A building plan intelligent generation method considering indoor environment

Abstract

This invention relates to an intelligent method for generating architectural plans that considers the indoor environment, comprising the following steps: 1) establishing a dataset of functional zoning and room topology relationships, and a dataset of functional rooms and traffic space figure-ground relationships; 2) training corresponding neural network models; 3) generating floor plan layouts; 4) generating functional floor plan layouts; 5) generating 3D geometry; 6) selecting a suitable scheme; 7) selecting the optimal scheme. This invention's intelligent method for generating architectural plans that considers the indoor environment solves the core problems of "complex existing constraints, difficulty in adapting to multiple objectives, and low design efficiency" in the renovation of industrial heritage, achieving an intelligent, efficient, and quantifiable technical solution for architectural plan design to overcome the aforementioned deficiencies in existing technologies.

Description

Technical Field

[0001] This invention relates to an architectural design method, specifically to an intelligent method for generating architectural plans that takes into account the indoor environment, belonging to the field of architectural simulation and analysis technology. Background Technology

[0002] Currently, revitalizing and transforming modern industrial heritage buildings (immovable cultural relics) is a trend, aiming to create mixed-use projects integrating cultural exhibitions, creative offices, and distinctive commercial spaces. This project faces multiple complex challenges: on the one hand, the main building structure (such as red brick load-bearing walls and concrete frame columns) must be strictly preserved and cannot be altered arbitrarily, and the spatial layout is severely restricted by the existing structure; on the other hand, it must simultaneously meet the high lighting requirements of the exhibition area, the comfortable ventilation conditions of the office area, and the pedestrian flow requirements of the commercial area, making it difficult to balance multi-functionality with the preservation of historical features and the comfort of the indoor environment.

[0003] The traditional manual design process involves first manually drawing a floor plan based on existing building drawings, then repeatedly adjusting it using CAD to adapt to structural constraints, followed by manually building a 3D model using modeling software, and finally importing it into simulation tools to verify the performance of the light and thermal environment. However, in practice, the traditional process has been found to have many bottlenecks: First, the design iteration efficiency is extremely low, with the floor plan alone requiring 2-3 weeks of repeated modifications. Each adjustment requires re-checking structural constraints and fire safety regulations, and compliance risks can easily arise if not careful. Second, adapting to multiple objectives is difficult, as manual design struggles to accurately balance "functional layout, historical preservation, and environmental comfort." Third, the simulation verification process is cumbersome, with discrepancies in the data connection between manual modeling and simulation tools. Each performance adjustment requires remodeling the model, further extending the design cycle and resulting in high time and labor costs.

[0004] Therefore, in order to solve the above-mentioned technical problems, it is indeed necessary to provide an innovative method for intelligent generation of building plans that takes into account the indoor environment, so as to overcome the core problems of "complex existing constraints, difficulty in adapting multiple objectives, and low design efficiency" in the transformation of industrial heritage. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent method for generating building plans that takes into account the indoor environment. Through an integrated process of "AI model collaborative generation - automated modeling - performance prediction closed-loop optimization", it realizes the intelligence, efficiency and quantification of building plan design, and provides reliable technical support for similar existing building revitalization and renovation projects.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: a method for intelligently generating architectural plans considering the indoor environment, comprising the following steps:

[0007] 1) Establish datasets: By acquiring a large amount of existing building renovation floor plan data, the data is divided into several dataset subsets based on the building's performance characteristics; each dataset subset is further divided into a functional zoning and room topology relationship dataset, and a functional room and circulation space map-ground relationship dataset.

[0008] 2) Train the corresponding neural network model: Input the functional zoning and room topology relationship dataset from step 1) into the GNN neural network and train the learning model A1; input the functional room and traffic space map-ground relationship dataset into the pix2pix model in the generative adversarial network and train the learning model A2.

[0009] 3) Generate floor plan: Input the existing building floor plan into the learning model A1 to obtain the functional zoning and room topology relationship dataset. After translation and cleaning into image format, input it together with the existing building floor plan into the learning model A2 to generate a floor plan that conforms to building codes.

[0010] 4) Generate a functional floor plan: Optimize the floor plan generated in step 3) by aligning wall lines and smoothing room outlines using the Douglas Peucker algorithm, and output a functional floor plan that meets the requirements;

[0011] 5) Generate 3D geometry: Use Grasshopper software to generate a 3D building model containing doors and windows from the building floor plan obtained in step 4).

[0012] 6) Select a suitable solution: Then use Grasshopper software to perform daylighting and ventilation prediction on the 3D building model obtained in step 5), and output the solution that meets the prediction results.

