Steel modular building structure performance prediction method and system based on graph neural network

Abstract

The application discloses a steel modular building structure performance prediction method and system based on a graph neural network, and belongs to the technical field of physical calculation. In view of the problems that multi-specialty data is split, index calculation relies on manual work, and numerical simulation is time-consuming and difficult to support rapid iteration in the prior art, integrated digital basic JSON data is first constructed based on DXF drawings; then building material cost is automatically calculated, structure performance indexes are extracted, and building cold and heat loads are calculated; finally, a graph convolution network structure proxy model and a heterogeneous graph convolution network energy consumption proxy model are trained by taking the calculated indexes as a supervision signal, the structure topology and the energy consumption heterogeneous relationship are accurately captured through the graph neural network, and rapid and accurate prediction of structure performance and building energy consumption is realized. The application realizes homologous integration of multi-specialty data and automatic calculation of indexes, and provides technical support for efficient design of steel modular building structures.

Description

Technical Field

[0001] This invention relates to the field of physical computing technology, and in particular to a method and system for predicting the structural performance of steel modular buildings based on graph neural networks. Background Technology

[0002] With the deepening of industrialization in construction, steel modular building structures, as a highly industrialized, systematic, and integrated building form, are gradually becoming a key focus of industry development. This building form uses modules as basic units, and through standardized design and factory prefabrication, it can achieve integrated construction, structure, electromechanical, and decoration disciplines, offering significant advantages such as fast construction speed, controllable quality, and green and low-carbon characteristics. However, the highly integrated nature of steel modular building structures places higher demands on the digitalization and intelligentization of their design phase.

[0003] In the design process of steel modular building structures, structural performance assessment and building energy consumption analysis are two crucial core tasks. Structural performance assessment typically employs the finite element method, establishing a refined structural analysis model to calculate mechanical indicators such as stress, displacement, and inter-story drift angle of components under load, verifying whether the design meets the requirements of standards such as the "Steel Structure Design Standard." Building energy consumption analysis relies on energy simulation software to establish models of building thermal zones, spaces, and the building envelope, simulating hourly heating and cooling loads throughout the year, providing a basis for energy-saving design. In traditional design models, these two tasks are primarily completed manually: designers must manually build BIM models based on architectural drawings, then import them into structural analysis and energy simulation software, repeatedly adjusting parameters, running calculations, and extracting results. This process is not only time-consuming and labor-intensive, but also prone to problems such as model mismatch and result deviations due to the independent construction of models by different disciplines and inconsistent data sources, severely restricting design efficiency and collaboration.

[0004] More importantly, with the increasing demands for design optimization, designers often need to compare performance and iterate to find the best solution from a large number of candidate options. In this process, the computational time required for finite element analysis and energy consumption simulation becomes prominent—a complete structural analysis and energy consumption simulation typically takes several minutes to tens of minutes. When evaluating thousands of design schemes, the total computation time increases exponentially, making it difficult to meet the timeliness requirements of design iteration. Therefore, how to quickly predict structural performance and building energy consumption without significantly sacrificing accuracy has become a core technical challenge facing the intelligent design of steel modular building structures.

[0005] To address these issues, scholars have recently attempted to introduce machine learning methods to construct surrogate models to replace time-consuming numerical simulations. For example, algorithms such as artificial neural networks, support vector machines, and random forests are used to learn the mapping relationship between design parameters and performance indicators, enabling rapid prediction of structural response or building energy consumption. However, existing surrogate model methods suffer from the following technical limitations: First, traditional machine learning methods struggle to effectively handle the graph structure features in building structural data—structural components form complex topological relationships through node connections. This spatial association information significantly impacts mechanical performance, but conventional machine learning models (such as multilayer perceptrons and random forests) cannot directly capture this structured information. Second, building energy consumption systems involve multiple entities, including thermal zones, spaces, envelope structures, ground, and outdoor environments, with complex multi-faceted relationships such as inclusion, connection, and attribution. Existing surrogate model methods lack the ability to model such heterogeneous graph structure data, resulting in limited prediction accuracy. Third, current research typically separates structural prediction from energy consumption prediction, failing to achieve collaborative modeling within a unified data framework, which hinders subsequent integrated applications of multi-objective optimization.

[0006] Therefore, there is an urgent need for a prediction method that can fully utilize the characteristics of structured data of steel modular building structures. By constructing a proxy model that adapts to the features of the structural system graph and the heterogeneous graph features of the energy consumption system, a rapid and accurate prediction of structural performance and building energy consumption can be achieved, providing support for efficient design optimization. Summary of the Invention

[0007] This invention provides a method and system for predicting the structural performance of steel modular buildings based on graph neural networks. It aims to solve the technical problems in the design of existing steel modular buildings, where structural performance evaluation and building energy consumption analysis rely on manual modeling and numerical simulation, which are too time-consuming. Furthermore, existing surrogate models cannot effectively handle the graph structural features of the structural system and the heterogeneous graph structural relationship of the energy consumption system, resulting in low prediction efficiency, limited accuracy, and difficulty in meeting the needs of large-scale design optimization.

[0008] To address the aforementioned technical problems, this invention provides a method for predicting the structural performance of steel modular buildings based on graph neural networks, comprising the following steps:

[0009] D1. Extract geometric data from DXF drawings, fill in the code and information fusion of the three disciplines of architecture, structure and energy consumption, and construct integrated digital basic JSON data;

[0010] D2. Based on the integrated digital foundation JSON data, automatically calculate the cost of building materials, automatically build a structural finite element analysis model and extract structural performance indicators, and automatically build an energy consumption analysis model and calculate the building's heating and cooling loads.

[0011] D3. Based on the integrated digital foundation JSON data, construct homogeneous graph data and heterogeneous graph data. Using the structural performance index and building heating and cooling load calculated in step D2 as supervision signals, construct and train a structural proxy model based on graph convolutional network and an energy consumption proxy model based on heterogeneous graph convolutional network, respectively. The structural proxy model is used to predict whether the structural performance exceeds the limit, and the energy consumption proxy model is used to predict the building heating and cooling load.

[0012] The present invention also provides a steel modular building structure performance prediction system based on graph neural networks, including an integrated digital foundation construction module, a multi-professional index automatic calculation module, and a graph neural network proxy model construction module, which are respectively used to execute steps D1, D2, and D3 in the steel modular building structure performance prediction method based on graph neural networks.

[0013] This invention provides a method and system for predicting the structural performance of steel modular buildings based on graph neural networks. Using DXF drawings as the sole input, it extracts information from the drawings and processes and fuses information from multiple disciplines to construct an integrated digital foundation JSON data covering architecture, structure, and energy consumption. This achieves unified integration and standardized storage of information from multiple disciplines. Based on this unified data source, it automatically performs hierarchical and cumulative calculations of building material costs, automatically builds structural finite element models and extracts mechanical performance indicators, automatically maps energy consumption models, and simulates annual heating and cooling loads, achieving fully automated and accurate calculations from data to underlying indicators. Furthermore, this invention constructs homogeneous and heterogeneous graph data based on the integrated digital foundation, training a graph convolutional network (GCN) structural proxy model and a heterogeneous graph convolutional network (HGCN) energy consumption proxy model. The graph neural network accurately captures the graph topological features of the structural system and the heterogeneous graph structural relationships of the energy consumption system, enabling rapid determination of whether structural performance exceeds limits and accurate prediction of building heating and cooling loads, replacing traditional time-consuming finite element analysis and energy consumption simulation. This invention effectively solves the technical problems of fragmented multi-disciplinary data, reliance on manual calculation of indicators, and excessive time consumption in numerical simulation that makes it difficult to support rapid iteration in the existing technology. It provides accurate, unified, and scalable technical support and theoretical basis for the collaborative design and efficient optimization of steel modular building structures. Attached Figure Description

[0014] Figure 1 A flowchart illustrating the performance prediction method for steel modular building structures based on graph neural networks provided in an embodiment of the present invention;

[0015] Figure 2 This is a schematic diagram of the parameterized filling process provided in an embodiment of the present invention;

[0016] Figure 3 This is a schematic diagram of the IFC file structure provided in an embodiment of the present invention;

[0017] Figure 4 This is a schematic diagram of the initial IFC model provided in an embodiment of the present invention. Figure 4 (a) is the initial IFC model corresponding to a standard floor building drawing. Figure 4 (b) is the initial IFC model corresponding to another standard floor building drawing;

[0018] Figure 5 This is a schematic diagram of the collision-corrected IFC model provided in an embodiment of the present invention. Figure 5 (a) is the modified IFC model corresponding to the first-floor architectural drawings. Figure 5 (b) is the modified IFC model corresponding to the two- to four-story building drawings;

[0019] Figure 6 This is a schematic diagram of independent and non-independent space modules provided in an embodiment of the present invention. Figure 6 (a) is a schematic diagram of an independent space module. Figure 6 (b) is a schematic diagram of a non-independent space module;

[0020] Figure 7 This is a tree structure fusion hierarchy framework diagram provided in an embodiment of the present invention;

[0021] Figure 8 An example of an embodiment of the present invention is a year-round outdoor temperature curve of a building.

