Method for optimizing lighting performance of irregular window of building based on machine learning agent model

Abstract

The application discloses a building irregular window daylighting performance optimization method based on a machine learning agent model and belongs to the technical field of green buildings. The method comprises the following steps: constructing an initial digital geometric model, performing light tracking simulation operation, and verifying an effective benchmark digital model; applying a rigid body collision radius force field to the window entity in the benchmark digital model to generate a non-overlapping layout scheme, obtaining a multi-dimensional space input feature vector and a multi-dimensional performance index label of each layout scheme respectively, and constructing a CatBoost regression prediction agent model; performing feature contribution degree calculation to obtain a feature interpretation result for verifying the logical and physical consistency of the prediction logic; constructing a multi-objective fitness evaluation function embedded with a forward maximum value penalty value to obtain an optimal layout scheme of the irregular window. The application solves the problems of traditional simulation, such as time-consuming inefficiency, many invalid solutions and lack of design guiding significance, realizes fast, accurate and feasible automatic optimization of irregular window layout, and significantly improves the design efficiency.

Description

Technical Field

[0001] This invention belongs to the field of building performance analysis and green building technology, specifically involving a method for analyzing and optimizing the daylighting performance of irregular windows in buildings based on machine learning proxy models and parameterized generation technology. Background Technology

[0002] In the fields of green building and modern architectural design, natural lighting design is a core element in reducing building energy consumption, improving indoor environmental quality, and enhancing user health. In recent years, driven by architectural aesthetics and personalized needs, building facade design has gradually evolved from traditional regular rectangular arrays to irregular or irregular window layouts. Irregular window layouts not only possess unique visual appeal but also theoretically offer a wider range of parameter design possibilities. By rationally configuring the density and distribution characteristics of windows, a better balance between uniform lighting and glare control can be achieved, demonstrating significant application potential and engineering value in modern architectural design.

[0003] Currently, the analysis and optimization of building daylighting performance mainly rely on physical optics radiative transfer simulation or conventional machine learning proxy models. However, traditional physical simulation is extremely time-consuming and difficult to perform large-scale iterative optimization within the actual engineering cycle when faced with large irregular geometric combinations. In addition, existing automated optimization algorithms lack underlying physical collision detection mechanisms, which can easily generate physically infeasible invalid solutions such as window overlap or boundary crossing. Conventional machine learning models are prone to overfitting when dealing with the high-dimensional nonlinear characteristics of irregular windows, and as "black box" models, they cannot explain the physical mapping mechanism between geometric features and daylighting indicators.

[0004] The physical simulation calculations for analyzing and optimizing the daylighting performance of irregular windows are time-consuming, conventional optimization lacks geometric boundary constraints and is prone to overlapping and failing solutions, and data-driven surrogate models are prone to overfitting under small sample sizes and cannot guarantee physical rationality due to the drawbacks of "black boxes". Therefore, the existing optimization process has problems such as being time-consuming and inefficient, having many invalid solutions, and lacking design guidance. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention provides a method and system for optimizing the daylighting performance of irregular windows in buildings based on a machine learning surrogate model. By constructing a CatBoost (Categorical Boosting) regression prediction surrogate model to replace the time-consuming ray tracing simulation, a rigid body collision radius force field is introduced in the dataset generation stage, a positive maximum penalty function is embedded in the optimization stage, and a physical consistency verification module based on the SHAP (Shapley Additive exPlanations) mechanism is embedded, thus achieving multi-objective global parallel optimization that balances computational efficiency, physical feasibility, and interpretability.

[0006] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution: A method for optimizing the daylighting performance of irregular windows in buildings based on a machine learning agent model includes: Acquire architectural feature information and measured light environment data of the target room; An initial digital geometric model is constructed based on the architectural feature information; Light tracing simulations are performed on the initial digital geometric model to obtain simulated illuminance data; The accuracy of the measured light environment data is verified by comparing it with the simulated illuminance data, and a valid benchmark digital model is obtained. Apply a rigid body collision radius force field to the window entities in the baseline digital model to generate a geometric scheme library consisting of non-overlapping layout schemes; The spatial topology parameters of each layout scheme in the geometric scheme library are analyzed to obtain a multi-dimensional spatial input feature vector. Light tracing simulations were performed on each layout scheme in the geometric scheme library to obtain multi-dimensional performance index labels. The multidimensional input feature vector and the multidimensional performance index label are used to perform regression mapping training to construct a CatBoost regression prediction proxy model. Extract the internal mapping relationship of the CatBoost regression prediction surrogate model to calculate the feature contribution, and obtain the feature interpretation results used to verify the logical-physical consistency of the prediction. Based on the CatBoost regression prediction surrogate model, a multi-objective fitness evaluation function with an embedded positive maximum penalty value is constructed to find the optimal layout scheme for irregular windows.

[0007] Furthermore, obtaining architectural feature information and measured light environment data of the target room includes extracting the three-dimensional spatial scale of the target room's boundaries, the construction of the wall enclosure structure, the orientation, and the initial geometric contour boundary of the irregular window openings on the facade.

[0008] Further, illumination tracing simulations are performed on the initial digital geometric model to obtain simulated illuminance data, including: The standard meteorological documents for the historical typical meteorological year corresponding to the target building location are introduced, and meteorological parameters are obtained based on the measured experimental cycle. Based on the interior surface material of the room, the reflectivity of the interior surface material can be obtained; The meteorological parameters and the preset indoor surface material reflectivity are imported into the initial digital geometric model for attribute assignment, resulting in a digital geometric model with environmental attributes. Virtual sensors are arranged at preset coordinates on the working surface inside the digital geometric model with environmental attributes to obtain a simulated geometric model with measurement point configuration; The physical ray tracing engine is used to perform time-by-time ray tracing calculations on the simulated geometric model with measurement point configuration to obtain the hourly illuminance values ​​of each measurement point, and the combined output is the simulated illuminance data.

