BIM (Building Information Modeling) forward design method based on full-process AI (Artificial Intelligence) driving

Abstract

The invention relates to the technical field of building construction, in particular to a BIM (Building Information Modeling) forward design method based on full-process AI (Artificial Intelligence) driving, which comprises the following steps of: digitizing a demand, and extracting a quantitative design index through a multi-modal demand analysis module; parameter generation: generating and evaluating a first-order BIM seed model by using an AI parameter generation engine driven by a generative adversarial network; iterative optimization is carried out, and automatic form, performance and layout optimization is carried out on the model through topological optimization, energy consumption prediction and reinforcement learning; performing collaborative verification, and generating collaborative correction suggestions by adopting a conflict detection algorithm fusing a boundary box and a GJK algorithm; and result delivery: performing simulation verification through a digital twin engine and outputting a final result. According to the method, intelligence and automation of the whole process from demand to delivery are achieved, the problems that a traditional method depends on manpower, collaboration lags behind, performance analysis and design are disjointed and the like are solved, and the design efficiency, quality and scientificity are remarkably improved.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of building construction, in particular to a BIM forward design method based on full-process AI driving. BACKGROUND

[0002] As the core of the digital transformation of the construction industry, the building information model (BIM) technology has been widely used in the design, construction and operation stages of the project. BIM forward design refers to the design method that uses BIM models for three-dimensional design from the scheme design stage, and directly derives construction drawings, statistical reports and other results. Compared with the traditional "two-dimensional first and three-dimensional later" model, it has significant advantages in ensuring design quality and improving the efficiency of professional collaboration.

[0003] However, the current BIM forward design process still has several bottleneck problems that need to be solved, mainly in the following aspects:

[0004] 1. Low degree of design early decision-making intelligence, excessive reliance on manual experience. In the scheme generation stage of traditional BIM forward design, architects still need to manually convert the text requirements such as the design task book of the owner into specific architectural form parameters (such as volume rate, functional partition, spatial scale, etc.). This process highly depends on the personal experience and subjective judgment of the designer, lacks scientific and quantitative decision support, resulting in low efficiency of scheme generation, and it is difficult to comprehensively and accurately consider multi-dimensional constraints such as building codes and regional climate. The scientificity and rationality of the scheme are inherently deficient.

[0005] 2. Performance analysis and design process are disconnected, and optimization feedback is not timely. In the current process, the performance analysis of the scheme such as structure, energy consumption and lighting is usually used as a verification link after the design, rather than an optimization guide integrated into the design process. The analysis results are usually presented in the form of a report, and the designer needs to manually understand the report and modify the model again. The whole process is long and slow in iteration, and cannot realize real-time and online performance feedback and closed-loop optimization, making it difficult to ensure the optimal comprehensive performance of the final scheme.

[0006] 3. Limited automation, unable to form a full-process closed loop. Although some generative design or parametric design tools have appeared on the market to assist in scheme generation, these tools often only focus on the scheme concept stage, and the generated model has low precision (usually only LOD200), which is disconnected from the subsequent detailed design and construction drawing output. From the concept model to the high-precision BIM model (LOD300 and above) that can be used for construction and operation, a large amount of manual intervention and modeling work is still required, and full-process AI driving and automation from demand input to final results cannot be achieved.

[0007] Therefore, there is an urgent need in the art for a new BIM forward design method that can overcome the above-mentioned defects, realize the intelligent and automated driving of the whole process from demand to result, and fundamentally improve the design efficiency and quality. SUMMARY

[0008] In view of the deficiencies of the prior art, the purpose of the present application is to provide a BIM forward design method based on whole-process AI driving, which can realize the intelligent and automated driving of the whole process from demand to result and fundamentally improve the design efficiency and quality.

[0009] The technical solution adopted by the present application to solve its technical problems is:

[0010] A BIM forward design method based on whole-process AI driving, comprising the following steps:

[0011] Step 1, digitalization of demand: inputting design task book and field survey data into a multi-modal demand analysis module; the multi-modal demand analysis module analyzes the design task book through a natural language processing (NLP) unit, extracts building function, construction scale, and performance requirement key information, simultaneously automatically associates building specifications and regional climate two types of constraint conditions through a knowledge graph, and finally outputs three types of quantitative design indexes of plot ratio, functional zoning, and space demand;

[0012] Step 2, parameter generation: inputting the quantitative design indexes into an AI parameter generation engine; the AI parameter generation engine constructs a parameter prediction model based on a generative adversarial network, and the parameter prediction model contains two sub-networks of a generator and a discriminator;

[0013] The generator generates building shape basic parameters according to the quantitative design indexes, and matches two types of configuration schemes of structural system and mechanical and electrical system according to the building shape basic parameters, and combines and outputs the building shape basic parameters and the optimal configuration scheme to form a complete scheme;

[0014] The discriminator adopts a multi-layer perception machine model, takes historical engineering cases as benchmark data, and quantitatively scores by calculating the Euclidean distance value of the complete scheme and the benchmark data; when the quantitative score is less than a preset pass line, it is fed back to the generator, the generator adjusts the conversion coefficient of the parameter mapping algorithm, and generates a complete scheme again; when the quantitative score is greater than or equal to the preset pass line, the generator and the discriminator perform adversarial training based on the complete scheme, optimize the weight coefficient of the multi-layer perception machine model, and output the first-order BIM seed model meeting the preset requirements when the iteration termination condition is met;

[0015] Step 3, iterative optimization: importing the first-order BIM seed model into an intelligent optimization decision module, and the intelligent optimization decision module performs the following optimization,