[0013] 7) Select the optimal solution: Sort the normalized comprehensive scores of all qualified solutions from high to low, and finally select the optimal solution.

[0014] The intelligent generation method for building plans considering indoor environment of the present invention further includes the following steps: In step 1), the process of establishing the functional zoning and room topology relationship dataset is as follows: The existing building renovation floor plan is annotated using the Labelme annotation tool, and the rooms in different functional zones are annotated with different colors, and the rooms in the same functional zone are annotated with the same color, thus clarifying the functional attributes of the rooms and their zoning affiliation; the JSON format data output by Labelme annotation is parsed, and the functional labels, location coordinates and adjacency relationships of the rooms are extracted, and they are reconstructed into graph structure data adapted by graph neural network (GNN). The nodes in the neural network are rooms, the node attributes include functional labels and location information, and the edges are the adjacency relationships between rooms.

[0015] The intelligent generation method for building plans considering the indoor environment of the present invention is further described as follows: In step 1), the process of establishing the figure-ground relationship dataset of functional rooms and traffic spaces is as follows: using Photoshop software, all functional rooms in the existing building renovation plan are marked with the same color, and traffic spaces are marked with another color, generating a figure-ground relationship diagram corresponding to "functional room-traffic space"; the images of the existing building renovation plan and the figure-ground relationship diagram are matched, and a figure-ground relationship dataset of functional rooms and traffic spaces is formed in the form of "existing figure / figure-ground relationship diagram".

[0016] The intelligent generation method for building plans considering the indoor environment of the present invention is further described as follows: In step 2), the specific training method for the learning model A1 is as follows: The functional zoning and room topology relationship dataset is input into the GNN neural network. First, the "json library" in Python is called to read the label names and locations of the rooms in the json file generated by the labelme module; second, the "OpenCV library" in Python is called to read the color information of each room; finally, the "Imgraph library" in Python is used to generate the GNN graph structure; thereby training the learning model A1 for the "functional zoning-single room topology relationship".

[0017] The intelligent generation method for building plans considering indoor environments in this invention is further described as follows: In step 2), the specific training method for the learning model A2 is as follows: The dataset of the figure-ground relationship between functional rooms and traffic spaces is input into the pix2pix model in the GAN adversarial network. The U-Net structure generator in the pix2pix model extracts the features of the input existing building image through encoder downsampling, and then upsampling and skip connections through decoder to output a preliminary draft of the renovation plan with the same size as the target image. The PatchGAN discriminator in the pix2pix model is input as a spliced ​​tensor of "input image + target image" and outputs an N×N score matrix. The true probability is set for each element to constrain the rationality of local details. The training process adopts an alternating optimization strategy. First, the discriminator is trained by inputting real data pairs, namely the existing building structure map and the artificially designed real renovation map, and generated data pairs, namely the original building structure map and the renovation map output by the generator. Through the loss function, it is made capable of distinguishing the authenticity of local patches. In this way, the learning model A2 is trained to learn the figure-ground relationship mapping law of functional rooms and traffic spaces in the building plan.

[0018] The intelligent building plan generation method of the present invention, which takes into account the indoor environment, is further described as follows: Step 3) specifically involves: inputting the existing building plan with constraint data and functional requirement data into the learning model A1; the learning model A1 automatically retrieves matching case data and outputs a functional partition and room topology relationship dataset containing topologically structured data of room nodes and adjacency relationships; after being translated and cleaned into image format by the visualization library "networkx" in Python, it is input into the learning model A2 along with the existing building plan; the Dijkstra algorithm is embedded, which takes the center point of the room as the starting point and the exit as the ending point, sets the room as an inaccessible area, and uses "the evacuation distance of the public area does not exceed 30m" as a constraint condition to calculate the shortest path from each room to the safety exit in real time, generating 100 floor plans that meet fire evacuation compliance.

[0019] The intelligent building plan generation method of the present invention, which takes into account the indoor environment, further includes: step 5) specifically as follows:

[0020] 5-1) Use the "File Path" component of Grasshopper software to read the architectural floor plan output in step 4) and extract all room boundary lines; connect all room boundary lines to the "Closed" operator (curve closure judgment component), and use the "List Item" component (list element extraction component) to filter out non-closed curves separately; for non-closed curves, connect the "Simplify Curve" component (curve simplification component), which automatically connects curve breakpoints and automatically repairs defects in non-closed curves;

[0021] 5-2) Using the "Region Union" operator (planar region union operation component) in Grasshopper software, automatically extract "closed room polygons" from the building floor plan and group and mark the functional zones according to the functional zone division and labeling rules in step 1); then use the "Data Tree" component (tree data structure component) to classify the curves of the same functional zone to form a tree data of "zone name-boundary curve";

[0022] 5-3), preset the floor height parameter to 3.6m using the "Slider" component (parameter slider component) in Grasshopper;

[0023] 5-4) Connect the "room boundary curve of tree data" in step 5-2) and the "floor height parameter" in step 5-3) to the "Extrude" operator (stretch generation component), and geometrically stretch each room closed polygon along the vertical direction. The stretching height is the preset floor height parameter to generate the room volume.