[0009] Furthermore, a rigid body collision radius force field is applied to the window entities in the baseline digital model to generate a geometric scheme library consisting of non-overlapping layout schemes, including: Set the total window area, number of windows, reserved structural gaps, and geometric constraints on window shape; The collision radius of the rigid outer circle of each window entity is calculated based on the geometric constraints. The collision radius and rigid body physical properties are assigned to the window entity in the baseline digital model to obtain a window entity with physical properties. A restoring force field is applied to the window entity with physical properties to prevent its geometric center from exceeding the wall boundary, thus obtaining the window entity under boundary constraint. A collision repulsion force field, triggered when the distance between two windows is less than a safety threshold, is applied to the window entity under the boundary constraint state to construct a dynamic mechanical equilibrium system. Perform an iterative calculation to minimize the energy state of the dynamic mechanical equilibrium system to obtain the set of non-interference window coordinates when the total kinetic energy of the system returns to zero; The set of non-interference coordinates of the windows is generated in batches under different window count conditions, and the geometric scheme library is output by combining them.

[0010] Furthermore, based on the window entity with physical properties, a two-dimensional constraint control domain that completely corresponds to the physical boundary of the wall is constructed, and two types of core target force fields are introduced into each window within the domain: including fence force and collision repulsion force; The fence force is used to detect the geometric centroid of the window entity in real time. Once the centroid or physical outline boundary of the window entity exceeds the boundary of the facade constraint control domain, the fence force applies a reverse restoring force to its geometric center to forcibly push it back into the boundary. The collision repulsion force is used to retrieve the Euclidean distance between two window entities in real time. Once the spatial distance is less than the sum of the collision radii of the two entities, a nonlinear interference repulsion force is triggered to prevent the components from overlapping, driving them to separate.

[0011] Furthermore, the step of analyzing the spatial topological parameters of each layout scheme in the geometric scheme library to obtain a multi-dimensional spatial input feature vector includes: Extract the number of windows for each layout scheme in the geometric scheme library for each orientation to obtain the quantity features; Calculate the normalized mean values ​​of the centroid coordinates of each facade window in the horizontal and vertical directions in each layout scheme to obtain the location center characteristics; Calculate the normalized standard deviation of the centroid coordinate distribution of windows on each facade in each layout scheme to obtain the spatial dispersion characteristics; The quantitative features, location center features, and spatial dispersion features are concatenated into a matrix according to a preset multidimensional order to generate a multidimensional spatial input feature vector.

[0012] Furthermore, ray tracing simulations are performed on each layout scheme in the geometric scheme library to obtain multi-dimensional performance index labels, including: The physical ray tracing engine is used to simulate and calculate each layout scheme in the geometry scheme library to obtain the daylight factor (DF), which reflects the degree of lighting sufficiency. The effective natural illuminance index (UDI) is obtained by statistically analyzing the proportion of time during which the illuminance of the working face for each layout scheme falls within the effective operating illuminance range. The proportion of measurement points whose working surface illuminance exceeds the preset illuminance threshold and whose duration throughout the year exceeds the preset time threshold is statistically analyzed to obtain the annual solar exposure index (ASE). The light-transmitting coefficient, effective natural light intensity index, and annual solar exposure index are mapped and spliced ​​according to the corresponding scheme, and the multidimensional performance index labels are output.

[0013] Further, the spatial input feature vector and the multidimensional performance index labels are trained through regression mapping to construct a CatBoost regression prediction surrogate model, including: The recursive feature elimination algorithm combined with cross-validation is used to evaluate and iteratively eliminate feature weights in the training set, and the optimal feature subsets for the daylight factor, effective natural light intensity index and annual solar exposure index are separated and output. The dataset containing the multidimensional spatial input feature vector and the multidimensional performance index label is divided according to a preset ratio to obtain a training set and a validation set; The optimal feature subset is used as an independent variable and input into the CatBoost regression network for nonlinear mapping. The loss function for network training is configured as root mean square error to obtain the initial prediction model. An early stopping mechanism based on real-time monitoring of the validation set error is introduced into the training loop of the initial prediction model to prevent overfitting, and a converged CatBoost regression prediction surrogate model is output.

[0014] Further, the step of extracting the internal mapping relationship of the CatBoost regression prediction surrogate model and calculating the feature contribution to obtain the feature interpretation results used to verify the logical-physical consistency of the prediction includes: Input the test sample data into the SHAP interpreter and calculate the Shapley value of each spatial distribution parameter in the optimal feature subset for the prediction result of the daylighting performance prediction surrogate model; Data reconstruction is performed based on the Shapley values ​​to generate feature summary maps and dependency maps that quantify the positive and negative contribution weights of each spatial distribution parameter to different daylighting performance indicators. The positive and negative contribution weights are compared with the prior architectural optical evolution law, and the feature interpretation result is output when they match.