[0016] S301: Adjust the shape of the primary BIM seed model through a topology optimization algorithm, adjust the shape of the components, and output a first optimization intermediate model;

[0017] S302: Identify the energy consumption exceeding area coordinates in the first optimization intermediate model through an energy consumption prediction model and give energy-saving optimization suggestions; after adjusting the first optimization intermediate model according to the energy-saving optimization suggestions, output a second optimization intermediate model after energy consumption optimization;

[0018] S303: Call the reinforcement learning agent, take the second optimization intermediate model as the initial state, and through iteration, select the optimal space layout and pipeline synthesis scheme, and output the optimized BIM seed model;

[0019] Step 4, collaborative verification: import the optimized BIM seed model into the multi-specialty collaborative platform; the multi-specialty collaborative platform uses a conflict detection algorithm combining bounding box collision detection and GJK algorithm, first filters possible component pairs through bounding box collision detection, then filters out component pairs that will collide through GJK algorithm, and generates two types of collaborative correction suggestions, pipeline detour and component position adjustment, for component pairs that will collide, adjust the optimized BIM seed model according to the collaborative correction suggestions, and output the corrected BIM seed model;

[0020] Step 5, deliver the results: import the corrected BIM seed model into the full life cycle verification system; the full life cycle verification system uses a digital twin engine for construction simulation and operation prediction, and generates final results based on simulation and prediction results, including final construction drawings and delivery BIM model.

[0021] As preferred, the further technical scheme of the present application is:

[0022] Preferably,

[0023] Preferably, in step 2, the work of the generator includes,

[0024] S201: Convert the plot ratio to the initial values of the building floor area and the number of floors through a parameter mapping algorithm, convert the functional zoning to the area plane size parameters, and convert the space requirements to the headroom height parameters;

[0025] S202: Generate building shape basic parameters such as floor height and column grid based on the initial values of building floor area and number of floors, area plane size parameters, and headroom height parameters;

[0026] S203: Call the building specifications and regional climate constraints associated with step 1, extract the specification constraint parameters and climate characteristic parameters; extract the structure-mechanical system adaptation data of similar buildings from the historical engineering case library, and establish a candidate scheme set;

[0027] S204: Calculate three indicator scores for each candidate scheme in the candidate scheme set, including specification compliance score, climate adaptability score, and case matching degree score, and calculate a comprehensive score according to the specification compliance score, the climate adaptability score, and the case matching degree score, and select the scheme with the highest comprehensive score as the optimal configuration scheme;

[0028] The work of the discriminator includes,

[0029] S205: The discriminator adopts a multi-layer perception model, takes the parameter intervals of the same type of building in the historical engineering cases and the configuration scheme as the benchmark data, combines the building shape basic parameters output by the generator with the optimal configuration scheme into a complete scheme, quantitatively scores the complete scheme by calculating the Euclidean distance value of the complete scheme and the benchmark data, and quantitatively scores the complete scheme by calculating the Euclidean distance value of the complete scheme and the benchmark data. The calculation formula of the quantitative score is,

[0030] (1) ;

[0031] Wherein, is the Euclidean distance value, is an adjustment coefficient, the value range is 0.8-1.2, and is dynamically set according to the building type, the residential building takes 0.8, the public building takes 1.0, and the industrial building takes 1.2; when the preset qualified line, the corresponding complete scheme is determined as a valid scheme;

[0032] S206: The generator and the discriminator perform adversarial training based on the valid scheme, optimize the weight coefficient of the multi-layer perception model, and output the primary BIM seed model meeting the preset requirements after meeting the iteration termination condition.

[0033] Preferably, in step S204, the calculation formula of the specification compliance score is,

[0034] (2) ;

[0035] Wherein, is the candidate scheme number;

[0036] The calculation formula of the climate adaptability score is,

[0037] (3) ;

[0038] Wherein, is the number of climate influence factors, is the weight of the th climate influence factor, is the adaptation degree score of the th scheme to the th climate influence factor, 0-100.

[0039] Case matching degree score The calculation formula is,

[0040] (4) ;

[0041] Comprehensive score The calculation formula is,

[0042] (5).

[0043] Preferably, the calculation formula of the climate adaptability score The selection of the climate impact factor in the calculation formula adopts the following method,

[0044] First, a basic factor pool is constructed, including annual average temperature, average temperature of the coldest month, average temperature of the hottest month, extreme minimum temperature, extreme maximum temperature, annual average relative humidity, annual precipitation, maximum hourly precipitation, annual average wind speed, winter dominant wind direction wind speed, summer dominant wind direction wind speed, and annual total solar radiation;

[0045] Then, according to the climate region where the project is located, the basic factor pool is reduced in dimension through principal component analysis, the correlation degree of each climate parameter with building design is calculated, and the top core evaluation factors with a variance contribution ratio of ≥85% are selected as the core evaluation factors.

[0046] Preferably, the calculation rule of the adaptability score is as follows,

[0047] For the th climate impact factor, the optimal solution , is the optimal design parameter value corresponding to the climate impact factor in the historical case, if the parameter value of the climate impact factor in the th scheme is , if ≤0, let =0.

[0048] Preferably, in step S205, the calculation formula of the Euclidean distance value is,

[0049] (6) ;

[0050] wherein, is the total number of parameters participating in the calculation, is the th parameter value of the generated scheme, is the The mean of each parameter.