[0022] Figure 9 This is a year-round temperature curve of a certain hot zone provided in an embodiment of the present invention;

[0023] Figure 10 This is a heat load curve diagram of a certain hot zone provided in an embodiment of the present invention;

[0024] Figure 11 This is a cooling load curve diagram inside a certain hot zone provided in an embodiment of the present invention;

[0025] Figure 12 A schematic diagram of structural performance indicators provided for embodiments of the present invention;

[0026] Figure 13 The confusion matrix diagram of the structural proxy model GCN provided in the embodiments of the present invention;

[0027] Figure 14 The graph shows the fitting curves between the predicted and actual values ​​of the HGCN energy consumption proxy model provided in this embodiment of the invention. Detailed Implementation

[0028] To make the technical problems, technical solutions, and beneficial effects to be solved by this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific examples described herein are merely illustrative and not intended to limit the scope of this application.

[0029] Specifically, such as Figure 1 As shown in the flowchart, this embodiment provides a method for predicting the performance of steel modular building structures based on graph neural networks, which includes the following steps:

[0030] D1. Constructing an integrated digital foundation: Extract geometric data from DXF (Drawing Exchange Format) drawings, fill in the data with codes from the three disciplines of architecture, structure, and energy consumption, and integrate the information to construct an integrated digital foundation JSON (JavaScript Object Notation) data.

[0031] D2. Automatic Calculation of Multiple Professional Indicators: Based on the integrated digital foundation JSON data, automatically calculate the cost of building materials, automatically build a structural finite element analysis model and extract structural performance indicators, and automatically build an energy consumption analysis model and calculate the building's heating and cooling loads.

[0032] D3. Constructing a graph neural network proxy model for prediction: Based on the integrated digital foundation JSON data, construct homogeneous graph data and heterogeneous graph data, and use the structural performance index and building heating and cooling load calculated in step D2 as supervision signals to construct a structural proxy model based on a graph convolutional network and an energy consumption proxy model based on a heterogeneous graph convolutional network, respectively, and train them. The structural proxy model is used to predict whether the structural performance exceeds the limit, and the energy consumption proxy model is used to predict the building heating and cooling load.

[0033] This invention addresses the technical problems of fragmented multi-disciplinary data, reliance on manual calculation of indicators, and the excessive time required for traditional structural analysis and energy consumption simulation in the design phase of steel modular building structures, which hinders rapid iteration. First, step D1 automatically extracts geometric data from DXF drawings. This data is then integrated into a unified digital foundation JSON data structure through coding and tree-structure fusion of architectural, structural, and energy consumption disciplines. This achieves unified integration and standardized storage of information from multiple disciplines, providing a unified data source for subsequent proxy model training. Based on this, step D2 automatically calculates building material costs hierarchically and cumulatively using this unified data source, automatically builds an OpenSees structural finite element analysis model, and extracts inter-story drift angles and mechanical indicators such as component strength, stiffness, and stability. The EnergyPlus energy consumption model was used to simulate the annual heating and cooling loads, achieving fully automated and accurate calculations from data to underlying indicators, providing high-quality training samples for the surrogate model. Step D3 then constructed isomorphic and heteromorphic graph data based on an integrated digital infrastructure, abstracting the building structure as a isomorphic graph and the building energy consumption system as a heteromorphic graph. Graph convolutional network (GCN) structural surrogate model and heteromorphic graph convolutional network (HGCN) energy consumption surrogate model were trained respectively. Through graph neural networks, the nonlinear mapping relationships between structural parameters and structural compliance, and between building envelope parameters and building heating and cooling loads, were accurately learned. This enabled rapid determination of whether structural performance exceeded limits and accurate prediction of building heating and cooling loads, providing efficient alternative model support for subsequent multi-objective optimization iterations.

[0034] Step D1: Building an Integrated Digital Infrastructure

[0035] Step D1 specifically includes the following steps:

[0036] D11. Drawing Information Extraction and Processing: Input DXF drawing, extract geometric data of grid, walls, doors and windows, module frames and hot zone frames according to line type layer based on Dxfgrabber library, calculate center line coordinates and endpoints, and output initial component data in JSON format;

[0037] D12. Multi-disciplinary information processing: The initial JSON data is processed by three disciplines: the architecture discipline performs 19-bit encoding, parameter filling, collision correction and component classification; the structural discipline calculates room load and fills in the mechanical materials of components; the energy consumption discipline performs 23-bit encoding, space identification and enclosure structure filling, and outputs JSON data for each discipline.

[0038] D13. Multi-disciplinary information integration: Based on the five principles of cross-disciplinary integration, multi-level, classification, master-slave dependency, and coding collaboration, a tree-structured integration method is adopted to integrate the JSON data of the three disciplines according to the hierarchy and output integrated digital basic JSON data.

[0039] The integrated digital infrastructure built in this step effectively breaks down data barriers between multiple disciplines, realizes intensive data management and efficient collaboration, solves the problems of poor data reusability and low correlation, verifies the feasibility of data extraction, processing and fusion methods, and lays a solid foundation for the full life cycle collaboration and design optimization of steel modular building structures.

[0040] (1) D11, Drawing Information Extraction and Processing

[0041] This step can be divided into the following sub-steps:

[0042] D111. Obtain the floor plan file of the steel module building structure as the initial file;

[0043] D112. Preprocess the initial file;

[0044] D113. Extract and process drawing information from the preprocessed initial file.

[0045] In step D111, to ensure the efficient transfer of multi-disciplinary information extracted from the drawings to subsequent multi-disciplinary information processing and fusion stages, a unified data format needs to be clearly defined. This embodiment selects JavaScript Object Notation (JSON) as the data format for integrating multi-disciplinary information in steel modular building structures. JSON uses key-value pairs as its core data structure, with keys represented by strings and values ​​that can be flexibly selected (strings, numbers, booleans, arrays, objects, etc.). Its syntax is clear, with low redundancy, and requires no complex parsing logic.

[0046] Considering the cross-disciplinary, hierarchical, and diverse data characteristics of steel modular building structures, the adaptability of the JSON format is mainly reflected in three aspects: First, it has strong openness in transmission, and can encapsulate heterogeneous data such as building geometry data, structural mechanical parameters, and energy consumption and thermal performance without discrimination, solving the industry pain point of diverse data sources from multiple disciplines; second, it has good cross-platform and cross-language compatibility, and can be easily parsed and generated by various mainstream programming languages ​​(JavaScript, C / C++, Python, etc.), realizing barrier-free data interaction between different professional design software and data processing platforms, and adapting to multi-disciplinary collaborative scenarios; third, it is standardized, concise, and highly extensible, and can dynamically expand data attributes and integrate complex hierarchical related data through object nesting and array combination, which can match the fusion needs of multi-level data in the design stage of steel modular building structures.

[0047] This embodiment uses the CAD (Computer-Aided Design) floor plan file of the steel module building structure as the initial file. All kinds of building components (walls, columns, doors and windows, dimensions, etc.) are stored in line layers according to industry standards. Therefore, the extraction method based on vector graphics is adopted to directly extract the structured data of the building drawings.

[0048] Currently, there are two main formats for CAD drawings. The DXF format meets the requirements of this embodiment for extraction accuracy, convenience, and data integrity. Therefore, this embodiment extracts building-related data based on CAD drawings in DXF format.

[0049] In step D112, the drawings are preprocessed to facilitate subsequent information extraction. The drawings need to be divided according to standard floors. For example, the drawings for floor 1 and floors 2-4 represent two standard floors, and two standard floor DXF vector drawings are created. Furthermore, the line types of the drawings need to be standardized: the grid lines should use the "DOTE" line type, walls the "WALL" line type, doors the "DOOR" line type, windows the "WINDOW" line type, and text the "PUB_TEXT" line type. This embodiment optimizes the analysis based on a determined module layout. To facilitate subsequent information extraction, determined module frames are added to the steel module building structure DXF drawings, using the "MODULE" line type. To facilitate the extraction of thermal zone information, determined thermal zone frames are added, using the "ZONE" line type.

[0050] In step D113, Dxfgrabber is a lightweight DXF graphics file parsing library developed using the Python programming language. Its core purpose is to enable fast reading, parsing, and data extraction of DXF format files generated by professional design software such as AutoCAD. It does not rely on the runtime environment of commercial design software such as AutoCAD and is one of the commonly used tools for processing DXF vector files in the open source field.

[0051] This embodiment uses the floor plans of a steel modular integrated dormitory building (floors 2-4) as an analysis case. It utilizes the Dxfgrabber library in Python programming to automatically extract component information from DXF vector drawings.

[0052] (2) D12, Multidisciplinary Information Processing

[0053] This embodiment specifically divides information into three categories: architectural, structural, and energy consumption, and processes each category of information using different processes.

[0054] Architectural information is the core of integrated data. This part of information processing includes four parts: component coding, architectural design information filling, module information processing, and component information filling.