[0015] Furthermore, the step of constructing a multi-objective fitness evaluation function with an embedded positive maximum penalty value based on the CatBoost regression prediction surrogate model to find the optimal layout scheme for the irregular window includes: Weighted calculations are performed on dimensionless normalized performance values, and a multi-objective weighted normalization evaluation function is configured. Set the number of windows facing each direction as the outer discrete independent variable, and perform a combined traversal within the integer domain range already learned by the daylight performance prediction proxy model to obtain the current traversal condition. Under the current traversal conditions, the differential evolution algorithm is used to perform mutation and crossover operations in the continuous coordinate domain to generate candidate decision vectors. Perform a geometrical interferometry test on the candidate decision vectors to obtain the interferometry test results; When the interference test result is a violation of the physical displacement constraint, a preset maximum penalty constant is superimposed on the original fitness value, and the current fitness value is updated and changed to the fitness value. Based on the changed fitness value, a selection operation is performed to eliminate individuals that violate the rules and trigger a maximum penalty constant. Finally, the optimal layout scheme of the irregular window with the highest comprehensive score is output iteratively.

[0016] Compared with the prior art, the present invention has at least the following beneficial effects: This invention provides a method for optimizing the daylighting performance of irregular windows in buildings based on a machine learning surrogate model, effectively solving the problems of time-consuming and inefficient processes, numerous invalid solutions, and lack of design guidance in existing technologies. On one hand, this invention replaces time-consuming ray tracing simulations with a CatBoost regression prediction surrogate model, combined with a two-layer optimization architecture, achieving efficient and rapid optimization in the high-dimensional solution space. On the other hand, it introduces a rigid body collision radius force field during the dataset generation stage and embeds a positive maximum penalty function based on Chebyshev absolute difference matrix judgment during the optimization stage, enabling the system to automatically eliminate geometrically ineffective individuals and ensuring the physical feasibility of the final layout scheme. Furthermore, by embedding a physical consistency verification module based on the SHAP mechanism, it quantitatively analyzes the contribution weights of spatial topology parameters to various daylighting indicators, reproducing the architectural optical laws of orientation-differentiated light control, and providing architects with physically logical facade design guidance. This invention's method balances computational efficiency, physical feasibility, and interpretability, providing a transparent and robust implementation path for performance-driven irregular facade generation design. Attached Figure Description

[0017] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a schematic diagram of the overall process of the present invention; Figure 2 This is a detailed flowchart of the geometric scheme library consisting of non-overlapping layout schemes corresponding to step S5 in the overall process of the present invention. Figure 3 This is a detailed flowchart of step S8 in the overall process of this invention, which corresponds to the construction of the CatBoost regression prediction surrogate model. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0019] This invention provides a method for optimizing the daylighting performance of irregular windows in buildings based on a machine learning agent model. The overall execution process of this method relies on a digital software engine and an automated computing platform deployed in a parametric building design workstation.

[0020] like Figure 1As shown in the figure, the present invention provides a method for optimizing the daylighting performance of irregular windows in buildings based on a machine learning agent model, which specifically includes the following steps: S1: Obtain architectural feature information and measured light environment data of the target room.

[0021] In this embodiment, step S1 specifically includes the following steps: By conducting on-site physical surveys or consulting existing digital design drawings, the three-dimensional spatial dimensions of the target room's boundaries, the construction of the wall enclosure structure, the orientation, and the initial geometric outline of the irregular window openings on the facade are extracted as architectural feature information.

[0022] At a predetermined height on the working surface inside the target room, several physical illuminance sensors are deployed according to a predetermined spatial grid spacing. During a predetermined experimental continuous period, the hourly illuminance measurement values ​​of each measuring point are synchronously collected using an environmental quality data acquisition instrument. After matrix aggregation, a measured dataset is constructed.

[0023] S2: Construct an initial digital geometric model based on building feature information.

[0024] In this embodiment, step S2 specifically includes the following steps: On the digital 3D parametric design platform, based on the architectural feature information extracted in step S1, the corresponding closed 3D spatial geometric structure of the target room is replicated and constructed at the same size. Based on the actual material reflection properties of the interfaces of various components inside the target room, the corresponding prior optical reflectivity parameters are assigned to each physical surface entity inside the three-dimensional spatial geometric structure.

[0025] S3: Perform illumination tracing simulation on the initial digital geometric model to obtain simulated illuminance data.

[0026] In this embodiment, step S3 specifically includes the following steps: Retrieve historical typical meteorological year standard meteorological files of the target building's geographical location, and extract historical hourly cloud cover, direct solar radiation, and diffuse solar radiation meteorological parameters within the preset experimental period; Based on the actual weather conditions recorded by the local meteorological station within the preset experimental period, the solar radiation and sky shading parameters in the standard meteorological file are scaled and fine-tuned to calculate highly realistic corrected meteorological parameters. The corrected meteorological parameters and the optical reflectivity parameters assigned in step S2 are imported into the initial digital geometric model as boundary conditions to assign attributes, resulting in a digital geometric model with environmental attributes. Inside the initial digital geometric model with environmental attributes, corresponding virtual sensor components are arranged at the same relative positions of the three-dimensional spatial coordinates of each physical illuminance sensor in step S1 to obtain a simulated geometric model with measurement point configuration. The ray tracing engine based on the principle of physical optics radiation transmission is launched to perform time-by-time reverse ray tracing and luminous flux conservation simulation calculations on the simulated geometric model with measurement point configuration, so as to obtain the hourly illuminance values ​​of each measurement point and extract the hourly simulated illuminance data of each virtual sensor in the corresponding measured period.

[0027] S4: Compare the measured data of the light environment with the simulated illuminance data to verify the accuracy and obtain a valid benchmark digital model.

