are normalized: ; in, Indicates the first The candidate three-dimensional space optimization schemes are in the... Normalized evaluation values ​​under each evaluation indicator; Indicates the first The candidate three-dimensional space optimization schemes are in the... The original evaluation values ​​under each evaluation indicator; This indicates the type of evaluation index, with values ​​including sunshine evaluation, wind environment evaluation, skyline evaluation, and space compactness evaluation. This represents the smallest positive number that prevents the denominator from being zero.

[0093] Furthermore, the information entropy weight of each evaluation indicator is calculated: ; ; ; in, Indicates the first Under the evaluation index, the first The normalized proportion of each candidate 3D spatial optimization scheme; Indicates the number of candidate solutions; Indicates the first Information entropy of each evaluation indicator; Indicates the first The weight of each evaluation indicator; This indicates the number of evaluation indicators. Penalties for violating hard constraints are as follows: ; in, Indicates the first Penalty terms for hard constraint violations of candidate 3D space optimization solutions; Indicates the number of hard constraints; Indicates the first The penalty coefficient for a hard constraint; Indicates the first The candidate three-dimensional space optimization schemes are compared with the first... The amount of violation of a hard constraint, when the first When a hard constraint is satisfied When the first When a hard constraint is not met The final overall evaluation value is: ; in, Indicates the first The comprehensive evaluation value of each candidate three-dimensional space optimization scheme; Indicates the first The weight of each evaluation indicator; Indicates the first The candidate three-dimensional space optimization schemes are in the... Normalized evaluation values ​​under each evaluation indicator; This indicates the penalty weight for violating hard constraints; This indicates a penalty for violating a hard constraint; Indicates the penalty weight for wind environment uncertainty; This indicates a penalty for uncertainty in the wind environment.

[0094] The candidate 3D spatial optimization schemes are ranked according to their comprehensive evaluation values, resulting in a comprehensive evaluation sequence: ; in, Indicates a comprehensive evaluation sequence; A function that sorts users by their comprehensive evaluation values ​​from highest to lowest. S78: Generative model feedback guidance processing is performed based on the comprehensive evaluation sequence to obtain evaluation guidance values ​​for subsequent model iterations. Specifically, the comprehensive evaluation values ​​are used as feedback guidance signals for the generative model to update the latent variables of the generative model. ; in, Indicates the first Generative model latent variables at the next iteration; Indicates the first Generative model latent variables at the next iteration; Indicates the step size for feedback guidance; The parameters are Urban design space generation model; Represents the design constraint vector; This indicates that the generative model is used to represent the latent variables. and design constraint vector The comprehensive evaluation value of the output scheme under the influence of the action; Represents the latent variable Find the gradient.

[0095] Through the above processing, the evaluation results are no longer just used as the basis for the backend ranking of candidate schemes, but are further used as a guiding factor for the iteration of the generation model, so that the model converges towards the scheme area with peak daily targets, comfortable wind environment, continuous skyline, compact spatial form and fewer hard constraint violations in the subsequent generation process.

[0096] Depend on Figure 6 As can be seen, with the increase of feedback optimization iterations, the overall evaluation value shows a continuous upward trend, gradually increasing from 0.540 in the initial stage to 0.947 around the 20th round, indicating that the candidate schemes are continuously converging towards better spatial organization results under the guidance of feedback. Correspondingly, the penalty value for hard constraint violations gradually decreased from the initial 0.355 to 0.042, indicating that the model continuously reduced violations of bottom-line constraints such as setbacks, height limits, density, or sunlight in subsequent iterations; the number of manual adjustment rounds also gradually decreased from 6.90 to about 1.18, indicating that the matching degree between the output scheme and the design goal gradually improved, and the subsequent manual correction work was significantly reduced. The thin line and local jitter in the figure reflect the fluctuation of the sampling results in each round in the actual calculation, while the thick line indicates that the overall convergence trend is relatively stable. This shows that this application does not just rank the candidate schemes once, but can use the evaluation results to guide the model iterative optimization in reverse, so that the scheme quality, planning compliance and human-machine collaboration efficiency improve simultaneously in the continuous feedback process.