[0055] 1) Component coding

[0056] This embodiment encodes architectural data into 19-digit long numeric codes, employing a hierarchical segmented interpretation rule to define the core attributes of components digit by digit, thereby achieving accurate identification of steel modular integrated building components: The coding rules are shown in Table 1:

[0057] Table 1. Architectural Professional Data Coding Rules

[0058]

[0059] For example, the code 1010101010101010101 represents the first column in the southwest direction of the first geometric type of column component in the first module of the first module in the first standard floor of the steel modular building structure.

[0060] 2) Filling in architectural design information

[0061] Architectural design information is a comprehensive description of a building, including its name, area, height, number of floors, etc. This information is entered directly into JSON fields.

[0062] 3) Module Information Processing

[0063] The information extracted from the drawings only includes module frames and wall / door / window information. Further wall line cutting is needed according to the module layout diagram to obtain module data, including module location and the geometric position of internal wall / door / window components. Since the modules are produced according to standardized procedures, they are matched and categorized based on their data. Modules with the same or equivalent geometric features are grouped into the same category, resulting in a total of six equivalent geometric cases. For example, mirror images of each other belong to the same category, and rotated versions of each other belong to the same category of features.

[0064] 4) Component information filling

[0065] 1. Assignment of parameterized information and preliminary calculation

[0066] The information extracted from the drawings mainly includes the length and width dimensions of the modules, and the centerline information of walls, doors, and windows. Further parametric filling is required, primarily involving the cross-sections, materials, geometric information, and location distribution parameters of walls, doors, windows, and other components (floors, roofs, etc., filled according to actual needs). For the parametric data, the starting point position, stretching length, and stretching direction of each specific component are calculated based on predefined module construction rules and component distribution locations. Then, specific information about the component is assigned based on its cross-sectional shape, cross-sectional dimensions, material type, material name, and whether it has openings. The process is as follows: Figure 2 As shown.

[0067] 2. Collision Correction

[0068] Parametric calculations only preliminarily determine the positions of components. Since the physical relationships between components are not yet clearly established, collision problems may occur, thus failing to effectively guide actual production and construction. To solve this critical problem, this embodiment first performs systematic collision detection on the model to accurately identify collision areas between components. Subsequently, based on the detection results, the geometric parameters and spatial positions of the building components are adjusted to achieve adaptive adjustment of the model, ultimately generating conflict-free building data information that meets engineering requirements.

[0069] This embodiment employs an axis-aligned bounding box (AABB) collision detection algorithm. This algorithm creates an axis-aligned bounding box for each component and calculates the minimum and maximum coordinates of each bounding box. By projecting the bounding boxes of two components onto three coordinate planes, if the projections on all three coordinate axes overlap, the two components are determined to have collided.

[0070] To accelerate detection, a recursive method is used to construct a BVH (Bounding Volume Hierarchy) tree. This organizes the AABB bounding boxes of all components in the module into a tree structure based on parent-child hierarchies. Upper-level nodes are "large bounding boxes" containing multiple lower-level nodes, and lower-level nodes are "small bounding boxes" containing a single object. During collision detection, it first checks if the upper-level large bounding boxes intersect. If they do not intersect, the detection of all lower-level small bounding boxes is skipped, thus improving detection efficiency.

[0071] In practice, a physical model of the building must first be obtained, and then bounding boxes are constructed based on the geometric patch data of the components to achieve collision detection. To ensure the model has standardized universality, this embodiment uses IFC standard files for geometric data processing. IFC model files are based on a plain text file format, follow the ISO10303 standard (STEP standard), and are characterized by high readability, high compatibility, and clear structure, and are widely used in the exchange and sharing of engineering data. Such files can be viewed and edited using ordinary text editors, and have good readability and universality. Their file structure begins with "ISO-10303-21;" and ends with "END-ISO-10303-21;", and is divided into two main parts: a header segment and a data segment, such as... Figure 3 As shown.

[0072] The header segment begins with "HEADER;" and ends with "ENDSEC;". The code snippet between them contains metadata information about the IFC file, describing its basic attributes. For example, "FILE_DESCRIPTION" describes the file's purpose, and "FILE_NAME" contains the file name, timestamp, and authoring tools. "FILE_SCHEMA(('IFC4'));" indicates that the IFC file uses the IFC4 standard. The data segment begins with "DATA;" and ends with "ENDSEC;", and is the core of the IFC file, containing specific project information and object definitions. Each object is labeled with a unique identifier, and its attributes and relationships are recorded in STEP file format. Figure 3 In the STEP file notation, object "#16" is defined as IFCCOLUMN (representing a column in the Building Information Model). Among the definition information contained in this field, "GlobalId" (a unique identifier) ​​is used to uniquely identify the column object; "OwnerHistory" is typically used to record the creator, software version, and timestamp; "Name" represents the column's name; "Description" describes the column's information; "ObjectType" typically provides a more specific component classification definition; "ObjectPlacement" defines the column's spatial location, described using identifier #44 in the file; and "Representation" defines the column's geometry, defined using identifier #38 in the example file. Additionally, the "Tag" field is a user-defined tag used to attach extra identification information; the PredefinedType property is empty.

[0073] In component geometry generation, sweep modeling is a typical and widely used modeling method. This method efficiently generates 3D solid models by defining the component's 2D closed cross-section and specifying the sweep path. Furthermore, to achieve precise placement of modular components, the generated solids are first positioned in a local coordinate system and then positioned in the global target space through affine transformations (such as translation, rotation, and mirroring), ensuring the correctness and consistency of their spatial pose. The IFC standard provides a comprehensive coordinate transformation mechanism. Using the IfcLocalPlacement method, the position of the component in the local coordinate system can be explicitly specified, and the origin and orientation of this coordinate system can be described using IfcAxis2Placement. This mechanism allows the component to be transformed from the local coordinate system to the global coordinate system.

[0074] IfcOpenShell is a free and open-source code library that provides analysis and creation functions. It can be used to process industrial basic IFC files, parse and retrieve relevant data within them, and perform tasks such as reading, creating, and modifying IFC files. This embodiment uses the Python-based IfcOpenShell library to map the previously obtained initial building JSON information to IFC components, resulting in an initial IFC model as follows: Figure 4 As shown, Figure 4 In the middle (a) and (b), the initial IFC models corresponding to one standard floor building drawing (floor 1) and another standard floor building drawing (floors 2-4) are respectively.

[0075] After obtaining the initial solid model, the Geom function from the IfcOpenShell library is used to extract geometric mesh data from the IFC components to obtain triangular facet information. Then, AABB bounding boxes are constructed to achieve collision detection. After collision detection, the model returns detailed collision information, including the accurate location and relevant dimensional data of the collisions. Based on this information, an adaptive adjustment method is proposed. This method consists of two stages. In the first stage, components are grouped according to their category, and the starting point position, cross-sectional dimensions, and extension length of components such as columns, beams, walls, doors, roofs, and floors are adjusted respectively, and the module data is integrated and updated. In the second stage, collision calculations are performed between the adjusted walls and the original door and window IFC component solids, and door and window opening information is added to the walls.

[0076] The first stage of adaptive adjustment requires defining the adjustment priorities of components, from highest to lowest: intra-module connectors, frame components, and wall panel components. When collisions occur between components of different priorities, the components are adjusted one by one in order of priority from lowest to highest. For components that have collided, a Boolean difference operation is performed between the low-priority solid model and the collision area to achieve adaptive adjustment of the components.

[0077] In the second stage of adaptive adjustment, the priority order from high to low is set as door and window components, then wall panel components. Openings in the wall panels are then made based on the collision information between doors / windows and wall panels. After the two-stage adaptive adjustment, JSON data is obtained. The collision-corrected IFC model based on the IfcOpenshell library mapping is as follows: Figure 5 As shown, Figure 5 In the middle (a) and (b), the modified IFC models corresponding to the 1st floor building drawings and the 2nd to 4th floor building drawings are respectively.

[0078] 3. Component Classification

[0079] To classify the geometric features of building components within the module, this embodiment first extracts the three-dimensional vertex coordinates of the components using the Geom function in the IfcOpenShell library. Based on this, it iterates through the equivalent geometric data of the components' X, Y, and Z axes, including their mirror images and rotations. For each type of geometric data, the SHA256 algorithm is used to calculate a hash value. Finally, the minimum hash value among the components themselves and all their mirror images is taken as the component's geometric fingerprint. Components with the same hash calculation value are classified as belonging to the same category.

[0080] Structural information processing mainly includes three aspects: filling in structural design information, calculating module loads, and filling in structural material information for components.

[0081] 1) Filling in structural design information

[0082] This section provides a comprehensive description of the building's structural characteristics, including basic structural design information such as site type, seismic intensity, and functional room dead / live coefficients. This information can be directly entered into JSON fields.