[0028] In this embodiment, step S4 specifically includes the following steps: The time-series simulated illuminance data corresponding to each measuring point are mapped and paired one-to-one with the measured dataset obtained in step S1 according to the time index. Using linear correlation analysis, the coefficient of determination R between the simulated illuminance sequence and the measured illuminance sequence at each measuring point throughout the entire period was calculated. 2 And root mean square error; Set a minimum accuracy threshold to determine the reliability of the model, and calculate the coefficient of determination R at each measurement point. 2 Perform a comparison with this precision threshold; If the predictive determination coefficient R of all measuring points is detected 2 If all values ​​are strictly greater than the preset accuracy threshold, then the temporal dynamic response of the current initial digital geometric model under dynamic light environment disturbance is determined to have physical authenticity. The current digital geometric model is then output and stored as a valid benchmark digital model.

[0029] S5: Apply a rigid body collision radius force field to the window entities in the baseline digital model to generate a geometric scheme library consisting of non-overlapping layout schemes.

[0030] In this embodiment, step S5 proceeds as follows: Figure 2 As shown, the specific steps include: In order to control the single independent variable of the total window area in the subsequent irregular shape evolution optimization, all window entities on the corresponding wall domain are abstracted into square solid components with rigid body physical properties and identical side lengths.

[0031] Set the total window area S Number of windows N Parameters for reserved structural gaps δ Obtain the geometric constraints of the window shape, including the preset side length of the square solid component. l and the parameters of the reserved structural gaps in the project δ Based on the principle of the geometric envelope of the circumcircle, through the formula = + The collision radius R of the circumscribed circle of the rigid body of each square solid window component was calculated. Randomly generate N window center points within the wall area; The calculated collision radius R and rigid body mass properties are attached to each window entity in the baseline digital model to obtain window entities with physical properties. A two-dimensional constraint control domain that completely corresponds to the physical boundary of the wall is then constructed inside the physics solver.

[0032] Two types of core target force fields are introduced into each window within the defined domain: one is the fence force, which is used to detect the geometric centroid of the window entity in real time. Once the centroid or physical outline boundary of the window entity exceeds the boundary of the facade constraint control domain, the fence force applies a reverse restoring force to its geometric center and forcibly pushes it back into the boundary; the other is the collision repulsion force, which is used to retrieve the Euclidean distance between two window entities in real time. Once the spatial distance is less than the sum of the collision radii of the two entities, a nonlinear interference repulsion force is triggered to prevent the component entities from overlapping, driving them to separate from each other.

[0033] Apply a collision repulsion force field triggered when the distance between two windows is less than a safety threshold to a window entity under boundary constraints to construct a dynamic mechanical equilibrium system.

[0034] The fence force and collision repulsion force are applied to each window entity in real time, and the energy minimization iterative solution loop is started. The solver adaptively adjusts the two-dimensional plane coordinates of each window entity until the total kinetic energy of the entire multi-rigid-body force system is completely reduced to zero and reaches a stable equilibrium state, thus obtaining the set of non-interference coordinates of the windows when the total kinetic energy of the system is reduced to zero. The set of non-interference coordinates of the windows is generated in batches under different window number conditions, and the set of non-interference coordinates of the windows when the total energy of the current system is the lowest is output.

[0035] Discretely preset combinations of topological working conditions with different window numbers, and perform independent mechanical collision iterations a preset number of times under each working condition based on different random seeds. Finally, batch combination outputs a geometric solution library composed of multiple sets of physically feasible layout samples.

[0036] S6: Spatial topological parameters of each layout scheme in the analytical geometry scheme library, to obtain spatial input feature vectors.

[0037] In this embodiment, step S6 specifically includes the following steps: Understand and analyze the absolute geometric coordinate set of each irregular window opening sample scheme in the geometric scheme library, and extract the number of windows on each orientation facade as a quantitative feature; Extract the two-dimensional components of the geometric centroid coordinates of all windows on each orientation facade in a single scheme in the vertical direction, calculate the mathematical mean and standard deviation of the component set, and divide them by the physical total height of the corresponding facade wall to perform normalization scaling, generating vertical center features and vertical dispersion features that can characterize the vertical height tendency and dispersion. Extract the two-dimensional components of the geometric centroid coordinates of all windows on each facing facade in the horizontal direction, calculate their mathematical mean, subtract half of the total width of the corresponding wall plane, and then divide by half of the total width to perform centering zero-point scaling, generating a horizontal center feature that represents the tendency of horizontal position offset; at the same time, calculate the standard deviation of all horizontal plane components, divide by half of the total width of the wall to perform proportional scaling, generating a horizontal dispersion feature that represents the degree of horizontal dispersion. The quantitative features, vertical center features, vertical dispersion features, horizontal center features, and horizontal dispersion features of each facade are matrix-concatenated according to a preset fixed multidimensional order, and the combined output is a multidimensional spatial input feature vector used to eliminate window generation sorting interference.

[0038] S7: Perform ray tracing simulation calculations on each layout scheme in the geometry scheme library to obtain multi-dimensional performance index labels.

[0039] In this embodiment, step S7 specifically includes the following steps: All irregular window opening scheme samples from the geometric scheme library are sequentially imported into the benchmark digital model as boundary conditions. The high-fidelity physical optics simulation batch calculation kernel is launched to perform ray tracing simulation calculations covering the entire year's time sequence for each irregular scheme of each layout; Extract and statistically analyze the multidimensional performance indicators on the indoor working surface for each group of schemes: The daylight factor (DF) measures the adequacy of natural daylight in an indoor space under static, overcast conditions. The Effective Natural Illuminance Index (UDI) is used to calculate the cumulative percentage of time during which the temporal illuminance of an indoor work surface falls within the effective operating illuminance range under dynamic daylighting throughout the year. The annual solar exposure index (ASE) is used to calculate the percentage of the working face that exceeds the standard due to the penetration of direct sunlight with high illuminance throughout the year, and is used to quantitatively indicate the risk of local heat overload and overexposure glare. The DF, UDI, and ASE indices of each irregular scheme calculated by simulation are concatenated in sequence to form a dependent variable label matrix, and the combined output is a multidimensional performance index label.