[0024] 5-5) Extract the top and bottom contours from the extruded room geometry using the "Brep Faces" operator (boundary representation face extraction component) in Grasshopper software. Use the "Slider" component to preset a floor slab thickness of 120mm, and use the "Extrude" operator to extrude along the Z-axis to generate solid floor slabs and ceilings. During the extrusion process, lock the Z-axis direction using the "Vector UnitZ" operator (Z-axis unit vector component).

[0025] (5-6) Use the "Point On Curve" calculator in Grasshopper software to locate the specific position of the opening on the wall and generate simplified 3D geometry of the door and window.

[0026] The intelligent generation method for building plans considering indoor environment in this invention is further described as follows: In step 6), the specific method for daylight prediction is as follows: First, the 3D geometry obtained in step 5) is divided into objects such as rooms, walls, floors, roofs, and window openings, and corresponding material properties are attached to each object, including surface reflectance and light transmittance, and then sent to the daylight simulation engine Radiance; then, the effective usable area inside the room is arranged into a rule matrix grid as sampling points, and the EPW file of the target city is loaded. Radiance simulates the outdoor light environment based on the data in the EPW file, calculates the effective daylight area inside the room by combining the 3D model of the building and material properties, and calculates the space daylight compliance rate sDA, where,

[0027] .

[0028] The intelligent building plan generation method of the present invention, which considers the indoor environment, is further described as follows: In step 6), the specific method for ventilation prediction is as follows: The 3D geometry obtained in step 5) is sent to the ventilation simulation engine "Butterfly", and the calculation domain is set according to the indoor walls, openings, and components that may affect airflow organization; then, the EPW file of the target city is loaded, and Radiance simulates the outdoor wind environment according to the data in the EPW file. Finally, the indoor air exchange rate (ACR) is calculated by using the "ANF zoneventilation air change rate" calculator (a zone air exchange rate calculation module based on air flow network) built into Butterfly and setting outdoor wind speed, wind direction, indoor and outdoor pressure conditions according to local natural climate information.

[0029] The formula for calculating indoor air exchange rate is: n = Q / V, where n is the number of air exchanges per hour; Q is the ventilation volume (m³). 3 / h; V is the room volume, in meters. 3 .

[0030] The intelligent architectural plan generation method of the present invention, which considers the indoor environment, can also be described as follows: Step 7) specifically involves: building an evaluation module in Grasshopper, using the sDA value and ACR value of a single scheme as input, and using the "Less Than" component (less than judgment component) to determine whether the sDA is lower than 75% and whether the ACR is less than 2 times / h, and then using the "Or" component to integrate the two judgment results. That is, if either condition is met, the scheme is directly marked as unqualified; if both indicators meet the standards, the sDA and ACR are normalized according to a preset reasonable range, and the calculation is completed by embedding the normalization formula through the "Python Script" component (Python script component). If the comprehensive score is lower than 0.6, it is still judged as an unqualified scheme. At the same time, a feedback mechanism is built with the help of Anemone (Anemone loop iteration plugin) or Hoopsnake loop plugin. Unqualified schemes automatically trigger the regeneration of the plan, and qualified schemes output the corresponding results; finally, the "Sort List" component (list sorting component) is called to sort the normalized comprehensive scores of all qualified schemes from high to low, and finally select the optimal scheme.

[0031] Compared with existing technologies, the present invention has the following beneficial effects: The intelligent generation method of building plan considering the indoor environment of the present invention accurately solves the core problems of "complex existing constraints, difficulty in adapting multiple objectives, and low design efficiency" in the renovation of industrial heritage. It realizes the intelligent, efficient and quantifiable design of building plan, provides reliable technical support for similar existing building revitalization and renovation projects, greatly shortens the project cycle compared with the traditional design process, and at the same time, the automated process ensures that all schemes comply with fire evacuation regulations and historical building structural constraints, avoiding compliance risks of manual design. Attached Figure Description

[0032] Figure 1 This is a flowchart of the intelligent generation method for building plans that takes into account the indoor environment, according to the present invention.