[0051] Preferably, step S206 specifically includes:

[0052] S2061, Calculate the loss function: Input the quantitative score corresponding to the effective solution. loss function Defined as The reciprocal of L, that is, L = 1 / ,in, The scoring results are from 0 to 100. 0.01- ;

[0053] S2062: The generator adjusts the transformation coefficients of the parameter mapping algorithm and the weight coefficients of the structure-electromechanical system matching through the backpropagation algorithm, so that the loss function continues to decrease;

[0054] S2063: The system continuously records the quantitative scores generated in each iteration. When 50 consecutive quantitative scores are all When the score is ≥85, the generator is considered to have stably output high-quality solutions.

[0055] S2064: Quantitative scoring of ≥85 points on the 50th attempt. The corresponding complete solution is imported into the BIM modeling engine, which automatically builds a 3D model according to the specified accuracy standards, forming a BIM seed model.

[0056] Preferably, S301: The initial BIM seed model is morphologically adjusted using a topology optimization algorithm, specifically including:

[0057] a. Using the structural components in the initial BIM seed model as the optimization objects, extract the initial values ​​of the element density for each component. Unit volume and material properties, among which, Number the units;

[0058] b. Set the topology optimization objective function as follows: ,in, The penalty factor is 3-5, and N is the total number of structural units;

[0059] c. Set the constraints as follows ,and ,in, For the first The actual stress of each unit For the allowable stress of the material, For the first The actual displacement of each unit To allow displacement;

[0060] d. Adopting the moving asymptote algorithm to solve the optimization problem, the unit density is iteratively updated by the following formula,

[0061] (7) ;

[0062] wherein, is the iteration number, is the step factor (initial value 0.01, gradually reduced to 0.001 with iteration convergence), is the gradient of the objective function at the th iteration;

[0063] e. When the difference of the objective function of 10 consecutive iterations is , stop iteration, and output the once-optimized intermediate model of the structural member form adjustment .

[0064] Preferably, S302, energy consumption optimization adjustment is performed through an energy consumption prediction model, specifically including:

[0065] f. Import into the computational fluid dynamics module to simulate and obtain the air flow field velocity , temperature field distribution and heat exchange coefficient , wherein, is the spatial coordinate;

[0066] g. Based on the above data, a long short-term memory network model for energy consumption prediction is constructed, and the spatial coordinate, window-wall ratio and envelope heat transfer coefficient are input, and the regional predicted energy consumption value is output; that is, the model learns the correlation between the input parameters and energy consumption, and outputs the prediction result of the energy consumption of each region in the building;

[0067] h. Set the loss function as,

[0068] Loss= (8) ;

[0069] wherein, is the sample number, which refers to the total number of independent energy consumption regions divided in the building, and each region is 1 sample; is the actual energy consumption value of the th region; is the predicted energy consumption value of the th region;

[0070] i. Adopting the Adam optimizer to minimize the loss function, and when Loss≤0.05, it is determined that the model converges, at this time the output by the model is consistent with the actual energy consumption value The deviation meets the accuracy requirements, and then the model, based on its converged predictive ability, outputs a first-optimized intermediate model. various regions in China And filter out The system simultaneously outputs the spatial coordinates of areas where energy consumption exceeds the standard; among which... The set standard energy consumption value is used to simultaneously output energy-saving optimization suggestions.

[0071] j. Suggestions for adjusting the window-to-wall ratio:

[0072] (9);

[0073] Recommendations for adjusting the heat transfer coefficient of the building envelope:

[0074] (10);

[0075] in, The adjusted window-to-wall ratio. The adjusted heat transfer coefficient of the building envelope;

[0076] Updated based on the above optimization suggestions. Based on the window-to-wall ratio and heat transfer coefficient parameters of the building envelope, recalculate the energy consumption of each area and confirm it. Then, the intermediate model with optimized energy consumption is output. .

[0077] Preferably, S303: Layout and pipeline optimization, specifically including,

[0078] k. with Initial state The state space of a reinforcement learning agent includes spatial layout parameters and pipeline parameters; the spatial layout parameters include room dimensions, and the pipeline parameters include pipe diameter and pipeline orientation angle.

[0079] l. Set the optimization objectives as reducing pipeline crossing conflicts and improving space utilization, and establish a dual objective function.

[0080] (11);

[0081] in, Pipeline conflict index , For the number of intersecting pipeline pairs, This represents the total number of pipeline pairs; To maximize space utilization, , To make effective use of the area, Total building area;

[0082] m. Iterative optimization is performed using a greedy algorithm.

[0083] m1. Randomly generate 10 sets of adjustment schemes containing spatial layout and pipeline parameters;

[0084] m2. Calculate the objective function value F for each group of schemes;

[0085] m3. Retain the 3 schemes with the highest F values, and generate 10 new schemes based on these;

[0086] m4. Repeat steps m1-m3 until the F-value increases by ≤0.01 for 5 consecutive iterations;

[0087] Select the solution with the highest final F-value and update. Spatial layout and pipeline parameters, output optimized BIM seed model .

[0088] The present invention, which adopts the above technical solution, has the following prominent features compared with the prior art:

[0089] 1. A fully automated closed loop has been built, from requirements analysis, solution generation, performance optimization to collaborative verification, which changes the situation in the traditional process where each link relies on manual labor and is isolated from each other.

[0090] 2. The core adopts a generative adversarial network framework. Through adversarial training and self-iteration between the generator and the discriminator, it ensures that the automatically generated schemes not only meet quantitative indicators but also have engineering rationality, laying the foundation for high-quality design.