[0040] S8: Train the CatBoost regression prediction surrogate model by performing regression mapping between the spatial input feature vector and the multidimensional performance index labels.

[0041] In this embodiment, step S8 proceeds as follows: Figure 3 As shown, the specific steps include: A recursive feature elimination algorithm (RFECV) combined with cross-validation is introduced to predict the coefficient of determination R. 2 As a scoring metric for iterative feature addition and deletion, the training set is subjected to a high-dimensional spatial dimension decreasing stepwise analysis. Redundant spatial variables with low explanatory power for each daylighting evaluation index are successively removed, and the optimal feature subset for DF, UDI and ASE prediction is separated and output. The dataset containing multidimensional input feature vectors and multidimensional performance index labels is randomly and mutually exclusively divided into training set and validation set according to a preset quantity ratio. The optimal feature subsets corresponding to each performance index are used as independent variables and input into the CatBoost regression machine learning network. The loss function for network training is configured as root mean square error (RMSE). The validation set data is introduced as a real-time anti-overfit validation input into the training loop of the regression machine learning network to activate the early stopping detection mechanism. During the network training iteration, the rate of change of the validation set on the RMSE loss function is monitored in real time. Once it is detected that the RMSE error of the validation set no longer decreases within a preset number of consecutive detection steps, the training blocking operation is automatically triggered, and the weight parameters inside the network are automatically rolled back to the optimal state corresponding to the lowest point of the validation set error. Finally, a CatBoost regression prediction surrogate model with converged multidimensional error and strong generalization prediction ability is output.

[0042] S9: Extract the internal mapping relationship of the CatBoost regression prediction surrogate model to calculate the feature contribution, and obtain the feature interpretation results used to verify the logical-physical consistency of the prediction.

[0043] In this embodiment, step S9 specifically includes the following steps: The independent test sample dataset is input into the feature interpretability analyzer SHAP based on the principle of cooperative game theory, and the Shapley marginal contribution value of each spatial topology parameter in the optimal feature subset to the final prediction output of the daylight performance prediction surrogate model is calculated. Data is reconstructed based on the calculated Shapley marginal contribution value, and feature summary response and dependency response data are generated for each daylighting performance index to quantitatively present the positive and negative driving characteristics of each spatial independent variable on the dependent variable on different daylighting performance indexes. The actual driving direction of each facade space independent variable extracted from the feature summary and dependent response data on the dependent variable is compared with the classic a priori evolution law of architectural physics optics to test whether the pure data-driven machine learning model has learned the real physical mechanism of light control internally. The model is deemed qualified and the daylighting performance prediction proxy model is officially output and stored only if the mathematical response characteristics of the current initial model are completely consistent with the virtual and real response states of the prior architectural optical evolution law, and it is confirmed that there is no blind mathematical fitting that violates the physical laws within the model.

[0044] S10: Based on the CatBoost regression prediction surrogate model, a multi-objective fitness evaluation function with embedded positive maximum penalty value is constructed to find the optimal layout scheme of irregular windows.

[0045] In this embodiment, step S10 specifically includes the following steps: Weighted calculation of dimensionless normalized performance values:

[0046] Among them, DF norm UDI norm and ASE norm These are the normalized daylight factor, effective natural light intensity, and annual solar exposure index, respectively. ω 1. ω 2. ω 3 represents the weighting coefficients for the normalized daylight factor, effective natural illuminance, and annual solar exposure, respectively.

[0047] Configure a multi-objective weighted normalized evaluation function; The number of windows on each facing facade is set as the discrete topological decision independent variable of the outer layer. In order to prevent the accuracy of the predictive agent model from collapsing when extrapolating in an unknown domain, the optimization interval is strictly limited to the discrete integer interval that the previous model has learned. The current traversal condition is determined by iterating through all discrete window number topological combinations in the outer loop structure. In each defined current traversal condition, the inner optimization program initiates a global search of the continuous solution space based on the differential evolution (DE) algorithm, generating candidate decision vectors within the continuous wall domain based on preset mutation and crossover operators. x ; The inner optimization process optimizes candidate decision vectors. x Real-time execution of embedded static physical feasibility constraint interference detection: extract the two-dimensional geometric center coordinates of any two window entities on the same wall, calculate the absolute value of their difference in the horizontal plane and the absolute value of their difference in the vertical plane, and extract the maximum absolute difference between the two as the spatial judgment distance for determining the distance between the two square entities. When the spatial judgment interval is detected to be less than the side length of the square window minus the preset micro-deviation tolerance threshold, it is determined that the currently generated window individual has exceeded the boundary or caused an overlap conflict in the physical structure of the facade. The inner program then activates the penalty function response mechanism, adding a preset maximum penalty constant to the original fitness value. M Update the current fitness value to the fitness value. F(x) – M This allows for the use of the natural evolutionary selection operation of the optimization algorithm to directly filter out and eliminate individuals with geometric failures that trigger the maximum penalty constant in the current generation.

[0048] The optimization algorithm continuously updates and iterates based on the final fitness value until the preset convergence relative tolerance threshold is met or the maximum number of evolution iterations is reached. After all inner and outer loops are completed, the algorithm integrates and outputs the optimal irregular window layout scheme with the minimum comprehensive fitness value, i.e. the highest comprehensive performance score.

[0049] The present invention will now be described with reference to specific embodiments.