[0033] Figure 2 This is a schematic diagram of the graph structure adapted by the graph neural network (GNN) in step 1) of the present invention.

[0034] Figure 3 This is the diagram showing the relationship between the functional rooms and the traffic space established in step 1) of this invention.

[0035] Figure 4 This is the plan layout diagram generated in step 3) of the present invention.

[0036] Figure 5 This is the planar functional layout diagram generated in step 4) of the present invention.

[0037] Figure 6 This is step 5-2 of the present invention, which forms a tree-shaped data diagram of "partition name - boundary curve".

[0038] Figure 7 It is the 3D building model containing doors and windows generated in step 5 of this invention. Detailed Implementation

[0039] Please refer to the instruction manual appendix. Figure 1 As shown, this invention is an intelligent method for generating architectural plans that takes into account the indoor environment, aiming to transform a modern industrial heritage building (formerly a textile factory, a non-movable cultural relic) into a complex project integrating cultural exhibitions, creative offices, and distinctive commercial spaces. The method includes the following steps:

[0040] 1) Establishing a dataset: By acquiring a large amount of existing building renovation floor plan data, the dataset is divided into several subsets based on the building's performance characteristics (e.g., factories, scenic spots, ancient towns, etc.). Each subset is further divided into two subsets based on the characteristics of the building's functional floor plan design: a dataset on functional zoning and room topology, and a dataset on the map-ground relationship between functional rooms and traffic spaces.

[0041] The process of establishing the functional zoning and room topology relationship dataset is as follows: The existing building renovation floor plan is annotated using the Labelme annotation tool. Rooms in different functional zones (such as exhibition areas, office areas, commercial areas, etc.) are labeled with different colors, while rooms within the same functional zone are labeled with the same color, clearly defining the room's functional attributes and zoning affiliation. The JSON format data output by Labelme is parsed to extract the room's functional labels, location coordinates, and adjacency relationships, and this data is reconstructed into graph structure data adapted to a graph neural network (GNN) (as shown in the instruction manual). Figure 2 As shown in the diagram, nodes in the neural network represent rooms, with node attributes including functional labels and location information, and edges representing the adjacency relationships between rooms.

[0042] The process of establishing the figure-ground relationship dataset for functional rooms and circulation spaces is as follows: Using Photoshop, all functional rooms in the existing building renovation floor plan are labeled with the same color, and circulation spaces with a different color, generating a figure-ground relationship diagram corresponding to "functional room - circulation space"; the images of the existing building renovation floor plan and the figure-ground relationship diagram are matched, and a dataset of the figure-ground relationship between functional rooms and circulation spaces is formed in the form of "existing figure / figure-ground relationship diagram" (as shown in the instruction manual). Figure 3 (As shown).

[0043] 2) Train the corresponding neural network model: Input the functional partition and room topology relationship dataset from step 1) into the GNN neural network and train the learning model A1; input the functional room and traffic space map-ground relationship dataset into the pix2pix model in the generative adversarial network and train the learning model A2.

[0044] Specifically, the training method for the learning model A1 is as follows: input the dataset of functional zoning and room topology relationship into the GNN neural network; first, call the "json library" in Python to read the room label names and locations in the JSON file generated by the labelme module; second, call the "OpenCV library" in Python to read the color information of each room; finally, use the "Imgraph library" in Python to generate the GNN graph structure; and train the learning model A1 for "functional zoning-single room topology relationship".

[0045] The specific training method for the learning model A2 is as follows: The dataset of the figure-ground relationship between functional rooms and traffic spaces is input into the pix2pix model in the generated GAN adversarial network. The U-Net structure generator in the pix2pix model extracts features from the input existing building image through encoder downsampling, then upsampling and skip connections through decoder, outputting a preliminary draft of the renovation plan with the same size as the target image. The PatchGAN discriminator in the pix2pix model takes a concatenated tensor of "input image + target image" as input and outputs an N×N score matrix, setting a true probability for each element to constrain the rationality of local details. The training process employs an alternating optimization strategy. First, the discriminator is trained by inputting real data pairs (existing building structure diagram and manually designed real renovation diagram) and generated data pairs (original building structure diagram and the renovation diagram output by the generator). Through a loss function, it is made capable of distinguishing between true and false local patches. This trains the learning model A2, which learns the figure-ground relationship mapping rules of functional rooms and traffic spaces in building floor plans.