[0083] 2) Module load calculation

[0084] To calculate the load of the module, room identification of the architectural drawings is required first. In this embodiment, a counterclockwise room outline detection method is used. This method first identifies the text and text coordinates within the grid area, then emits rays to the left based on the coordinates, takes the first intersecting line as the initial outline, and searches for the next outline in a counterclockwise direction from the endpoint to the starting point. This process is repeated until the initial outline is found, thereby determining the room outline.

[0085] Since a module may contain multiple rooms, or a room may contain multiple modules, we first traverse the rooms to find the modules they belong to. For rooms that contain multiple modules, we traverse the modules to find the modules contained in the room, thereby determining the room type and area range contained in each module.

[0086] For different room types, the corresponding dead and live load parameters are searched using the functional room dead and live load factor table. To calculate the area of ​​each room, this embodiment uses the Polygon.area function based on the Shapely library to calculate the area of ​​the enclosed region. Based on the weighted calculation of the room area and the room dead and live load factor, the average dead and live load factor of the ground is calculated using the trapezoidal triangle load distribution method to calculate the edge load.

[0087] 3) Filling of structural materials for components

[0088] The mechanical properties of structural components are crucial information for structural engineers. Therefore, it is necessary to supplement the structural performance of structural components, such as columns and beams, with information such as Poisson's ratio and modulus of elasticity.

[0089] Energy consumption professional information processing mainly includes five aspects: energy consumption information coding, energy consumption condition information filling, spatial identification, thermal zone spatial classification, and building envelope material filling.

[0090] 1) Energy consumption information coding

[0091] This embodiment encodes energy consumption data into a 23-bit long numeric code, employing a hierarchical segmented interpretation rule to define the core attributes of the object bit by bit, thereby achieving accurate identification of energy consumption information objects. The encoding rules are shown in Table 2 below.

[0092] Table 2 Energy Consumption Professional Data Coding Rules

[0093]

[0094] For example, the code 10101020101010101010101 indicates the first enclosure structure in the south direction of the first hot zone in the first hot zone in the first space in the first hot zone in the first hot zone in the first space in the first space in the first type of wall in the first type of wall in the first standard floor of the modular building.

[0095] 2) Filling in energy consumption information

[0096] This section describes the preliminary information for building energy consumption calculations, mainly including information required for energy consumption design such as altitude, average annual temperature, city location, and latitude and longitude. This information is directly entered as a JSON field.

[0097] 3) Spatial recognition

[0098] To simplify the spatial division of modular buildings, this embodiment defines a module space as the minimum space. Whether a module is an independent space is determined by whether it has walls in all four cardinal directions (north, south, east, and west). If it is an independent space, then that module is considered the space. For modules with fewer than four wall locations, the room to which the module belongs is queried, and that room is ultimately considered the space. A diagram illustrating independent and non-independent spaces is shown below. Figure 6 As shown, Figure 6 In the diagrams (a) and (b), independent space modules and non-independent space modules are shown respectively.

[0099] 4) Spatial classification of hot zones

[0100] To classify and integrate hot zones, the hot zones are compared based on the module component information corresponding to each space within the hot zone. Hot zones with equivalent geometric characteristics belong to the same category. Similarly, for each space within the same hot zone, the module information corresponding to the space is compared, and spaces with equivalent geometric characteristics are grouped into the same category.

[0101] 5) Filling in the information of the building envelope

[0102] Building envelope information is the core of energy consumption. In the energy consumption calculation model, the main building envelopes are five types: roof, ground, wall, door, and window. In this embodiment, combined with the standardized production characteristics of modular buildings, the building envelope materials are set consistently, and the materials and thermal performance are assigned according to the material names of each building envelope corresponding to the modules in the building profession.

[0103] (3) D13, Multi-disciplinary information integration

[0104] Cross-disciplinary integration: The mechanical performance information of structural components is integrated one-to-one with the building components in the architectural field that have completely overlapping geometric shapes and spatial positions; the thermal zone spatial division and thermal performance information of the building envelope are integrated in parallel with the data of the architectural and structural fields.

[0105] Multi-level integration: For architectural and structural engineering, integration is carried out in three levels: standard layer, module, and component; for energy consumption, integration is carried out in three levels: thermal zone, space, and building envelope.

[0106] Category-based fusion: Standard floor data is fused to achieve floor classification; for architecture or structural engineering, modules are classified according to the geometric distribution information of internal components, and components are classified according to their functional categories and geometric information; for energy consumption engineering, thermal zones are classified according to the spatial distribution of internal thermal zones and the information of the building envelope, spaces are classified according to the distribution of components within each space, and building envelopes are classified according to their types and geometric information.

[0107] Master-slave dependency integration: The overall building dimensions, module layout, and room functions are defined as the master model data; the geometric section parameters of components and the energy consumption model and building envelope materials are defined as slave model data, so that the slave model data is stored and managed in dependence on the corresponding master model data.

[0108] Coding Collaboration and Integration: A unified coding system is established, with building type identification, standard floor, and floor information as common components, professional categories, modules, hot zones, and spatial information as meso-level coding, and component and enclosure structure information as micro-level coding.

[0109] In the data from the architecture, structure, and energy consumption disciplines, the building type identifier, standard floor, and floor information are consistent, so they are designed as common areas. In subsequent information from each discipline, building and structural information is assigned based on components, while energy consumption information is assigned based on thermal zone space envelope, etc. Therefore, a tree structure is used for information fusion, with its hierarchical framework as follows: Figure 7 As shown.

[0110] Step D1 completes the construction of the integrated digital infrastructure, providing data support for the automatic calculation of indicators. Based on this infrastructure, the following steps will realize the automated calculation of three major categories of indicators: economic, energy consumption, and structural indicators for the integrated steel structure module.

[0111] II. Step D2: Automatic Calculation of Multiple Professional Indicators

[0112] Step D2 specifically includes:

[0113] D21. Automatic Calculation of Economic Indicators: Based on the module classification characteristics of integrated digital foundation JSON data, a step-by-step accumulation method from module level, standard level to building level is adopted to multiply the volume and unit price of structural components, enclosure structural components and other fixed components and sum them to calculate the total cost of building materials.

[0114] D22. Automatic calculation of building energy consumption indicators: Extract the spatial geometry information of the thermal zone and the thermal parameters of the building envelope from the integrated digital foundation JSON data, automatically generate IDF files through Eppy mapping, call EnergyPlus to perform building heat balance simulation, and calculate the sum of the annual cumulative heating load and cooling load.

[0115] D23: Automatic Calculation of Structural Indicators: Extracts geometric information and mechanical parameters of structural components from integrated digital foundation JSON data, automatically generates node elements, applies boundary conditions and seismic loads in OpenSees, solves node displacements and component internal forces through finite element analysis, and extracts inter-story drift angles and component strength, stiffness and stability indices.

[0116] It should be noted that there is no restriction on the execution order of D21, D22, and D23.

[0117] (I) D21, Principles of Economic Indicator Calculation

[0118] Material costs are a core component of total construction costs and account for a significant proportion of the total construction cost. Accurate calculation of material costs is not only a fundamental prerequisite for construction economic analysis, but also a direct reflection of the economic rationality of material input in construction. Therefore, this embodiment selects the material cost of building components as the core economic calculation indicator.

[0119] Based on the production characteristics of steel modular building structures, the total cost of its component materials covers three categories: structural components, enclosure structural components, and other fixed components. In this embodiment, the cost of all components is uniformly calculated based on volume (unit: yuan / m³), and its expression is as follows:

[0120] (1),

[0121] In formula (1): This refers to the total material cost of the steel modular building structure components; The total cost of structural components (columns, beams, and longitudinal beams); The total cost of the building envelope components (walls, doors, windows, roof, floor slab); This is the total cost of other fixed components (corner fittings, connectors, etc.).

[0122] (II) D22. Principles of Energy Consumption Index Calculation

[0123] The essence of building heating and cooling load is the theoretical heating and cooling capacity required to maintain a stable indoor thermal environment. Its calculation is based on the theory of unsteady thermal balance and the heat transfer theory of building envelope. It only reflects the thermal performance of the building itself and does not involve the energy efficiency of air conditioning equipment or system losses.

[0124] For any building thermal zone, its hourly heat balance equation is:

[0125] (2),

[0126] In formula (2): For a moment The building's cooling and heating loads are represented by positive values ​​for cooling loads and negative values ​​for heating loads. Total heat gain of the building; This represents the total heat loss of the building.

[0127] The total heat gain consists of solar radiation heat gain from the building envelope, internal heat gain, and air infiltration heat gain, and its expression is:

[0128] (3),

[0129] In formula (3): The amount of heat gained from solar radiation is determined by the solar radiation intensity, the solar radiation absorption coefficient of the building envelope, and the shading coefficient. For internal heat gain of personnel, lighting, equipment, etc.; The heat gain from air infiltration is calculated based on the indoor and outdoor temperature difference, the infiltration air volume, and the specific heat of the air.