[0091] 3. The multi-dimensional performance analysis of structure, energy consumption, pipelines, etc., has been transformed from traditional post-event verification to in-process intelligent optimization, realizing real-time linkage between design, simulation and optimization, and ensuring the comprehensive performance and compliance of the solution from the source.

[0092] 4. It adopts a fusion algorithm for accurate collision detection and can automatically generate actionable collaborative modification suggestions, transforming multi-disciplinary collaboration from passive post-event conflict checking to proactive intelligent collaboration, which greatly improves design efficiency and quality. Attached Figure Description

[0093] Figure 1 This is an overall flowchart of the BIM forward design method based on full-process AI-driven implementation in this embodiment of the invention. Detailed Implementation

[0094] The present invention will be further illustrated below with reference to specific embodiments. The purpose of this illustration is solely to provide a better understanding of the invention. Therefore, the examples given do not limit the scope of protection of the present invention.

[0095] like Figure 1As shown in the figure, this embodiment presents a BIM forward design method based on full-process AI-driven approach, including the following steps:

[0096] Step 1, Digitizing Requirements: Input the design brief and site survey data into the multimodal requirements analysis module; the multimodal requirements analysis module analyzes the design brief, extracts key information on building functions, construction scale, and performance requirements, and automatically associates building codes and regional climate constraints through a knowledge graph, ultimately outputting three types of quantitative design indicators: floor area ratio, functional zoning, and spatial requirements.

[0097] The Natural Language Processing (NLP) unit is used to parse the design task statement. The NLP unit is a pre-trained model based on the Transformer architecture.

[0098] The knowledge graph includes a building code knowledge graph (such as the "Unified Standard for Design of Civil Buildings" GB50352) and a regional climate database (derived from the "Special Meteorological Dataset for Thermal Environment Analysis of Chinese Buildings"). During automatic association, semantic similarity calculations are used to automatically associate the requirements extracted by NLP units with the clauses and climate parameters in the knowledge graph, thereby transforming abstract requirements such as "comfort" and "energy saving" into specific quantitative constraints such as "window-to-wall ratio ≤ 0.4" and "heat transfer coefficient of building envelope K ≤ 0.6 W / (m²·K)".

[0099] Step 2, Parameter Generation: Input the quantitative design indicators into the AI ​​parameter generation engine; the AI ​​parameter generation engine builds a parameter prediction model based on generative adversarial networks, which includes two sub-networks: a generator and a discriminator.

[0100] The generator generates basic building form parameters based on quantitative design indicators, and matches structural system and electromechanical system configuration schemes based on these parameters. It then merges the basic building form parameters with the optimal configuration scheme to output a complete solution. The generator's specific tasks include:

[0101] S201: The plot ratio is converted into the initial values ​​of building footprint and number of floors through a parameter mapping algorithm, the functional zoning is converted into the planar dimension parameters of each area, and the space requirements are converted into the net height parameters.

[0102] S202: Generate basic building form parameters, namely floor height and column grid, based on the initial values ​​of building footprint and number of floors, plan dimensions of each area, and net height parameters.

[0103] S203: Call the building codes and regional climate constraints associated in step 1, extract the code constraint parameters and climate characteristic parameters; extract the structural-mechanical system adaptation data of similar buildings from the historical engineering case library, and establish a candidate scheme set.

[0104] S204: Calculate three index scores for each candidate scheme in the candidate scheme set, including standard compliance score, climate adaptability score, and case matching score. Calculate a comprehensive score based on the standard compliance score, climate adaptability score, and case matching score, and select the scheme with the highest comprehensive score as the optimal configuration scheme.

[0105] Among them, the standard compliance score The calculation formula is as follows:

[0106] (2);

[0107] in, Number the candidate scheme.

[0108] Climate adaptability score The calculation formula is as follows:

[0109] (3);

[0110] in, The number of climate influencing factors, For the first The weights of each climate influencing factor, For the first The first scheme is for the... The fit score for each climate influencing factor is 0-100.

[0111] Climate adaptability score The following method is used to select climate impact factors in the calculation formula.

[0112] First, a pool of basic factors is constructed, including annual average temperature, average temperature of the coldest month, average temperature of the hottest month, extreme minimum temperature, extreme maximum temperature, annual average relative humidity, annual precipitation, maximum hourly precipitation, annual average wind speed, prevailing wind direction and speed in winter, prevailing wind direction and speed in summer, and total annual solar radiation.

[0113] Then, based on the climate zoning of the project location, the dimensionality of the basic factor pool was reduced using principal component analysis, the correlation between each climate parameter and the architectural design was calculated, and the top factors with a variance contribution ratio of ≥85% were selected. These factors serve as the core evaluation factors.

[0114] Compatibility rating The calculation rules are as follows:

[0115] Regarding the first Determine the optimal solution based on several climate influencing factors. , For the optimal design parameter value corresponding to this climate impact factor in historical cases, if the... The value of the climate impact factor parameter in each scheme is [value]. ,but ,like ≤0, let =0.

[0116] Case matching score The calculation formula is as follows:

[0117] (4);

[0118] Overall score The calculation formula is as follows:

[0119] (5).

[0120] The discriminator uses a multilayer perceptron model and historical engineering cases as benchmark data. It performs quantitative scoring by calculating the Euclidean distance between the complete solution and the benchmark data. The discriminator's work includes:

[0121] S205: The discriminator adopts a multilayer perceptron model, using the parameter range and configuration scheme of similar buildings in historical engineering cases as benchmark data. It merges the basic parameters of the building shape output by the generator with the optimal configuration scheme into a complete scheme, and performs quantitative scoring by calculating the Euclidean distance between the complete scheme and the benchmark data.