[0046] 3) Generate floor plan: Input the existing building's current floor plan into learning model A1 to obtain a dataset of functional zoning and room topology relationships. After translation and cleaning, convert it into image format and input it along with the existing building's current floor plan into learning model A2 to generate a floor plan that conforms to building codes, as shown in the instruction manual. Figure 4 As shown.

[0047] Specifically, the existing building floor plan, containing constraint data (such as the building's outline, load-bearing structure location, and coordinates of unalterable historical components) and functional requirement data (such as 3 exhibition areas, 8 office areas, and 12 commercial shops, requiring clear separation between the exhibition areas and adjacent building facades, and between the office and commercial areas), is input into learning model A1. Learning model A1 automatically retrieves matching case data and outputs a functional partition and room topology relationship dataset containing topologically structured data of room nodes and adjacency relationships. This dataset is then translated and cleaned into image format using the Python visualization library "networkx" and input into learning model A2 along with the existing building floor plan, embedding the Dijkstra algorithm. This algorithm uses the room's center point as the starting point and the exit as the ending point, designating rooms as inaccessible areas, and using the constraint that "the evacuation distance in public areas does not exceed 30m" as a constraint. It calculates the shortest path from each room to the safe exit in real time, generating 100 floor plans that meet fire evacuation compliance requirements.

[0048] 4) Generate a functional floor plan: Optimize the floor plan generated in step 3) using the Douglas Peucker algorithm to align wall lines and smooth room outlines, outputting a functional floor plan that meets the requirements, as shown in the attached manual. Figure 5 As shown.

[0049] Specifically, the Douglas Peucker algorithm is used to extract continuous contour points of walls and rooms in the floor plan generated in step 3); key vertices of the contour (such as room corners and wall endpoints) are retained and redundant points with minimal impact on the overall shape are removed to achieve initial contour smoothing and eliminate jagged noise; the angles of each side of the simplified contour are calculated and constrained to a threshold range of preset target angles (such as 0° and 90°) to force wall lines to be horizontal and vertical, solving the problems of line tilting and misalignment; finally, the optimized contour point set is redrawn as a regular floor plan, ultimately optimizing the alignment of wall lines and the smoothing of room contours, and outputting a functional floor plan that meets the requirements.

[0050] 5) Generate 3D geometry: Use Grasshopper software to generate a 3D building model including doors and windows from the architectural floor plan obtained in step 4), as shown in the instruction manual. Figure 7 As shown.

[0051] This step is specifically as follows:

[0052] 5-1) Read the building floor plan output in step 4) using the "File Path" component of Grasshopper software, and extract all room boundary lines; connect all room boundary lines to the "Closed" operator, and use the "ListItem" component to filter out non-closed curves separately; for non-closed curves, connect the "Simplify Curve" component, which automatically connects curve breakpoints and automatically repairs non-closed curve defects.

[0053] 5-2) Using the "Region Union" function in Grasshopper software, automatically extract "closed room polygons" from the architectural floor plan. Then, group and label the functional zones according to the functional zoning and labeling rules in step 1), ensuring that the 3D volumes of different functional rooms are distinguishable. Next, use the "Data Tree" component to categorize the curves of the same functional zone, forming a tree-like data structure of "zone name - boundary curve," as shown in the instruction manual. Figure 6 As shown.

[0054] 5-3), preset the floor height parameter to 3.6m using the "Slider" component in Grasshopper.

[0055] 5-4) Connect the "room boundary curve of tree data" in step 5-2) and the "floor height parameter" in step 5-3) to the "Extrude" operator, and geometrically stretch each room's closed polygon along the vertical direction. The stretching height is the preset floor height parameter, and the room volume is generated.

[0056] 5-5) Extract the top and bottom contours from the extruded room geometry using the "Brep Faces" operator in Grasshopper software. Use the "Slider" component to preset a floor slab thickness of 120mm, and use the "Extrude" operator to extrude along the Z-axis to generate solid floor slabs and ceilings. During the extrusion process, use the "Vector Unit Z" operator to lock the Z-axis direction to ensure that all room volumes are perpendicular to the floor plan reference plane and avoid tilting or misalignment.

[0057] (5-6) Use the "Point On Curve" calculator in Grasshopper software to locate the specific position of the opening on the wall and generate simplified 3D geometry of the door and window.

[0058] 6) Select the acceptable schemes: Then use Grasshopper software to perform daylighting and ventilation prediction on the 3D building model obtained in step 5), output the space daylight autonomy value sDA, and obtain 12 acceptable schemes by passing the preset standard (sDA>75%).