[0050] In this embodiment, a real-world example of a building facade lighting refinement design is selected to provide detailed data illustration of the specific implementation of the present invention. A typical room in the "Milan Building" of Xi'an Jiaotong University, located in the Western China Science and Technology Innovation Port, is chosen as the research object. The facade of this building features a highly randomized irregular window design. A typical teaching room on the sixth floor of this building, with a space size of 5.1m × 3.3m × 2.8m, is selected as the specific target room.

[0051] In this embodiment, three typical measuring points A, B, and C were arranged at a preset height of 0.75m on the working surface inside the target room. Continuous illuminance data were collected from September 14th to 16th, 2025, using a METREL MI6401 indoor environmental quality comprehensive analyzer with a range of 0-20000 lux and an accuracy of 0.1 lux. The TMYx meteorological file of Xi'an Xianyang International Airport was incorporated into the simulation, and the parameters of the standard meteorological file were fine-tuned by scaling according to the specific weather conditions of the actual measurement days. A digital geometric model was constructed in the Ladybug Tools environment of Grasshopper, and the indoor material reflectance was set as follows: whitewashed walls: 0.75, dark wooden doors: 0.10, white glazed tiles: 0.80.

[0052] The simulated sensor values ​​were compared with the measured dataset using a Pearson correlation test. The coefficient of determination R0 between the simulated and measured values ​​at the three measurement points was calculated. 2 All values ​​are significantly higher than the judgment threshold of 0.70, indicating that the current geometric model can accurately capture the dynamic change trend of the indoor light environment over time, meets the benchmark accuracy requirements, and is stored as the benchmark digital model.

[0053] In this embodiment, to independently study the impact of "window spatial distribution" on lighting and avoid interference from size and aspect ratio, all windows on the same wall are set to the same square. To address the common issues of window overlap and boundary crossing in window layouts, a parameterized generation method integrating the Kangaroo2 physics solver is introduced. To prevent window entities from overlapping at the edges, the collision radius of the rigid circumcircle of the window entity is set to... = + (in l The side length of a square window (To construct the gap). The fence force and collision repulsion force are loaded into the solver. The solver eliminates all geometric conflicts by minimizing the total energy of the system. When the total kinetic energy of the system is zero, it outputs a non-overlapping rigid body physical coordinate solution.

[0054] Based on this method, 25 possible combinations of window counts N∈{1, 2, 4, 6, 8} on the west and south walls were defined. For each combination, 30 independent layout schemes were generated, resulting in a total of 750 physically feasible irregular window layout samples. The daylight factor (DF) and effective natural illuminance (UDI) of each scheme were then simulated in batches using the Radiance kernel. 100-2000lx ) and annual solar exposure index (ASE) 1000lx, 250h (Indicator value)

[0055] Daylight factor (DF) measures the sufficiency of natural daylight indoors under static, overcast conditions; Effective Natural Illuminance Index (UDI) 100-2000lx The percentage of effective time when the illuminance on the working surface falls within the range of 100-2000 lx; ​​annual solar exposure index (ASE) 1000lx, 250h () is the proportion of the area of ​​the measuring points where the working surface illuminance exceeds 1000 lx and the duration is greater than 250 h throughout the year.

[0056] To eliminate the interference of window sorting on model training, discrete coordinates are transformed into statistical feature vectors X:

[0057] in, N Indicates the number of windows; X and Y These represent the horizontal and vertical directions, respectively. norm This represents the normalized positional mean. std The standard deviation of the location distribution; w and s These represent the west wall and the south wall, respectively.

[0058] This completes the construction of the feature performance mapping dataset.

[0059] In this embodiment, the RFECV method is used to filter the 10-dimensional features. After filtering, the features targeted are DF and UDI. 100-2000lx and ASE 1000lx, 250h The optimal feature subsets have 4, 9, and 9 features respectively. The dataset is divided into training and testing sets in an 8:2 ratio. A surrogate model is built based on the CatBoost algorithm, with an RMSE loss function, a learning rate of 0.03, and an early stopping parameter of 50 detection steps introduced into the training loop. Evaluation results show that DF and UDI... 100-2000lx and ASE 1000lx, 250h The prediction model predicts the coefficient of determination R. 2 Above 0.7, the residuals exhibit a good normal distribution.

[0060] The SHAP method was used to internally map and deconstruct the model prediction results, calculate the Shapley value, and quantify the contribution weights of each distribution feature to the daylighting index. Physical consistency self-verification revealed that the south wall parameters affect DF and ASE. 1000lx, 250h The total importance percentages were 56.51% and 54.19% respectively, consistent with the physical law of southward-dominant diffuse total luminous flux; the west wall parameters at UDI 100-2000lx The predicted proportion is as high as 59.25%, and when the horizontal position of the west wall Xnorm_w is slightly to the left, it is beneficial to extend the reflection path of the incident light and expand the coverage of the medium illuminance area, which is consistent with the passive anti-glare law of avoiding strong direct sunlight from the southwest in the afternoon. The mathematical fitting direction of the current proxy model is completely consistent with the prior architectural optical evolution law, indicating that the model has high robustness and reliability.

[0061] In this embodiment, the maximum and minimum values ​​of each index in the feature performance mapping dataset from historical simulations are read. The trained daylighting performance prediction surrogate model is then used to predict the DF and UDI of any unknown scenario. 100-2000lx and ASE 1000lx, 250h The absolute value is calculated, and discrete maximum-minimum normalization is applied to it to obtain a dimensionless performance value that eliminates dimensional differences. DFnorm , UDInorm , ASEnorm .