[0130] The total heat loss is mainly due to heat transfer from the building envelope and heat loss through air infiltration, and its expression is as follows:

[0131] (4),

[0132] In equation (4): The calculation method for heat loss through air infiltration is exactly the same as that for heat gain through air infiltration, except that the direction of the temperature difference is reversed (or the signs of the calculation results are reversed). For the unsteady-state heat transfer of the building envelope, the Conduction Transfer Function (CTF) method is used for calculation. This method comprehensively considers the steady-state heat transfer and heat storage characteristics of the building envelope. The formula is as follows:

[0133] (5),

[0134] In equation (5): For the first Heat transfer coefficient of the building envelope; For the first Area of ​​the enclosure structure; For a moment Indoor air temperature; For a moment Outdoor air temperature; The transfer function coefficient represents the heat storage and unsteady heat transfer effects of the building envelope. For a historic moment The outdoor air temperature; This represents the total number of building envelope categories; The length of the historical period.

[0135] Under controlled indoor temperature conditions, in order to maintain the set heating temperature With cooling set temperature The hourly heating load and cooling load calculation expressions are as follows:

[0136] (6),

[0137] In formula (6): air density; The specific heat of air at constant pressure; For the volume of the hot zone; For time step; For a moment Heating load is only output when the indoor temperature is lower than the set heating temperature; For a moment Cooling load is only output when the indoor temperature is higher than the cooling set temperature.

[0138] By summing up the hourly loads over the entire year, we obtain the annual heating load, annual cooling load, and total annual heating and cooling load. The calculation expression is as follows:

[0139] (7),

[0140] (8),

[0141] (9),

[0142] In equations (7), (8), and (9): This represents the annual heating load. This represents the annual cooling load. For the first The number of days in a month; The total annual heating and cooling load of a building is a core indicator for evaluating building thermal performance and optimizing design. It indicates month, day, and hour.

[0143] During the IDF design process, data such as annual outdoor temperature, indoor temperature, and heating and cooling loads in hot zones can be output. Visualization is performed using the Matplotlib library within a Python programming environment. For example, the annual outdoor temperature of the building in the case study is shown below. Figure 8 As shown in the analysis results of a certain thermal region, the annual temperature inside the thermal region is as follows: Figure 9 As shown. The heat load results within the hot zone are as follows. Figure 10 As shown. The cooling load results within the hot zone are as follows. Figure 11 As shown.

[0144] (III) D23. Principles of Structural Index Calculation

[0145] For steel structure integrated buildings, this embodiment uses the finite element method to solve the structural response. The essence of structural static response calculation is to solve the finite element equilibrium equations. Its core logic is based on the displacement method, using nodal displacements as basic unknowns to construct the force equilibrium relationship. For linear elastic structures, the overall equilibrium equation can be expressed as:

[0146] (10),

[0147] In formula (10): This is the overall stiffness matrix of the structure; The nodal displacement vector; This represents the nodal load vector.

[0148] For the steel modular building structure in this embodiment, the base shear method in the "Code for Seismic Design of Buildings" is used to calculate the horizontal seismic action. This method is applicable to buildings with a height of [missing information]. For regular structures where shear deformation is the primary characteristic, the core principle is to convert seismic forces into equivalent horizontal loads using the seismic influence coefficient. The specific calculation process is as follows: The equivalent total gravity load of the structure is the basis for seismic force calculation, and it is obtained by weighting the representative gravity load values ​​of each floor.

[0149] (11),

[0150] In equation (11): For the first The representative value of the layer gravity load is the product of the mass of the nodes in that layer and the gravitational acceleration. , ); 0.85 is the equivalent total gravity load factor, which is a simplified factor specified in the standard. Total number of floors.

[0151] Horizontal earthquake influence coefficient Reflecting the coupling relationship between seismic intensity and structural dynamic characteristics, the calculation is performed in segments according to the period:

[0152] (12),

[0153] In equation (12): This represents the maximum value of the horizontal earthquake influence coefficient. The fundamental natural period of the structure; For the period of time; , The damping adjustment coefficient is as follows:

[0154] (13),

[0155] (14),

[0156] is the structural damping ratio.

[0157] The decay exponent is calculated using the following formula:

[0158] (15),

[0159] The formula for calculating the standard value of total horizontal seismic action is:

[0160] (16),

[0161] When the natural period of the structure When this occurs, additional seismic action at the top needs to be considered to correct for deviations in the high-rise response:

[0162] (17),

[0163] In equation (17): Add a coefficient to the top ( , ).

[0164] The standard value of horizontal seismic force for each floor is weighted by floor weight and height:

[0165] (18),

[0166] (19),

[0167] In equations (18) and (19): For the first The floor level is the height from the ground. The calculated floor loads are distributed to specific nodes, ultimately forming the load vector in the equilibrium equations. This completes the coupling between seismic load and finite element solution.

[0168] When designing modular integrated buildings, the safety and reliability of the structure must be considered. Because the structural load-bearing frame of a modular building mainly consists of modular beams, modular columns, and corner fittings connecting them, it is necessary to verify the performance indicators of bending members (beams) and bending-compression members (columns). According to the "Steel Structure Design Standard," the verification of components mainly considers three important indicators: strength, stiffness, and stability (overall stability and local stability). Whether the entire structure can meet the requirements under load is determined by the inter-story drift angle. Structural performance indicators include... Figure 12 As shown.

[0169] Steps D21, D22, and D23 use the integrated digital foundation for steel modular building structures established in step D1 as a unified data source. They focus on the automated calculation of three core professional indicators: building economics, energy consumption, and structure, and conduct systematic research. This opens up the channel between the digital foundation data and the calculation of various professional indicators, effectively solving the industry pain points of cumbersome traditional indicator calculation processes, low efficiency, disconnect between professional data and large amounts of manual intervention. This provides solid technical support and data guarantee for the digital analysis, performance evaluation and subsequent optimization design of steel modular building structures.

[0170] III. Step D3: Constructing a Graph Neural Network Agent Model

[0171] Graph structures are the core data carriers of graph neural network models, and their mathematical definition is a triple. ,in Represents a set of nodes. Denotes the set of edges. Represents the node feature matrix ( For the number of nodes, , where R represents the set of real numbers (where R is the node feature dimension). Based on the differences in the types of nodes and edges, graph structures can be divided into two main categories: isomorphic graphs and heteromorphic graphs. The core difference between the two lies in whether there is a type distinction between nodes and edges.

[0172] A homogeneous graph is a graph structure containing only one type of node and one type of edge. All nodes have a unified feature dimension and physical meaning, and all edges have the same association logic. It is suitable for scenarios where node attributes are uniform and the association relationships are simple. The adjacency matrix of an isogeneous graph... Satisfying symmetry, that is ,in Represents a node With nodes There is a connection. This indicates no connection relationship. To preserve the characteristics of each node, self-loops are usually added to the adjacency matrix, resulting in an adjacency matrix with self-loops. ( (as the identity matrix), and through the degree matrix (used for normalizing the adjacency matrix) To avoid excessively large feature values, normalization is performed to prevent numerical deviations during feature propagation.

[0173] A heterogeneous graph is a graph structure containing multiple node types and edge types. Different node types have different feature dimensions and physical meanings, and different edge types correspond to different association logics (such as inclusion, connection, and belonging). It is suitable for scenarios with diverse node types and complex association logics. The mathematical representation of a heterogeneous graph needs to be extended to quadruples. ,in These represent the set of node types and the set of edge types, respectively. Unlike isomorphic graphs, heterogeneous graphs require independent feature processing and propagation rules for each type of node and edge. By distinguishing different types of neighborhood information, more accurate feature extraction can be achieved.

[0174] Graph Neural Networks (GNNs) are a class of deep learning models specifically designed for processing graph-structured data. Their core idea stems from the fusion of traditional graph theory and neural networks. They aim to learn the feature representations of nodes and the global structure in a graph by simulating information interaction and propagation between nodes, thus solving the learning challenges of non-Euclidean space data. GNNs are widely used in fields such as social networks, recommender systems, bioinformatics, and architectural engineering.

[0175] The core framework of GNN can be summarized as a neighborhood aggregation-feature propagation loop. Its basic process is as follows: In the initial state, each node only contains its own original features; through multi-layer aggregation operations, each node merges the features of its neighboring nodes to achieve iterative feature updates; after multiple rounds of propagation, the features of each node not only contain its own attributes, but also integrate global graph structure information, which can ultimately be used for downstream classification, regression and other tasks at the node level, edge level or graph level.

[0176] Graph Convolutional Networks (GCNs) are the most classic and widely used specific models in GNNs. Their core idea is to use neighborhood aggregation and feature propagation to allow each node to integrate its own features with the features of neighboring nodes, gradually generating an embedding vector that can represent the global graph structure.

[0177] The basic form of the inter-layer feature update formula for GCN is:

[0178] (20),

[0179] In equation (20): For the first The node feature matrix of the layer, the initial layer (Original node features); For the first The learnable weight matrix of the layer is used to map the feature dimensions of the nodes; Non-linear activation functions (such as ReLU) are used to introduce non-linear features and improve the expressive power of the model.