[0059] The specific method for predicting daylight is as follows: First, the 3D geometry obtained in step 5) is divided into objects such as rooms, walls, floors, roofs, and window openings, and corresponding material properties are attached to each object, including surface reflectance and light transmittance. This is then sent to the daylight simulation engine Radiance. Next, a rule matrix grid is arranged within the effective usable area of ​​the room as sampling points, and the EPW file (i.e., meteorological data file) of the target city is loaded. Radiance simulates the outdoor light environment based on the data in the EPW file, calculates the effective daylight area indoors based on the building's 3D model and material properties, and calculates the spatial daylight compliance rate (sDA).

[0060] .

[0061] For example, the wall reflectivity is set to 0.5, the floor reflectivity to 0.2, the ceiling reflectivity to 0.8, and the glass transmittance to 0.65, and the data is sent to the daylighting simulation engine Radiance. Then, a regular matrix grid is arranged within the effective usable area of ​​the room as sampling points, with a sampling point spacing of 0.8 meters and a sampling height of 0.8 meters. Finally, the EPW file from Hangzhou is loaded. Radiance simulates the outdoor light environment based on the data in the EPW file, and, combined with the building's 3D model and material properties, calculates the effective daylighting area to be 68.4㎡ and the total room area to be 80㎡. According to the formula sDA (spatial daylight autonomy value) = effective daylighting area / total room area, the building's sDA (spatial daylight compliance rate) is calculated to be 85.5%.

[0062] The specific method for ventilation prediction is as follows: The 3D geometry obtained in step 5) is sent to the ventilation simulation engine "Butterfly", and the calculation domain is set according to the indoor walls, openings, and components that may affect airflow organization; then, the EPW file of the target city is loaded, and Radiance simulates the outdoor wind environment based on the data in the EPW file. Finally, the indoor air exchange rate (ACR) is calculated by using the built-in "ANF zone ventilation air change rate" calculator in Butterfly and setting outdoor wind speed, wind direction, indoor and outdoor pressure conditions according to local natural climate information.

[0063] The formula for calculating indoor air exchange rate is: n = Q / V, where n is the number of air exchanges per hour; Q is the ventilation volume (m³). 3 / h; V is the room volume, in meters. 3 .

[0064] For example, based on the natural climate information of a certain location, setting outdoor wind speed of 2.8 m / s, wind direction of 135° (southeast), and outdoor temperature of 22.1℃, the simulated ventilation volume is 980 m³ / s. 3 / h, room volume is 420m³ 3 The calculated air exchange rate (ACR) was 2.33 times / h.

[0065] 7) Select the optimal solution: Sort the normalized comprehensive scores of all qualified solutions from high to low, and finally select the optimal solution.

[0066] This step involves building an evaluation module in Grasshopper. Using the sDA and ACR values ​​of a single solution as input, the "Less Than" component determines whether the sDA is below 75% and the ACR is less than 2 times / hour. The "Or" component then integrates these two results: if either condition is met, the solution is directly marked as unqualified; if both indicators meet the criteria, sDA and ACR are normalized within a preset reasonable range. The normalization formula is embedded using the "Python Script" component for calculation. If the overall score is below 0.6, the solution is still considered unqualified. A feedback mechanism is built using Anemone or Hoopsnake loop plugins: unqualified solutions automatically trigger plane regeneration, while qualified solutions output their corresponding results. Finally, the "Sort List" component is called to sort all qualified solutions by their normalized overall scores from highest to lowest, ultimately selecting the optimal solution.

[0067] The normalization formula is as follows:

[0068]

[0069] X is the original data, and min(X) and max(X) are the minimum and maximum values ​​of the data, respectively.

[0070] In summary, the intelligent architectural plan generation method of the present invention, which takes into account the indoor environment, accurately solves the core problems of "complex existing constraints, difficulty in adapting multiple objectives, and low design efficiency" in the transformation of industrial heritage through an integrated process of "AI model collaborative generation - automated modeling - performance prediction closed-loop optimization". It realizes the intelligent, efficient and quantifiable design of architectural plans and provides reliable technical support for similar existing building revitalization and transformation projects.

[0071] The above-described specific embodiments are merely preferred embodiments of this invention and are not intended to limit this invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this invention should be included within the protection scope of this invention.