[0122] The Multilayer Perceptron (MLP) model contains three hidden layers: the first hidden layer has 128 neurons, the second hidden layer has 64 neurons, and the third hidden layer has 32 neurons. All hidden layers use the ReLU activation function, and the output layer uses a linear activation function.

[0123] Quantitative scoring The calculation formula is:

[0124] (1);

[0125] in, The value is the Euclidean distance. To adjust the coefficient, the value ranges from 0.8 to 1.2, dynamically set according to the building type: 0.8 for residential buildings, 1.0 for public buildings, and 1.2 for industrial buildings; when When setting a passing threshold, the corresponding complete solution will be judged as a valid solution.

[0126] Specifically, Euclidean distance value The calculation formula is as follows:

[0127] (6);

[0128] in, The total number of parameters involved in the calculation. For the generation scheme of the first Parameter values, The first in the historical case benchmark data The mean of each parameter.

[0129] When the quantitative score is less than the preset passing score, feedback is sent to the generator, which readjusts the conversion coefficients of the parameter mapping algorithm and regenerates the complete solution.

[0130] When the quantitative score is greater than or equal to the preset passing score, the generator and discriminator conduct adversarial training based on the complete scheme. The specific steps of the adversarial training are as follows.

[0131] S206: The generator and discriminator undergo adversarial training based on effective schemes to optimize the weight coefficients of the multilayer perceptron model. Upon meeting the iteration termination condition, a preliminary BIM seed model meeting preset requirements is output. Specifically, this includes:

[0132] S2061, Calculate the loss function: Input the quantitative score corresponding to the effective solution. loss function Defined as The reciprocal of L, that is, L = 1 / ,in, The scoring results are from 0 to 100. 0.01- .

[0133] S2062: The generator adjusts the transformation coefficients of the parameter mapping algorithm and the weight coefficients of the structure-electromechanical system matching through the backpropagation algorithm, so that the loss function continues to decrease.

[0134] S2063: The system continuously records the quantitative scores generated in each iteration. When 50 consecutive quantitative scores are all When the score is ≥85, it is determined that the generator has stably output high-quality solutions.

[0135] S2064: Quantitative scoring of ≥85 points on the 50th attempt. The corresponding complete solution is imported into the BIM modeling engine, which automatically builds a 3D model according to the specified accuracy standards, forming a BIM seed model.

[0136] Step 3, Iterative Optimization: Import the initial BIM seed model into the intelligent optimization decision module. The intelligent optimization decision module performs the following optimizations:

[0137] S301: Perform a shape adjustment on the initial BIM seed model using a topology optimization algorithm, specifically including:

[0138] a. Using the structural components in the initial BIM seed model as the optimization objects, extract the initial values ​​of the element density for each component. Unit volume and material properties, among which, Number the units;

[0139] b. Set the topology optimization objective function as follows: ,in, The penalty factor is 3-5, and N is the total number of structural units;

[0140] c. Set the constraints as follows ,and ,in, For the first The actual stress of each unit For the allowable stress of the material, For the first The actual displacement of each unit To allow displacement;

[0141] d. The optimization problem is solved using the moving asymptote algorithm, and the element density is iteratively updated using the following formula:

[0142] (7);

[0143] in, For the number of iterations, The step size factor (initially 0.01, gradually decreasing to 0.001 as the iteration converges). For the objective function at the th The gradient of the next iteration;

[0144] e. When the difference of the objective function after 10 consecutive iterations The iteration stops at a certain point, and the first optimized intermediate model after adjusting the shape of the structural components is output. .

[0145] S302: Energy consumption optimization and adjustment through energy consumption prediction models, specifically including:

[0146] f. will Import the computational fluid dynamics module to simulate and obtain the airflow velocity inside the building. Temperature field distribution and heat exchange coefficient ,in, Spatial coordinates;

[0147] g. Based on the above data, construct a long short-term memory network model for energy consumption prediction, inputting spatial coordinates and window-to-wall ratio. Heat transfer coefficient of building envelope Output region predicted energy consumption value That is, the model learns the relationship between input parameters and energy consumption, and outputs the prediction results of energy consumption in each area of ​​the building.

[0148] h. Set the loss function as follows:

[0149] Loss= (8);

[0150] in, The sample size refers to the total number of independent energy consumption zones within the building, with each zone constituting one sample. For the first The actual energy consumption value of each region; For the first Predicted energy consumption values ​​for each region;

[0151] i. The Adam optimizer is used to minimize the loss function, with the learning rate set to 0.001. The model is considered converged when Loss ≤ 0.05, at which point the model outputs... Compared with actual energy consumption The deviation meets the accuracy requirements, and then the model, based on its converged predictive ability, outputs a first-optimized intermediate model. various regions in China And filter out The system simultaneously outputs the spatial coordinates of areas where energy consumption exceeds the standard; among which... The set standard energy consumption value is used to simultaneously output energy-saving optimization suggestions.

[0152] j. Suggestions for adjusting the window-to-wall ratio:

[0153] (9);

[0154] Recommendations for adjusting the heat transfer coefficient of the building envelope:

[0155] (10);

[0156] in, The adjusted window-to-wall ratio. The adjusted heat transfer coefficient of the building envelope;

[0157] Updated based on the above optimization suggestions. Based on the window-to-wall ratio and heat transfer coefficient parameters of the building envelope, recalculate the energy consumption of each area and confirm it. Then, the intermediate model with optimized energy consumption is output. .