[0062] Since the global optimization algorithm defaults to a cost-value minimization search strategy, in order to balance the effective acquisition of natural lighting and strictly suppress the risk of glare and overexposure, the multi-objective fitness objective function is configured using the following equal-weight combination formula:

[0063] Among them, DF norm UDI norm and ASE norm These are the normalized daylight factor, effective natural light intensity, and annual solar exposure index, respectively. ω 1. ω 2. ω 3 represents the weighting coefficients for the normalized daylight factor, effective natural illuminance, and annual solar exposure, respectively, set to 0.3, 0.5, and 0.2.

[0064] The number of windows Nw on the west-facing facade and the number of windows Ns on the south-facing facade are defined as the outer discrete topological decision variables. To prevent accuracy collapse of the machine learning model during extrapolation to unknown domains, the optimization interval is strictly limited to the integer interval Nw∈[1, 8] and Ns∈[1, 8] that the model has previously learned. The program iterates through these 64 discrete window number topological combinations sequentially using an outer double loop structure to determine the current iteration condition.

[0065] In each defined current traversal condition, the inner optimization program initiates a continuous spatial search based on the differential evolution (DE) algorithm. The decision variables are configured as the two-dimensional geometric center coordinate sequence of each window entity, and the differential evolution parameters are configured as follows: population size of 10, maximum number of evolution iterations of 100, and relative tolerance threshold of 0.01.

[0066] In each iteration of differential evolution, the algorithm generates candidate decision vectors within a preset boundary. x The inner optimization program incorporates a static geometric constraint detection and processing mechanism. The program automatically extracts candidate decision vectors. x For any two window entities on the same wall, calculate the absolute value of their difference in the horizontal direction and the absolute value of their difference in the vertical direction, and extract the maximum absolute difference between the two. The maximum absolute difference is used as the judgment distance for determining the distance between the two window entities.

[0067] When a window with a spacing smaller than the side length of a square window on the corresponding facade minus a preset micro-deviation tolerance threshold of 0.01 is detected, it is determined that the currently generated window has exceeded the boundary or overlapped in the physical structure of the facade. The inner program then initiates a penalty function response, adding a preset maximum penalty constant to the original fitness value. M = 99999, the current fitness value has been updated to... F(x) – M This allows the violating geometric individual to be directly eliminated in the natural selection operation of differential evolution.

[0068] When candidate decision vectors are detected xWhen no geometric interference is triggered and all facade physical feasibility constraints are met, the program calculates the corresponding 10-dimensional spatial statistical feature vector, inputs it into the daylighting performance prediction proxy model for millisecond-level performance prediction, calculates the dimensionless value, and outputs the true fitness value through a multi-objective adaptive objective function. The differential evolution algorithm performs crossover and selection updates based on the final fitness value, continuously converging towards the region that minimizes the fitness value (i.e., maximizes the overall daylighting performance score). After all inner and outer loop calculations are completed, the Pareto front solution set is integrated, and finally, the optimal irregular window layout scheme with the highest overall score and morphological convergence in the facade prototype of "a large number of discrete windows on the south wall and windows on the west wall arranged high and slightly to the left" is output.

[0069] This invention is not limited to the above embodiments. Based on the technical solutions disclosed in this invention, those skilled in the art can make some substitutions and modifications to some of the technical features without creative effort, and all such substitutions and modifications are within the protection scope of this invention.

Claims

1. A method for optimizing the daylighting performance of irregular windows in buildings based on a machine learning agent model, characterized in that, include: Acquire architectural feature information and measured light environment data of the target room; An initial digital geometric model is constructed based on the architectural feature information; Light tracing simulations are performed on the initial digital geometric model to obtain simulated illuminance data; The accuracy of the measured light environment data is verified by comparing it with the simulated illuminance data, and a valid benchmark digital model is obtained. Apply a rigid body collision radius force field to the window entities in the baseline digital model to generate a geometric scheme library consisting of non-overlapping layout schemes; The spatial topology parameters of each layout scheme in the geometric scheme library are analyzed to obtain a multi-dimensional spatial input feature vector. Light tracing simulations were performed on each layout scheme in the geometric scheme library to obtain multi-dimensional performance index labels. The multidimensional input feature vector and the multidimensional performance index label are used to perform regression mapping training to construct a CatBoost regression prediction proxy model. Extract the internal mapping relationship of the CatBoost regression prediction surrogate model to calculate the feature contribution, and obtain the feature interpretation results used to verify the logical-physical consistency of the prediction. Based on the CatBoost regression prediction surrogate model, a multi-objective fitness evaluation function with an embedded positive maximum penalty value is constructed to find the optimal layout scheme for irregular windows.

2. The method for optimizing the daylighting performance of irregular windows in buildings based on a machine learning agent model according to claim 1, characterized in that, Acquiring architectural feature information and measured light environment data of the target room includes extracting the three-dimensional spatial scale of the target room's boundary, the construction of the wall enclosure structure, the orientation, and the initial geometric contour boundary of the irregular window openings on the facade.

3. The method for optimizing the daylighting performance of irregular windows in buildings based on a machine learning agent model according to claim 1, characterized in that, Light tracing simulations are performed on the initial digital geometric model to obtain simulated illuminance data, including: Retrieve historical typical meteorological year standard meteorological files of the target building's geographical location and extract meteorological parameters within the preset experimental period; Based on the interior surface material of the room, the reflectivity of the interior surface material can be obtained; The meteorological parameters and the preset indoor surface material reflectivity are imported into the initial digital geometric model for attribute assignment, resulting in a digital geometric model with environmental attributes. Virtual sensors are arranged at preset coordinates on the working surface inside the digital geometric model with environmental attributes to obtain a simulated geometric model with measurement point configuration; The physical ray tracing engine is used to perform time-by-time ray tracing calculations on the simulated geometric model with measurement point configuration to obtain the hourly illuminance values ​​of each measurement point, and the simulated illuminance data is output by combining the results.