Claims

1. A method for intelligently generating architectural plans that takes into account the indoor environment, characterized in that: Includes the following steps: 1) Establish datasets: By acquiring a large amount of existing building renovation floor plan data, the data is divided into several dataset subsets based on the building's performance characteristics; each dataset subset is further divided into a functional zoning and room topology relationship dataset, and a functional room and circulation space map-ground relationship dataset. 2) Train the corresponding neural network model: Input the functional zoning and room topology relationship dataset from step 1) into the GNN neural network and train the learning model A1; input the functional room and traffic space map-ground relationship dataset into the pix2pix model in the generative adversarial network and train the learning model A2. 3) Generate floor plan: Input the existing building floor plan into the learning model A1 to obtain the functional zoning and room topology relationship dataset. After translation and cleaning into image format, input it together with the existing building floor plan into the learning model A2 to generate a floor plan that conforms to building codes. 4) Generate a functional floor plan: Optimize the floor plan generated in step 3) by aligning wall lines and smoothing room outlines using the Douglas Peucker algorithm, and output a functional floor plan that meets the requirements; 5) Generate 3D geometry: Use Grasshopper software to generate a 3D building model containing doors and windows from the building floor plan obtained in step 4). 6) Select a suitable solution: Then use Grasshopper software to perform daylighting and ventilation prediction on the 3D building model obtained in step 5), and output the solution that meets the prediction results. 7) Select the optimal solution: Sort the normalized comprehensive scores of all qualified solutions from high to low, and finally select the optimal solution.

2. The intelligent generation method for building plans considering indoor environment as described in claim 1, characterized in that: In step 1), the process of establishing the functional zoning and room topology relationship dataset is as follows: the existing building renovation floor plan is labeled using the Labelme annotation tool, and rooms in different functional zones are labeled with different colors, while rooms in the same functional zone are labeled with the same color, thus clarifying the functional attributes of the rooms and their zoning affiliation; the JSON format data output by Labelme is parsed, and the functional labels, location coordinates, and adjacency relationships of the rooms are extracted, and the data is reconstructed into graph structure data adapted by a graph neural network (GNN). The nodes in the neural network are rooms, and the node attributes include functional labels and location information, while the edges represent the adjacency relationships between rooms.

3. The intelligent building plan generation method considering indoor environment as described in claim 1, characterized in that: In step 1), the process of establishing the figure-ground relationship dataset of functional rooms and traffic spaces is as follows: using Photoshop software, all functional rooms in the existing building renovation floor plan are marked with the same color, and traffic spaces are marked with another color, generating a figure-ground relationship diagram corresponding to "functional room-traffic space"; the images of the existing building renovation floor plan and the figure-ground relationship diagram are matched, and a figure-ground relationship dataset of functional rooms and traffic spaces is formed in the form of "existing figure / figure-ground relationship diagram".

4. The intelligent building plan generation method considering indoor environment as described in claim 1, characterized in that: In step 2), the specific training method for learning model A1 is as follows: input the functional zoning and room topology relationship dataset into the GNN neural network. First, call the "json library" in Python to read the room label names and locations in the JSON file generated by the labelme module. Second, call the "OpenCV library" in Python to read the color information of each room. Finally, use the "Imgraph library" in Python to generate the GNN graph structure. In this way, the learning model A1 for "functional zoning-single room topology relationship" is trained.

5. The intelligent building plan generation method considering indoor environment as described in claim 1, characterized in that: In step 2), the specific training method for learning model A2 is as follows: the dataset of the figure-ground relationship between functional rooms and traffic spaces is input into the pix2pix model in the GAN adversarial network. The U-Net structure generator in the pix2pix model extracts the features of the input existing building image through encoder downsampling, and then upsampling and skip connections through decoder to output a draft of the renovation plan with the same size as the target image. The PatchGAN discriminator in the pix2pix model is input as a spliced ​​tensor of "input image + target image" and outputs an N×N score matrix. The true probability is set for each element to constrain the rationality of local details. The training process adopts an alternating optimization strategy. First, the discriminator is trained by inputting real data pairs, namely the existing building structure map and the artificially designed real renovation map, and generated data pairs, namely the original building structure map and the renovation map output by the generator. Through the loss function, it is made to have the ability to distinguish the authenticity of local patches. In this way, the learning model A2 is trained to learn the figure-ground relationship mapping law of functional rooms and traffic spaces in the building floor plan.

6. The intelligent building plan generation method considering indoor environment as described in claim 1, characterized in that: Step 3) specifically involves: inputting the existing building floor plan with constraint data and functional requirement data into learning model A1. Learning model A1 automatically retrieves matching case data and outputs a functional partition and room topology relationship dataset containing topologically structured data of room nodes and adjacency relationships. After being translated and cleaned into image format by the visualization library "networkx" in Python, the dataset, along with the existing building floor plan, is input into learning model A2. The Dijkstra algorithm is then embedded. This algorithm takes the center point of the room as the starting point and the exit as the ending point, sets the room as an inaccessible area, and uses "the evacuation distance of the public area does not exceed 30m" as a constraint to calculate the shortest path from each room to the safety exit in real time, generating 100 floor plans that meet fire evacuation compliance requirements.