[0097] The processing procedure in this embodiment differs from methods that rely solely on a planning indicator library for compliance assessment. Through a multi-scale spatial evaluation map, building blocks, open spaces, road interfaces, and green space systems in candidate schemes are transformed into a computable spatial relationship structure. Using a physical field neural operator surrogate model, wind environment evaluation can be completed quickly while preserving the physical field characteristics. Through a confidence penalty term, unstable areas predicted by the surrogate model will not be directly adopted as high evaluation results. Through entropy-weighted comprehensive evaluation and hard constraint violation penalties, multiple evaluation indicators can adaptively determine their weights based on the discriminatory power among candidate schemes, and will not mask violations of bottom-line planning constraints due to a high single indicator. Therefore, this application establishes a feedback computational link between the generation and evaluation stages, improving the reliability of candidate three-dimensional spatial optimization scheme selection and the targeted nature of subsequent optimization.

Claims

1. A generative intelligent computing method for urban design spatial optimization, characterized in that, Includes the following steps: Acquire multi-source heterogeneous data of the built environment, and perform standardization and semantic alignment processing on the multi-source heterogeneous data to obtain a basic urban spatial dataset; Based on the aforementioned urban spatial basic dataset, spatial entity recognition and spatial relationship analysis are performed to construct an urban spatial semantic map. Based on the urban spatial semantic map and urban planning control conditions, condition coding is performed to obtain a design constraint vector. The design constraint vector includes hard constraint coding to characterize the statutory planning control content and soft constraint coding to characterize the design style and spatial tendency. An urban design space generation model is established based on the urban spatial semantic map and the design constraint vector, and candidate three-dimensional spatial optimization schemes are generated using the urban design space generation model. The candidate 3D spatial optimization schemes are converted into 3D occupancy tensors and height tensors, and a multi-scale spatial evaluation map is constructed based on the 3D occupancy tensors, height tensors, and urban spatial semantic graph. Based on the multi-scale spatial evaluation map, a multi-objective performance evaluation is performed on the candidate three-dimensional spatial optimization scheme, including solar shading evaluation, physical field proxy wind environment evaluation, skyline continuity evaluation, and three-dimensional morphological compactness evaluation, to obtain the candidate scheme evaluation results; based on the candidate scheme evaluation results, the urban design space generation model is optimized by feedback, and the target urban design space optimization scheme is output.

2. The generative intelligent computing method for urban design spatial optimization according to claim 1, characterized in that, The multi-source heterogeneous data includes at least oblique photogrammetry 3D reality model data, geographic information system vector data, urban land use classification planning map data, and urban management regulations text data. The standardization process includes data format unification, coordinate system unification, scale normalization, and attribute field normalization. The semantic alignment process includes unified label mapping for building objects, road objects, green space objects, open space objects, and planning control objects.

3. The generative intelligent computing method for urban design spatial optimization according to claim 1, characterized in that, The construction of the urban spatial semantic map includes: Extract building entities, road entities, green space entities, open space entities, land parcel boundary entities, and public facility entities from the aforementioned urban spatial basic dataset; Based on the extracted spatial entities, calculate the adjacency relationships, line-of-sight relationships, environmental relationships, and functional relationships; Based on the spatial entities and their corresponding relationships, an urban spatial semantic map is constructed to represent the urban spatial organization structure.

4. The generative intelligent computing method for urban design spatial optimization according to claim 1, characterized in that, The condition encoding based on the urban spatial semantic map and urban planning control conditions to obtain the design constraint vector includes: The urban planning control conditions are classified by rules to obtain a set of hard constraints and a set of soft constraints; The set of hard constraints is mapped to a constraint mask matrix to represent the construction boundary, height limit, density limit, spacing requirements, and sunlight limit. The soft constraint set is mapped to a preference feature vector to characterize architectural style preference, interface continuity, open space preference, and skyline harmony. The constraint mask matrix and the preference feature vector are fused and encoded to obtain the design constraint vector.

5. The generative intelligent computing method for urban design spatial optimization according to claim 1, characterized in that, The establishment of the urban design space generation model includes: Based on the urban spatial semantic map, the current spatial features are extracted to obtain the spatial feature tensor; Based on the design constraint vector, conditional control features are extracted and injected into the generative model; Based on the spatial feature tensor and the conditional control features, conditional generation training is performed to obtain an initial generation model; The compliance of the scheme output by the initial generated model with the execution rules is checked, and the initial generated model is corrected based on the check results to obtain the urban design space generation model.

6. The generative intelligent computing method for urban design spatial optimization according to claim 5, characterized in that, The urban design space generation model is a conditional generative adversarial network model or a conditional diffusion generation model. The output of the urban design space generation model includes at least the building block layout results, public space outline results, building height distribution results, and green space organization results.