[0180] GCN uses multi-layer convolutional operations to gradually mine the first-order, second-order, and even higher-order neighborhood information of nodes, ultimately realizing the transformation of node features into global graph features, and adapting to various downstream tasks such as classification and regression.

[0181] To address the two core requirements of structural performance evaluation and energy consumption calculation in building design optimization, this embodiment uses Graph Convolutional Networks (GCN) to construct a structural analysis proxy model, which is suitable for feature extraction and binary classification tasks of homogeneous graph data; and uses Heterogeneous Graph Convolutional Networks (HGCN) to construct an energy consumption analysis proxy model, which is suitable for regression prediction tasks with multiple types of nodes and complex relationships. This ensures that the two types of proxy models can accurately and efficiently replace traditional numerical simulations and improve optimization efficiency.

[0182] The essence of a building structure is a whole composed of various components connected by specific relationships; therefore, a structural system can be abstracted as an isomorphic diagram. ,in For a set of nodes, The specific construction rules for the edge set are as follows:

[0183] ① Node definition: A node is a single component in a building structure. Each node corresponds to a structural member. A single node is denoted as _____. ( , (Total number of components). The core function of a node is to carry the component's own attribute information, providing a foundation for subsequent feature extraction.

[0184] ② Node Feature Construction: This embodiment constructs node features based on three aspects: basic component attributes, spatial pose, and cross-sectional dimensions. Basic attributes mainly include component type, cross-sectional type, material density, and tensile length; spatial pose mainly includes the three-dimensional coordinates (3D) of the starting point and the Euler angles (3D) of the spatial direction; cross-sectional dimensions mainly include five-dimensional cross-sectional dimension parameters, including cross-sectional length, width, wall thickness, and two-dimensional custom parameters to adapt to diverse cross-sectional shapes. To eliminate the influence of dimensional differences on model training, all features are normalized.

[0185] ③ Edge Relationship Construction: An edge represents a physical connection between two structural components. Edge construction uses the AABB algorithm, which determines the connection state between components by checking if their bounding boxes intersect, thus establishing a bidirectional connection. The specific steps are as follows: For each structural component... Construct an AABB bounding box whose boundary is determined by the extreme values ​​of the component's three-dimensional coordinates, denoted as . (min represents the minimum, max represents the maximum); for any two components and ( If the formula is satisfied:

[0186] (twenty one),

[0187] Then determine the component and A connection exists between the two; establish a bidirectional edge between them. If an edge is created, its weight is set to 1; otherwise, no edge is created, and its weight is set to 0. The final adjacency matrix is ​​denoted as... ,in Representation of components and Connected, This indicates that they are not connected.

[0188] To construct the graph data for the steel modular building structure, the building classification data of the integrated digital foundation is expanded. First, the classification module information of the standard floors needs to be expanded. For a certain type of module, the module's location information and internal component information are calculated based on the transformation matrix of other modules. Second, the information for each floor of the standard floors is calculated to obtain a full model JSON file. Graph nodes are then built based on the full component JSON file. Using the Ifcopenshell library, the overall building JSON information is mapped to the IFC model and the component connection relationships are determined. The connection relationships are written into the full component JSON, and the graph node and edge relationships are built based on the full component JSON file.

[0189] ④ Prediction task setting: The prediction task of the structural analysis proxy model is set as binary prediction of structural performance. The core objective is to determine whether the key performance data of the building structure exceeds the preset limit, so as to quickly determine whether the structural performance index is feasible or not.

[0190] This embodiment sets the prediction categories to 10 dimensions, namely: whether the inter-story drift angle in the X-direction of the building exceeds the limit; whether the inter-story drift angle in the Y-direction of the building exceeds the limit; whether the bending strength of the beam exceeds the limit; whether the shear strength of the beam exceeds the limit; whether the deflection of the beam exceeds the limit; whether the overall stability of the beam exceeds the limit; whether the strength of the column exceeds the limit; whether the slenderness ratio of the column exceeds the limit; whether the in-plane stability of the column exceeds the limit; and whether the out-of-plane stability of the column exceeds the limit. Let the 10-dimensional structural performance data be... Each dimension corresponds to a different structural performance index; the preset limit vector is... , For the first Permissible limits for dimensional performance indicators ( The binary classification prediction rule is as follows: for each performance index... ,like If the performance of that dimension is satisfactory, the label 0 is output; otherwise... If the performance in that dimension is deemed unsatisfactory, label 1 is output. The final model output is a 10-dimensional binary classification label vector. ,in It is used to characterize the qualified status of each structural performance index, providing a direct basis for structural design optimization.

[0191] During model training, binary cross-entropy (BCE) is used as the basic loss function:

[0192] (twenty two),

[0193] In equation (22): For the first The first sample The true label of performance indicators This represents the corresponding predicted probability.

[0194] The construction of heterogeneous graphs for building energy consumption systems follows the principle of "multi-type nodes + targeted edge relationships." This embodiment abstracts the energy consumption system as a heterogeneous graph. ,in For a set of heterogeneous nodes, For a set of heterogeneous edges, A collection of node types This is a collection of edge types, and the specific construction rules are as follows:

[0195] ① Node Definition: Based on the constituent units of the building energy consumption system, five types of heterogeneous nodes are classified, namely, hot zone nodes (… ), spatial nodes ( ), Enclosing structure nodes ( ), ground nodes ( ), outdoor nodes ( ), covering the core elements of energy consumption analysis.

[0196] ② Node Feature Construction: Since this embodiment predicts the building's heating and cooling loads and uses ideal cooling and heating equipment, energy consumption is only related to the building's performance itself. The features of the hot zone are defined as: standard floor number, coordinate position (3D), and volume; the features of the spatial nodes are defined as: floor number, coordinate position (3D), and volume; the features of the building envelope nodes are defined as: coordinate position (3D), surface area, thermal conductivity, density, specific heat capacity, thickness, solar radiation absorptivity, infrared emissivity, visible light transmittance (for windows), and boundary type; the features of the ground are defined as soil thermal conductivity, soil density, soil specific heat capacity, solar radiation absorptivity, and infrared emissivity; and the features of the outdoors are defined as air thermal conductivity, air density, air specific heat capacity, solar radiation absorptivity, and infrared emissivity.

[0197] ③ Edge Relationship Construction: The edge relationships of heterogeneous graphs are constructed based on the actual physical connections and spatial logic between various types of nodes. Combining the five types of nodes defined above—hot zone, space, enclosure structure, outdoor, and ground—with the calculated connection relationships of space, hot zone, enclosure structure, outdoor, and ground, a bidirectional edge relationship definition is adopted to fully characterize the interaction mechanism between nodes. The core of the edge relationship originates from the essential connection between nodes, specifically: hot zone contains space relationship, space contains enclosure structure relationship, enclosure structure connects to outdoor relationship, enclosure structure connects to ground relationship, enclosure structure connects to enclosure structure relationship, enclosure structure connects to hot zone relationship, and the reverse relationship of the above relationships, totaling 12 connection relationships.

[0198] The mathematical expression for edge relations can be uniformly described as follows: , Represents the set of edges in a heterogeneous graph. Indicates the first A set of edge relationships; each type of edge relationship is stored in the form of a two-dimensional tensor, i.e. ,in For the source node index set, The target node index set ensures that the model can accurately locate the relationships between nodes through edge indexes.

[0199] ④ Prediction Task Setting: The prediction task of the energy consumption analysis proxy model is set as the regression prediction of the total building heating and cooling load. The core objective is to accurately quantify the total building heating and cooling load over a certain period of time, providing a quantitative basis for building energy consumption optimization. The model output is one-dimensional continuous energy consumption data, i.e., the total building heating and cooling load. It covers the sum of the building's winter heating load and summer cooling load.

[0200] During model training, mean squared error (MSE) is used as the loss function to measure the deviation between the model's predicted values ​​and the actual simulated values. The expression for the loss function is as follows:

[0201] (twenty three),

[0202] In equation (23): The number of training samples. For the first The model predicts the load value for each sample. For the first The actual simulated load values ​​of each sample. By minimizing this loss function, the HGCN model can fully learn the impact of various node characteristics and relationships on energy consumption, achieving accurate quantitative prediction of building heating and cooling loads, replacing the traditional and cumbersome energy consumption simulation process.

[0203] In practical applications, the Graph Convolutional Network (GCN) structural proxy model and the Heterogeneous Graph Convolutional Network (HGCN) energy consumption proxy model are used to replace finite element analysis and energy consumption simulation, enabling rapid prediction of new combinations of design variables and outputting the judgment results of whether the structural performance exceeds the limit and the predicted values ​​of building heating and cooling loads.

[0204] Taking a four-story school dormitory building as an example, this building adopts a steel modular construction structure. The building comprises two standard floors and is located in a certain city. According to meteorological data, the coldest month in this city is January, with an average temperature of 7.8℃. There are approximately 10 days with an average daily temperature ≤5℃, classifying it as a temperate region. The main structure of this steel modular building is designed for a service life of 50 years. The building's safety level is Level II, its seismic fortification category is Class C, the site category is Class II, the seismic fortification intensity is 8 degrees, and the ground roughness is Class B. This steel modular building structure has two standard floors: one floor and floors 2-4.