[0158] S303: = Layout and piping optimization, specifically including...

[0159] k. with Initial state The state space of the reinforcement learning agent includes spatial layout parameters and pipeline parameters; the spatial layout parameters include room dimensions, and the pipeline parameters include pipe diameter and pipe routing angle. The reinforcement learning agent employs the Proximal Policy Optimization (PPO) algorithm.

[0160] l. Set the optimization objectives as reducing pipeline crossing conflicts and improving space utilization, and establish a dual objective function.

[0161] (11);

[0162] in, Pipeline conflict index , For the number of intersecting pipeline pairs, This represents the total number of pipeline pairs; To maximize space utilization, , To make effective use of the area, Total building area;

[0163] m. Iterative optimization is performed using a greedy algorithm.

[0164] m1. Randomly generate 10 sets of adjustment schemes containing spatial layout and pipeline parameters;

[0165] m2. Calculate the objective function value F for each group of schemes;

[0166] m3. Retain the 3 schemes with the highest F values, and generate 10 new schemes based on these;

[0167] m4. Repeat steps m1-m3 until the F-value increases by ≤0.01 for 5 consecutive iterations;

[0168] Select the solution with the highest final F-value and update. Spatial layout and pipeline parameters, output optimized BIM seed model .

[0169] Step 4, Collaborative Verification: The platform integrates with a multi-disciplinary collaborative platform. This platform employs a collision detection algorithm combining bounding box collision detection and the GJK algorithm. The detection process includes:

[0170] S401, Collision Detection Parameter Initialization: Set the minimum safe distance between components. Define the set of professional components to be tested: architectural (walls, floors, doors and windows), structural (beams, columns, foundations), and mechanical and electrical (water supply pipes, drainage pipes, air ducts, cable trays); assign a unique identifier to each component. and the three-dimensional bounding box parameters ( , , , , , ),in, Number the component.

[0171] S402, Bounding Box Collision Detection: For any two components Calculate the overlap of its bounding box;

[0172] (12);

[0173] like If the result is positive, proceed to S403; otherwise, skip the component pair and continue performing boundary box collision detection on the remaining undetected components.

[0174] S403, Then, a secondary detection is performed using the GJK algorithm.

[0175] calculate minimum distance ,like If so, it is determined to be a collision, and the collision pair is recorded. , )and .

[0176] S404 provides collaborative correction suggestions for two types of components that may collide: pipeline detour and component position adjustment. For example, it generates electromechanical pipeline offset and pipeline elevation adjustment, or receives natural language change instructions from multidisciplinary designers (such as "raise the Z-axis of the duct by 200mm"), which are then parsed into parameter adjustment values ​​by the NLP module.

[0177] S405: Adjust and optimize the BIM seed model based on the collaborative correction suggestions; repeat S402-S405 until no collision-free component pairs are found, and finally output the corrected BIM seed model. .

[0178] Step 5, Deliverables: Input the full lifecycle verification system; the full lifecycle verification system uses a digital twin engine for construction simulation and operation and maintenance prediction.

[0179] The specific functions of the digital twin engine include:

[0180] ① Real-time mapping function: Access the Internet of Things (IoT) platform through industrial standard protocols (such as MQTT, OPC UA) to receive sensor data (such as tower crane load, personnel location, temperature and humidity) and progress data from the construction site in real time, realize the real-time mapping between physical entities and virtual models, with a time delay of ≤500ms.

[0181] ② Simulation and prediction function: based on the input BIM model It integrates a discrete event simulation (DES) engine and a system dynamics (SD) model with historical construction / operation and maintenance data. The DES engine simulates future construction processes based on construction plans, resource constraints, and real-time schedules, predicting the risk of project delays; the SD model predicts equipment lifespan and failure probability during the operation and maintenance phase based on equipment operation data and energy consumption data, and outputs visualization charts of the prediction results.

[0182] ③ Optimization suggestion generation function: When the digital twin model detects construction progress deviations (such as time lag exceeding the preset value) or equipment abnormalities (such as equipment temperature exceeding the threshold), it automatically calls the built-in rule base to generate optimization suggestions such as adjusting construction resource configuration and equipment maintenance plan.

[0183] ④ Digital twin model update: Based on actual data from the construction and operation and maintenance phases, continuously update the geometric, physical, and behavioral parameters of the digital twin model to ensure that the digital twin model remains consistent with the actual building throughout its entire life cycle.

[0184] An updated digital twin model (final construction drawings and delivered BIM model) is generated based on the simulation and prediction results, along with construction and operation optimization suggestions.

[0185] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of the invention. All equivalent changes made based on the description and drawings of the present invention are included within the scope of the present invention.