7. The generative intelligent computing method for urban design spatial optimization according to claim 1, characterized in that, Based on the multi-scale spatial evaluation map, a multi-objective performance evaluation is performed on the candidate three-dimensional spatial optimization schemes, including solar shading evaluation, physical field proxy wind environment evaluation, skyline continuity evaluation, and three-dimensional morphological compactness evaluation, to obtain the candidate scheme evaluation results, including: The candidate three-dimensional spatial optimization scheme is subjected to spatial rasterization to obtain a three-dimensional occupancy tensor and a height tensor. The three-dimensional occupancy tensor is used to characterize the occupancy state of the building block in the three-dimensional spatial grid, and the height tensor is used to characterize the building height distribution corresponding to the planar grid position. Based on the three-dimensional occupancy tensor, height tensor, and urban spatial semantic map, multi-scale spatial evaluation maps are constructed at the block scale, plot scale, and building cluster scale. The multi-scale spatial evaluation map includes a set of spatial nodes, a set of relation edges, an adjacency matrix, and a node feature matrix. The set of spatial nodes includes at least building nodes, road interface nodes, green space nodes, and open space nodes. The set of relation edges includes at least adjacency relation edges, sight relation edges, and ventilation corridor relation edges. Based on the three-dimensional occupancy tensor, the candidate three-dimensional spatial optimization scheme is evaluated for solar shading to obtain a solar shading evaluation value. A wind environment input tensor is constructed based on the three-dimensional occupancy tensor, height tensor, incoming wind speed, incoming wind direction, and surface roughness. The wind environment input tensor is then input into the physical field neural operator surrogate model to obtain the predicted wind speed field at the pedestrian height. Finally, the wind environment evaluation value is obtained based on the predicted wind speed field and the wind speed prediction confidence penalty term. Based on the height tensor and multi-scale spatial evaluation map, the skyline height sequence under different viewpoints is extracted, and the skyline height sequence is continuously compared with the reference skyline sequence to obtain the skyline evaluation value. The spatial compactness evaluation value is obtained based on the total volume of building blocks, the minimum volume of three-dimensional envelope, and the number of connected components of building clusters in the candidate three-dimensional spatial optimization schemes. The solar radiation evaluation value, wind environment evaluation value, skyline evaluation value, and space compactness evaluation value are normalized, and a comprehensive evaluation value is calculated based on the information entropy weight, hard constraint violation penalty term, and wind environment uncertainty penalty term. The candidate three-dimensional spatial optimization schemes are ranked according to the comprehensive evaluation value to obtain the evaluation results of the candidate schemes. The comprehensive evaluation value is then used as a feedback guide for the iterative optimization of the subsequent urban design spatial generation model.

8. The generative intelligent computing method for urban design spatial optimization according to claim 1, characterized in that, The extraction of the preferred solution set based on the evaluation results of the candidate solutions and the feedback guidance includes: Based on the evaluation results of the candidate solutions, a set of preferred solutions is extracted; Collect user actions such as retaining, deleting, replacing, and partially editing the preferred scheme set to form a design preference sample set; Based on the design preference sample set, preference ranking learning is performed to obtain preference update parameters; The urban design space generation model is iteratively updated based on the preference update parameters and the evaluation results of the candidate schemes.

9. The generative intelligent computing method for urban design spatial optimization according to claim 1, characterized in that, Also includes: The training and inference tasks of the urban design space generation model are deployed on a distributed computing framework. Containerize and encapsulate training and inference tasks; Based on cloud computing resources, the packaged training and inference tasks are elastically scheduled and executed in parallel to accelerate model training and candidate solution generation.

10. A generative intelligent computing system for urban design spatial optimization, characterized in that, include: The data construction unit is used to acquire multi-source heterogeneous data of the built environment and perform standardization processing, semantic alignment processing and spatial relationship parsing processing to construct an urban spatial semantic map; Conditional encoding unit, used to generate design constraint vectors based on urban planning control conditions; A computational unit is used to establish an urban design space generation model based on the urban spatial semantic map and the design constraint vector, and to generate candidate three-dimensional spatial optimization schemes. The evaluation and optimization unit is used to perform multi-objective performance evaluation on the candidate three-dimensional space optimization schemes and to perform feedback optimization on the urban design space generation model based on the evaluation results. The interactive display unit is used to receive user input parameters and output a 3D visualization of the target city's spatial design optimization scheme.

Images