7. The intelligent building plan generation method considering indoor environment as described in claim 1, characterized in that: Step 5) specifically involves: 5-1) Read the architectural floor plan output in step 4) using the "File Path" component of Grasshopper software, and extract all room boundary lines; connect all room boundary lines to the "Closed" operator, and use the "List Item" component to filter out non-closed curves separately; for non-closed curves, connect the "Simplify Curve" component, which automatically connects curve breakpoints and automatically repairs defects in non-closed curves. 5-2) Using the "Region Union" calculator in Grasshopper software, automatically extract "closed room polygons" from the building floor plan and group and mark the functional zones according to the functional zone division and labeling rules in step 1); then use the "Data Tree" component to classify the curves of the same functional zone to form a tree data of "zone name-boundary curve". 5-3), preset the floor height parameter to 3.6m using the "Slider" component in Grasshopper; 5-4) Connect the "room boundary curve of tree data" in step 5-2) and the "floor height parameter" in step 5-3) to the "Extrude" operator, and geometrically stretch each room's closed polygon along the vertical direction. The stretching height is the preset floor height parameter to generate the room volume. 5-5) Extract the top and bottom contours from the stretched room geometry using the "Brep Faces" operator in Grasshopper software. Use the "Slider" component to preset the floor slab thickness to 120mm. Use the "Extrude" operator to stretch along the Z-axis to generate solid floor slabs and ceilings. During the stretching process, lock the Z-axis direction using the "Vector Unit Z" operator. (5-6) Use the "Point On Curve" calculator in Grasshopper software to locate the specific position of the opening on the wall and generate simplified 3D geometry of the door and window.

8. The intelligent generation method for building plans considering indoor environment as described in claim 1, characterized in that: In step 6), the specific method for daylight prediction is as follows: First, the 3D geometry obtained in step 5) is divided into objects such as rooms, walls, floors, roofs, and window openings, and corresponding material properties are attached to each object, including surface reflectance and light transmittance. Then, this is sent to the daylight simulation engine Radiance. Next, a regular matrix grid of the effective usable area inside the room is arranged as sampling points, and the EPW file of the target city is loaded. Radiance simulates the outdoor light environment based on the data in the EPW file, calculates the effective daylight area indoors by combining the 3D model of the building and material properties, and calculates the space daylight compliance rate sDA. 。 9. The intelligent generation method for building plans considering indoor environment as described in claim 8, characterized in that: In step 6), the specific method for ventilation prediction is as follows: the 3D geometry obtained in step 5) is sent to the ventilation simulation engine Butterfly, and the calculation domain is set according to the indoor walls, openings, and components that may affect airflow organization; then, the EPW file of the target city is loaded, Radiance simulates the outdoor wind environment based on the data in the EPW file, and finally, the indoor air exchange rate (ACR) is calculated by simulating the air exchange rate using the built-in "ANF zone ventilation air change rate" calculator in Butterfly and setting outdoor wind speed, wind direction, indoor and outdoor pressure conditions according to local natural climate information. The formula for calculating indoor air exchange rate is: n = Q / V, where n is the number of air exchanges per hour; Q is the ventilation volume (m³). 3 / h; V is the room volume, in meters. 3 .

10. The intelligent generation method for building plans considering indoor environment as described in claim 9, characterized in that: Step 7) specifically involves: building an evaluation module in Grasshopper, using the sDA and ACR values ​​of a single solution as input, and using the "Less Than" component to determine whether the sDA is below 75% and the ACR is less than 2 times / hour, respectively. Then, using the "Or" component, the two judgment results are integrated. That is, if either condition is met, the solution is directly marked as unqualified; if both indicators meet the standards, the sDA and ACR are normalized according to a preset reasonable range, and the normalization formula is embedded in the "Python Script" component to complete the calculation. If the comprehensive score is below 0.6, it is still judged as an unqualified solution. At the same time, a feedback mechanism is built with the help of Anemone or Hoopsnake loop plugins. Unqualified solutions automatically trigger the regeneration of the plane, and qualified solutions output the corresponding results. Finally, the "Sort List" component is called to sort the normalized comprehensive scores of all qualified solutions from high to low, and finally select the optimal solution.

Images