Claims

1. A BIM forward design method based on full-process AI-driven approach, characterized in that, Includes the following steps: Step 1, Digitize Requirements: Input the design brief and site survey data into the multimodal requirements analysis module; The multimodal requirements analysis module analyzes the design brief, extracts key information on building function, construction scale, and performance requirements, and automatically associates building codes and regional climate constraints through a knowledge graph, ultimately outputting three types of quantitative design indicators: floor area ratio, functional zoning, and spatial requirements. Step 2, Parameter Generation: Input the quantitative design indicators into the AI ​​parameter generation engine; The AI ​​parameter generation engine builds a parameter prediction model based on generative adversarial networks, which includes two sub-networks: a generator and a discriminator. The generator generates basic building form parameters based on quantitative design indicators, and matches two types of configuration schemes: structural system and electromechanical system, based on the basic building form parameters. It then merges the basic building form parameters and the optimal configuration scheme to output a complete scheme. The discriminator uses a multilayer perceptron model and historical engineering cases as benchmark data. It quantifies the score by calculating the Euclidean distance between the complete scheme and the benchmark data. When the quantified score is less than the preset pass line, it is fed back to the generator. The generator readjusts the transformation coefficients of the parameter mapping algorithm and regenerates the complete scheme. When the quantified score is greater than or equal to the preset pass line, the generator and discriminator conduct adversarial training based on the complete scheme to optimize the weight coefficients of the multilayer perceptron model. After meeting the iteration termination condition, it outputs a preliminary BIM seed model that meets the preset requirements. Step 3, Iterative Optimization: Import the initial BIM seed model into the intelligent optimization decision module. The intelligent optimization decision module performs the following optimizations: S301: The initial BIM seed model is morphologically adjusted using a topology optimization algorithm to adjust the component shape and output an optimized intermediate model. S302: Identify the coordinates of energy-consuming areas exceeding the standard in the first-stage optimization intermediate model through the energy consumption prediction model and provide energy-saving optimization suggestions. After adjusting the first-stage optimization intermediate model according to the energy-saving optimization suggestions, output the second-stage optimization intermediate model after energy consumption optimization. S303: Call the reinforcement learning agent, with the ability to optimize the intermediate model as the initial state, and through iteration, select the optimal spatial layout and pipeline integration scheme, and output the optimized BIM seed model; Step 4, Collaborative Verification: Integrate the optimized BIM seed model into the multi-disciplinary collaborative platform. The multi-disciplinary collaborative platform adopts a conflict detection algorithm that combines boundary box collision detection and GJK algorithm. First, it filters out component pairs that may collide through boundary box collision detection. Then, it uses GJK algorithm to filter out component pairs that will collide from the possible collision component pairs. For component pairs that will collide, it generates two types of collaborative correction suggestions: pipeline detour and component position adjustment. Based on the collaborative correction suggestions, it adjusts and optimizes the BIM seed model and outputs the corrected BIM seed model. Step 5, Deliverables: Input the revised BIM seed model into the full lifecycle verification system; the full lifecycle verification system uses a digital twin engine to perform construction simulation and operation and maintenance prediction, and generates the final deliverables based on the simulation and prediction results. The final deliverables include the final construction drawings and the delivered BIM model.

2. The BIM forward design method based on full-process AI-driven design according to claim 1, characterized in that: In step 2, the generator's work includes: S201: The plot ratio is converted into the initial values ​​of building footprint and number of floors through the parameter mapping algorithm, the functional zoning is converted into the plan dimension parameters of each area, and the space requirements are converted into the net height parameters. S202: Generate basic building form parameters for two types of buildings: floor height and column grid, based on the initial values ​​of building footprint and number of floors, plan dimensions of each area, and net height parameters. S203: Call the building codes and regional climate constraints associated in step 1, and extract the code constraint parameters and climate characteristic parameters; extract the structural-mechanical system adaptation data of similar buildings from the historical engineering case library, and establish a candidate scheme set; S204: Calculate three index scores for each candidate scheme in the candidate scheme set, including standard compliance score, climate adaptability score, and case matching score, and calculate a comprehensive score based on the standard compliance score, climate adaptability score, and case matching score, and select the scheme with the highest comprehensive score as the optimal configuration scheme. The discriminator's operation includes, S205: The discriminator employs a multilayer perceptron model, using parameter ranges and configuration schemes of similar buildings in historical engineering cases as benchmark data. It merges the basic building form parameters output by the generator with the optimal configuration scheme to form a complete scheme. A quantitative score is then calculated by comparing the Euclidean distance between this complete scheme and the benchmark data. The calculation formula is as follows: （1）； in, The value is the Euclidean distance. To adjust the coefficient, the value ranges from 0.8 to 1.2, dynamically set according to the building type: 0.8 for residential buildings, 1.0 for public buildings, and 1.2 for industrial buildings; when When setting a passing score, the corresponding complete solution will be judged as a valid solution; S206: The generator and discriminator are trained adversarially based on effective schemes to optimize the weight coefficients of the multilayer perceptron model. After meeting the iteration termination condition, the generator outputs a preliminary BIM seed model that meets the preset requirements.

3. The BIM forward design method based on full-process AI-driven design according to claim 2, characterized in that, In step S204, the specification compliance score is calculated. The calculation formula is as follows: （2）； in, Number the candidate solutions; Climate adaptability score The calculation formula is as follows: （3）； in, The number of climate influencing factors, For the first The weights of each climate influencing factor, For the first The first scheme is for the... The fit score for each climate influencing factor is 0-100; Case matching score The calculation formula is as follows: （4） ； Overall score The calculation formula is as follows: （5）。 4. The BIM forward design method based on full-process AI-driven design according to claim 3, characterized in that: Climate adaptability score The following methods are used to select climate impact factors in the calculation formula. First, construct a pool of basic factors, including annual average temperature, average temperature of the coldest month, average temperature of the hottest month, extreme minimum temperature, extreme maximum temperature, annual average relative humidity, annual precipitation, maximum hourly precipitation, annual average wind speed, prevailing wind direction and speed in winter, prevailing wind direction and speed in summer, and total annual solar radiation. Then, based on the climate zoning of the project location, the dimensionality of the basic factor pool was reduced using principal component analysis, the correlation between each climate parameter and the architectural design was calculated, and the top factors with a variance contribution ratio of ≥85% were selected. These factors serve as the core evaluation factors.