[0205] Information was extracted and preprocessed from the architectural drawings of each standard floor. After equivalent feature matching and classification of the obtained module information, a total of 22 module categories were obtained. Information on additional components such as columns, beams, roofs, and floor slabs was defined through manual interaction to fill in the information of the integrated digital foundation. The final result is a 4314KB JSON file, representing the IFC model of the module within each standard floor.

[0206] Compared to traditional manual BIM modeling, structural modeling, and energy consumption modeling based on architectural drawings, this integrated digital infrastructure integrates and expresses data from various disciplines, and categorizes and integrates information to suit the standardized production characteristics of modular buildings, resulting in a significant advantage in file size. As shown in Table 3 below, the total size of the architectural IFC model, the HDF5 model file of the structure in OpenSees, and the IDF file of energy consumption is 32707KB, while the size of the integrated digital infrastructure proposed in this embodiment is only 4314KB, saving 86.81% of storage space.

[0207] Table 3 Manual Modeling and Digital Foundation File Sizes

[0208]

[0209] The integrated digital foundation is a digital representation of steel modular building structures. Regarding the accuracy verification of the digital foundation, this embodiment systematically compares the building material costs, finite element model analysis results, and energy consumption model analysis results obtained based on the integrated digital foundation with the calculation results of traditional manual modeling to verify the accuracy and reliability of the proposed method. The root mean square error of variation (CV(RMSE)) is used as the core indicator in this accuracy evaluation.

[0210] The material cost, energy consumption, and maximum inter-story drift angle of the structural analysis are compared, as shown in Table 4 below.

[0211] Table 4. Results of manual and automatic calculations

[0212]

[0213] As shown in Table 4, the CV (RMSE) values ​​for material cost, finite element model analysis results, and energy consumption model analysis results are all 0. This result indicates that the proposed method for calculating indicators based on an integrated digital foundation is consistent with the results of manual modeling calculations. Therefore, it can be concluded that the integrated digital foundation framework proposed in this embodiment can meet the professional accuracy requirements achievable through manual model building and can serve as an effective alternative to manual modeling calculations.

[0214] Regarding the overall efficiency of the integrated digital infrastructure, the total time to complete the entire process is only 623 seconds. In contrast, the traditional manual method requires engineers to separately build a building BIM model, a structural OpenSees model, and write an energy consumption IDF file, all of which involve manual data interaction and setup, accumulating approximately 10 hours of time. By comparison, the integrated digital infrastructure improves the overall efficiency of professional modeling and performance analysis by approximately 60 times, achieving a reduction from hours to minutes.

[0215] For structural compliance classification prediction, in the initial training based on LHS (Latin Hypercube Sampling) samples, the confusion matrix of the GCN model is as follows: Figure 13 As shown. From Figure 13 As can be seen, the GCN model has an accuracy of up to 96.4%.

[0216] For building heating and cooling load regression prediction, the fitting curves between the predicted and actual values ​​of the HGCN surrogate model on the test set are as follows: Figure 14As shown. From Figure 14 It can be seen that the GCN model also has high prediction accuracy.

[0217] Based on the above method, this embodiment also provides a steel modular building structure performance prediction system based on graph neural networks. The key features are: an integrated digital foundation construction module, a multi-professional index automatic calculation module, and a graph neural network proxy model construction module, which are respectively used to execute steps D1, D2, and D3 in the steel modular building structure performance prediction method based on graph neural networks.

[0218] Furthermore, this embodiment also provides a computer-readable storage medium storing a computer program thereon, the key point being that when the program is executed by a processor, it implements the aforementioned method for predicting the performance of steel modular building structures based on graph neural networks or the aforementioned system for predicting the performance of steel modular building structures based on graph neural networks.

[0219] It should be noted that, through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to implement the methods or systems described in the various examples or some parts of the examples.

[0220] In summary, the steel modular building structure performance prediction method and system based on graph neural networks provided in this invention uses DXF drawings as the sole input. Through drawing information extraction, multi-disciplinary information processing and fusion, it constructs an integrated digital foundation JSON data covering three major disciplines: architecture, structure, and energy consumption, achieving homogeneous integration and standardized storage of multi-disciplinary information. Based on this unified data source, it automatically completes the hierarchical cumulative accounting of building material costs, the automatic construction of structural finite element models and extraction of mechanical performance indicators, and the automatic mapping of energy consumption models and simulation of annual heating and cooling loads, achieving fully automated and accurate calculation from data to underlying indicators. Furthermore, based on the integrated digital foundation, this invention constructs homogeneous and heterogeneous graph data respectively, trains a graph convolutional network (GCN) structural proxy model and a heterogeneous graph convolutional network (HGCN) energy consumption proxy model, and uses graph neural networks to accurately capture the graph topological features of the structural system and the heterogeneous graph structural relationships of the energy consumption system, enabling rapid determination of whether structural performance exceeds limits and accurate prediction of building heating and cooling loads, replacing traditional time-consuming finite element analysis and energy consumption simulation. This invention effectively solves the technical problems of fragmented multi-disciplinary data, reliance on manual calculation of indicators, and excessive time consumption in numerical simulation that makes it difficult to support rapid iteration in the existing technology. It provides accurate, unified, and scalable technical support and theoretical basis for the collaborative design and efficient optimization of steel modular building structures.

[0221] The above are merely preferred examples of this application and are not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application shall be included within the scope of protection of this application.

Claims

1. A method for predicting the structural performance of steel modular buildings based on graph neural networks, characterized in that, Including the following steps: D1. Extract geometric data from DXF drawings, fill in the code and information fusion of the three disciplines of architecture, structure and energy consumption, and construct integrated digital basic JSON data; D2. Based on the integrated digital foundation JSON data, automatically calculate the cost of building materials, automatically build a structural finite element analysis model and extract structural performance indicators, and automatically build an energy consumption analysis model and calculate the building's heating and cooling loads. D3. Based on the integrated digital foundation JSON data, construct homogeneous graph data and heterogeneous graph data. Using the structural performance index and building heating and cooling load calculated in step D2 as supervision signals, construct and train a structural proxy model based on graph convolutional network and an energy consumption proxy model based on heterogeneous graph convolutional network, respectively. The structural proxy model is used to predict whether the structural performance exceeds the limit, and the energy consumption proxy model is used to predict the building heating and cooling load.

2. The method for predicting the structural performance of steel modular buildings based on graph neural networks according to claim 1, characterized in that, The method for constructing isomorphic graph data in step D3 includes: Each individual component in a building structure is considered a node, and the node characteristics include the component's basic properties, spatial orientation, and cross-sectional dimensions. The Axis Aligned Bounding Box (AABB) algorithm is used to determine the connection relationship between components. If the projections of the bounding boxes of two components on the X, Y, and Z coordinate axes overlap, a connection relationship is determined to exist, and a bidirectional edge is established. The building structure is abstracted as an isomorphic graph. The prediction task of the graph convolutional network (GCN) structural surrogate model is a binary classification prediction of structural performance. It outputs a 10-dimensional binary label vector, which corresponds to whether the inter-story drift angle in the X direction of the building exceeds the limit, whether the inter-story drift angle in the Y direction of the building exceeds the limit, whether the bending strength of the beam exceeds the limit, whether the shear strength of the beam exceeds the limit, whether the deflection of the beam exceeds the limit, whether the overall stability of the beam exceeds the limit, whether the strength of the column exceeds the limit, whether the slenderness ratio of the column exceeds the limit, whether the in-plane stability of the column exceeds the limit, and whether the out-of-plane stability of the column exceeds the limit.

3. The method for predicting the structural performance of steel modular buildings based on graph neural networks according to claim 2, characterized in that, The method for constructing heterogeneous graph data in step D3 includes: The nodes are classified into five categories: hot zone nodes, spatial nodes, building envelope nodes, ground nodes, and outdoor nodes. Twelve edge relationships are constructed, including hot zone containing space relationship, space containing enclosure structure relationship, enclosure structure connecting to the outside relationship, enclosure structure connecting to the ground relationship, enclosure structure connecting to enclosure structure relationship, enclosure structure connecting to hot zone relationship and its reverse relationship. The building energy consumption system is abstracted as a heterogeneous graph, and the prediction task of the Heterogeneous Graph Convolutional Network (HGCN) energy consumption proxy model is to regress and predict the total building cooling and heating load.

4. The method for predicting the structural performance of steel modular buildings based on graph neural networks according to any one of claims 1 to 3, characterized in that, Step D1 specifically includes the following steps: D11. Input DXF drawing, extract geometric data of grid, wall, door and window, module frame and hot zone frame according to line type layer based on Dxfgrabber library, calculate center line coordinates and endpoints, and output initial component data in JSON format; D12: The architecture specialty performs 19-bit coding, parameter filling, collision correction, and component classification; the structural specialty calculates room loads and fills in the mechanical materials of components; the energy consumption specialty performs 23-bit coding, space identification, and enclosure structure filling, and outputs JSON data for each specialty. D13. Based on the five principles of cross-disciplinary integration, multi-level, classification, master-slave dependency, and coding collaboration, a tree-structure integration method is adopted to integrate the JSON data of the three disciplines according to the hierarchy and output integrated digital basic JSON data.