5. The BIM forward design method based on full-process AI-driven design according to claim 4, characterized in that: Compatibility rating The calculation rules are as follows: Regarding the first Determine the optimal solution based on several climate influencing factors. , For the optimal design parameter value corresponding to this climate impact factor in historical cases, if the... The value of the climate impact factor parameter in each scheme is [value]. ,but ,like ≤0, let =0.

6. The BIM forward design method based on full-process AI-driven design according to claim 2, characterized in that: In step S205, the Euclidean distance value The calculation formula is as follows: （6）； in, The total number of parameters involved in the calculation. For the generation scheme of the first Parameter values, The first in the historical case benchmark data The mean of each parameter.

7. The BIM forward design method based on full-process AI-driven design according to claim 2, characterized in that: Step S206 specifically includes: S2061, Calculate the loss function: Input the quantitative score corresponding to the effective solution. loss function Defined as The reciprocal of L, that is, L = 1 / ,in, The scoring results are from 0 to 100. 0.01- ; S2062: The generator adjusts the transformation coefficients of the parameter mapping algorithm and the weight coefficients of the structure-electromechanical system matching through the backpropagation algorithm, so that the loss function continues to decrease; S2063: The system continuously records the quantitative scores generated in each iteration. When 50 consecutive quantitative scores are all When the score is ≥85, the generator is considered to have stably output high-quality solutions. S2064: Quantitative scoring of ≥85 points on the 50th attempt. The corresponding complete solution is imported into the BIM modeling engine, which automatically builds a 3D model according to the specified accuracy standards, forming a BIM seed model.

8. The BIM forward design method based on full-process AI-driven design according to claim 1, characterized in that, S301: Perform a shape adjustment on the initial BIM seed model using a topology optimization algorithm, specifically including: a. Using the structural components in the initial BIM seed model as the optimization objects, extract the initial values ​​of the element density for each component. Unit volume and material properties, among which, Number the units; b. Set the topology optimization objective function as follows: ,in, The penalty factor is 3-5, and N is the total number of structural units; c. Set the constraints as follows ,and ,in, For the first The actual stress of each unit For the allowable stress of the material, For the first The actual displacement of each unit To allow displacement; d. The optimization problem is solved using the moving asymptote algorithm, and the element density is iteratively updated using the following formula: （7）； in, For the number of iterations, The step size factor (initially 0.01, gradually decreasing to 0.001 as the iteration converges). For the objective function at the th The gradient of the next iteration; e. When the difference of the objective function after 10 consecutive iterations The iteration stops at a certain point, and the first optimized intermediate model after adjusting the shape of the structural components is output. .

9. The BIM forward design method based on full-process AI-driven approach according to claim 8, characterized in that, S302, energy consumption optimization and adjustment are performed through an energy consumption prediction model, specifically including: f. will Import the computational fluid dynamics module to simulate and obtain the airflow velocity inside the building. Temperature field distribution and heat exchange coefficient ,in, Spatial coordinates; g. Based on the above data, construct a long short-term memory network model for energy consumption prediction, inputting spatial coordinates and window-to-wall ratio. Heat transfer coefficient of building envelope Output region predicted energy consumption value That is, the model learns the relationship between input parameters and energy consumption, and outputs the prediction results of energy consumption in each area of ​​the building. h. Set the loss function as follows: Loss= (8); in, The number of samples refers to the total number of independent energy consumption zones within the building, with each zone representing one sample. For the first The actual energy consumption value of each region; For the first Predicted energy consumption values ​​for each region; i. Minimize the loss function using the Adam optimizer. The model is considered converged when Loss ≤ 0.

05. At this point, the model output... Compared with actual energy consumption The deviation meets the accuracy requirements, and then the model, based on its converged predictive ability, outputs a first-optimized intermediate model. various regions in China And filter out The system simultaneously outputs the spatial coordinates of areas where energy consumption exceeds the standard; among which... The set standard energy consumption value is used to simultaneously output energy-saving optimization suggestions. j. Suggestions for adjusting the window-to-wall ratio: (9)； Recommendations for adjusting the heat transfer coefficient of the building envelope: (10)； in, The adjusted window-to-wall ratio. The adjusted heat transfer coefficient of the building envelope; Updated based on the above optimization suggestions. Based on the window-to-wall ratio and heat transfer coefficient parameters of the building envelope, recalculate the energy consumption of each area and confirm the results. Then, the intermediate model with optimized energy consumption is output. .

10. A BIM forward design method based on full-process AI-driven design according to claim 9, characterized in that, S303: Layout and pipeline optimization, specifically including... k. with Initial state The state space of a reinforcement learning agent includes spatial layout parameters and pipeline parameters; the spatial layout parameters include room dimensions, and the pipeline parameters include pipe diameter and pipeline orientation angle. l. Set the optimization objectives as reducing pipeline crossing conflicts and improving space utilization, and establish a dual objective function. (11)； in, Pipeline conflict index , For the number of intersecting pipeline pairs, This represents the total number of pipeline pairs; To maximize space utilization, , To make effective use of the area, Total building area; m. Iterative optimization using a greedy algorithm. m1. Randomly generate 10 adjustment schemes containing spatial layout and pipeline parameters; m2. Calculate the objective function value F for each group of schemes; m3. Retain the 3 schemes with the highest F-values, and generate 10 new schemes based on these; m4. Repeat steps m1-m3 until the F-value increases by ≤0.01 for 5 consecutive iterations; Select the solution with the highest final F-value and update. Spatial layout and pipeline parameters, output optimized BIM seed model .

Images