5. The method for predicting the structural performance of steel modular buildings based on graph neural networks according to claim 4, characterized in that, In step D11, extracting the geometric data of the wall and calculating the centerline coordinates and endpoints includes: Wall line matching steps: Calculate the wall thickness; match the wall lines based on the wall thickness; Centerline calculation steps: Calculate the centerline based on the paired two wall lines, and correct the endpoints of the centerline based on the grid and the half-wall thickness spacing between the inner walls. The correction process is as follows: Based on grid correction: Define axis coordinates. If the starting point of the centerline is less than the axis coordinate and the absolute value of the difference between the starting point and the axis coordinate is equal to half the wall thickness, then the starting point is corrected to the axis coordinate. Based on internal wall correction: Extract the endpoint set of all original wall lines, calculate the Euclidean distance between any pair of endpoints, determine the endpoint pairs with a distance equal to half the wall thickness as valid endpoint pairs, perform deduplication on the valid endpoint pairs, generate new wall lines based on the deduplicated valid endpoint pairs, and merge the new wall lines with the original wall lines.

6. The method for predicting the structural performance of steel modular buildings based on graph neural networks according to claim 4, characterized in that, In step D12: The collision corrections performed by the architectural professionals include: An IFC model is built based on the initial JSON data before correction. The triangular facet information of the IFC components is extracted using the Geom function of the IfcOpenShell library, and an axis-aligned bounding box is built for each component. A hierarchical bounding box tree is constructed using a recursive method, and the axis-aligned bounding boxes of all components of the module are organized into a tree structure according to the parent-child hierarchy; Collision detection is performed by determining whether the projections of two components on the X, Y, and Z coordinate axes overlap. If they overlap, a collision is determined, and the spatial geometry of the collision area is calculated. The components are adaptively adjusted based on the collision detection results, including first-stage adjustment and second-stage adjustment. The first stage of adjustment is as follows: the component adjustment priority is defined from high to low as the module connector, frame component, and wall panel component. When a collision occurs between components of different priorities, the components are adjusted one by one in order of priority from low to high. Boolean difference operation is performed between the solid model of the low priority component and the collision area. The second stage of adjustment is as follows: based on the wall after the first stage adjustment, collision calculation is performed between the wall and the original door and window IFC component entities. The adjustment priority is defined from high to low as door and window components and wall panel components. Door and window opening information is added to the wall according to the collision information between the door and window and the wall panel. The component classification performed by architectural professionals includes: Extract the 3D vertex coordinates of the component using the Geom function from the IfcOpenShell library; Traverse the equivalent geometric data of the component's mirror image and rotation along the X, Y, and Z axes, and calculate the hash value for each type of geometric data using the SHA256 algorithm; Take the minimum hash value among itself and all its mirror combinations as the geometric fingerprint of the component, and classify components with the same hash value as the same type; The structural engineering calculations of the room loads include: The room outline is identified by a counterclockwise room outline detection method. Text and text coordinates are identified from the axis grid area. Rays are emitted to the left based on the coordinates. The first line of intersection is taken as the initial outline. The next outline is found counterclockwise from the end point to the starting point. The process is repeated until the initial outline is found, and the room outline is determined. Traverse the rooms and find the modules contained in each room. For a room containing multiple modules, traverse the modules to determine the modules contained in the room, thereby determining the room type and area range contained in each module. Search the functional room constant and live load coefficient table to find the constant and live load parameters of the corresponding room type, and calculate the enclosed area of ​​each room based on the Polygon.area function of the Shapely library; Based on the average dead and live coefficient of the floor load in the weighted calculation module of room area and room dead and live coefficient, the edge load is calculated by the trapezoidal triangle load distribution method. The spatial identification conducted by energy consumption professionals includes: Define the module as the minimum space; Whether a module is an independent space is determined by whether there are walls in the four directions of east, south, west, and north. If there are walls in all four directions, then it is an independent space. If it is an independent space, then the module is used as the space; for modules with fewer than 4 wall orientations, query the room to which the module belongs and use that room as the space. The building envelope filling performed by the energy management professionals includes: The building envelope is made of the same material as the overall building, which includes five categories: roof, ground, walls, doors, and windows. Based on the names of the building envelope materials corresponding to the modules in the architectural profession, the material and thermal performance are assigned. The 19-digit code for the architectural profession adopts a hierarchical segmented interpretation rule, defining the building type identifier, standard floor number, floor number, professional type, module type, module number, component type, geometric type, spatial orientation, and component number for each digit. The 23-bit code for energy consumption uses a hierarchical segmented interpretation rule, defining the building type identifier, standard floor number, floor number, professional type, thermal zone type, thermal zone number, space type, space number, building envelope type, building envelope geometry type, building envelope orientation, and building envelope number bit by bit.

7. The method for predicting the structural performance of steel modular buildings based on graph neural networks according to claim 4, characterized in that, In step D13, the implementation method of the five principles includes: Cross-disciplinary integration: The mechanical performance information of structural components is integrated one-to-one with the building components in the architectural field that have completely overlapping geometric shapes and spatial positions; the thermal zone spatial division and thermal performance information of the building envelope are integrated in parallel with the data of the architectural and structural fields. Multi-level integration: For architectural and structural engineering, integration is carried out in three levels: standard layer, module, and component; for energy consumption, integration is carried out in three levels: thermal zone, space, and building envelope. Category-based fusion: Standard floor data is fused to achieve floor classification; for architecture or structural engineering, modules are classified according to the geometric distribution information of internal components, and components are classified according to their functional categories and geometric information; for energy consumption engineering, thermal zones are classified according to the spatial distribution of internal thermal zones and the information of the building envelope, spaces are classified according to the distribution of components within each space, and building envelopes are classified according to their types and geometric information. Master-slave dependency integration: The overall building dimensions, module layout, and room functions are defined as the master model data; the geometric section parameters of components and the energy consumption model and building envelope materials are defined as slave model data, so that the slave model data is stored and managed in dependence on the corresponding master model data. Coding Collaboration and Integration: A unified coding system is established, with building type identification, standard floor, and floor information as common components, professional categories, modules, hot zones, and spatial information as meso-level coding, and component and enclosure structure information as micro-level coding.

8. The method for predicting the structural performance of steel modular buildings based on graph neural networks according to any one of claims 1 to 3, characterized in that, Step D2 specifically includes: D21. Based on the module classification characteristics of integrated digital foundation JSON data, a step-by-step accumulation method from module level, standard level to building level is adopted to multiply the volume and unit price of structural components, enclosure structural components and other fixed components and sum them to calculate the total cost of building materials. D22. Extract the spatial geometry information of the thermal zone and the thermal parameters of the building envelope from the integrated digital foundation JSON data, generate IDF files through Eppy automatic mapping, call EnergyPlus to perform building heat balance simulation, and calculate the sum of the annual cumulative heating load and cooling load. D23. Extract the geometric information and mechanical parameters of structural components from the integrated digital foundation JSON data, automatically generate node elements, apply boundary conditions and seismic loads in OpenSees, solve the node displacements and component internal forces through finite element analysis, and extract the inter-story drift angles and component strength, stiffness and stability indices.

9. The method for predicting the structural performance of steel modular buildings based on graph neural networks according to claim 8, characterized in that, Step D22 specifically includes: Extract the spatial geometry information of the thermal zone and the thermal parameters of the building envelope from the integrated digital foundation JSON data. The geometric information includes the vertex coordinates of the thermal zone, space, and building envelope, and the thermal parameters include the material name and thermal performance of each building envelope. Based on three-dimensional spatial topology analysis, the external boundary conditions of each enclosure structure are automatically calculated to determine the attributes of cross-thermal zone connection, outdoor connection or ground connection. The thermal zone spatial geometry information is mapped to the thermal zone, spatial and enclosure structure geometry models in the IDF file using the Eppy library, and the thermal parameters of the enclosure structure are mapped to the corresponding material properties. EnergyPlus was used to perform building thermal balance simulation on the IDF file, and the hourly cooling load and hourly heating load of each thermal zone were calculated using the conduction transfer function method. The hourly cooling load is added up over the entire year to obtain the total annual cooling load, and the hourly heating load is added up over the entire year to obtain the total annual heating load. The two are then summed to obtain the total annual cooling and heating load.

10. A performance prediction system for steel modular building structures based on graph neural networks, characterized in that, It includes an integrated digital infrastructure construction module, a multi-professional indicator automatic calculation module, and a graph neural network proxy model construction module, which are respectively used to execute steps D1, D2, and D3 in the graph neural network-based steel modular building structure performance prediction method according to any one of claims 1 to 9.

Images