4. The method for optimizing the daylighting performance of irregular windows in buildings based on a machine learning agent model according to claim 1, characterized in that, Applying a rigid body collision radius force field to the window entities in the baseline digital model generates a geometric scheme library consisting of non-overlapping layout schemes, including: Set the total window area, number of windows, reserved structural gaps, and geometric constraints on window shape; The collision radius of the rigid outer circle of each window entity is calculated based on the geometric constraints. The collision radius and rigid body physical properties are assigned to the window entity in the benchmark digital model to obtain a window entity with physical properties. A restoring force field is applied to the window entity with physical properties to prevent its geometric center from exceeding the wall boundary, thus obtaining the window entity under boundary constraint. A collision repulsion force field, triggered when the distance between two windows is less than a safety threshold, is applied to the window entity under the boundary constraint state to construct a dynamic mechanical equilibrium system. Perform an iterative calculation to minimize the energy state of the dynamic mechanical equilibrium system to obtain the set of non-interference window coordinates when the total kinetic energy of the system returns to zero; The non-interference coordinate set of the windows is generated in batches under different window count conditions, and the resulting geometric scheme library is combined and output.

5. The method for optimizing the daylighting performance of irregular windows in buildings based on a machine learning agent model according to claim 4, characterized in that, Based on the window entity with physical properties, a two-dimensional constraint control domain that completely corresponds to the physical boundary of the wall is constructed, and two types of core target force fields are introduced into each window in the domain: including fence force and collision repulsion force. The fence force is used to detect the geometric centroid of the window entity in real time. Once the centroid or physical outline boundary of the window entity exceeds the boundary of the facade constraint control domain, the fence force applies a reverse restoring force to its geometric center to forcibly push it back into the boundary. The collision repulsion force is used to retrieve the Euclidean distance between two window entities in real time. Once the spatial distance is less than the sum of the collision radii of the two entities, a nonlinear interference repulsion force is triggered to prevent the components from overlapping, driving them to separate.

6. The method for optimizing the daylighting performance of irregular windows in buildings based on a machine learning agent model according to claim 1, characterized in that, The spatial topological parameters of each layout scheme in the geometric scheme library are analyzed to obtain a multi-dimensional spatial input feature vector, including: Extract the number of windows for each layout scheme in the geometric scheme library on each facing facade to obtain the quantitative features; Calculate the normalized mean values ​​of the geometric centroid coordinates of all windows on each orientation facade in each layout scheme in the horizontal and vertical directions to obtain the location center characteristics; Calculate the normalized standard deviation of the centroid coordinate distribution of windows on each facade in each layout scheme to obtain the spatial dispersion characteristics; The quantitative features, location center features, and spatial dispersion features of each facade are matrix-concatenated according to a preset multidimensional order to generate a multidimensional spatial input feature vector.

7. The method for optimizing the daylighting performance of irregular windows in buildings based on a machine learning agent model according to claim 1, characterized in that, Light tracing simulations were performed on each layout scheme in the geometry scheme library to obtain multi-dimensional performance index labels, including: The physical ray tracing engine is used to simulate and calculate the layout schemes in the geometric scheme library to obtain the daylight factor, which reflects the degree of lighting sufficiency. The effective natural illuminance index is obtained by calculating the cumulative proportion of time during which the illuminance of the working face for each layout scheme falls within the effective operating illuminance range. The annual cumulative excess area of ​​the working face corresponding to each layout scheme that is penetrated by high-intensity direct sunlight is calculated to obtain the annual solar exposure index. The light-transmitting coefficient, effective natural light intensity index, and annual solar exposure index are mapped and spliced ​​according to the corresponding scheme to output multi-dimensional performance index labels.

8. The method for optimizing the daylighting performance of irregular windows in buildings based on a machine learning agent model according to claim 1, characterized in that, The spatial input feature vector and the multidimensional performance index labels are used to perform regression mapping training to construct a CatBoost regression prediction surrogate model, including: The recursive feature elimination algorithm combined with cross-validation is used to evaluate and iteratively eliminate feature weights in the training set, and the optimal feature subsets for the daylight factor, effective natural light intensity index and annual solar exposure index are separated and output. The dataset containing multidimensional input feature vectors and multidimensional performance index labels is divided into training set and validation set according to a preset ratio. The optimal feature subset is input as an independent variable into the CatBoost regression learning network for nonlinear mapping. The loss function for network training is configured as root mean square error to obtain the initial prediction model. An early stopping mechanism based on real-time monitoring of the validation set error is introduced into the training loop of the initial prediction model to prevent overfitting, and a converged CatBoost regression prediction surrogate model is output.

9. The method for optimizing the daylighting performance of irregular windows in buildings based on a machine learning agent model according to claim 1, characterized in that, Extracting the internal mapping relationships of the CatBoost regression prediction surrogate model and calculating feature contributions yields feature interpretation results used to verify the logical-physical consistency of predictions, including: Input the test sample data into the SHAP interpreter and calculate the Shapley marginal contribution value of each spatial distribution parameter in the optimal feature subset to the prediction result of the daylight performance prediction surrogate model. Data is reconstructed based on the Shapley marginal contribution value to generate a feature summary map and a dependency map that quantifies the positive and negative contribution weights of each spatial distribution parameter to different daylighting performance indicators. The positive and negative contribution weights are compared with the prior architectural optical evolution law, and the feature interpretation result is output when the match is consistent.

10. The method for optimizing the daylighting performance of irregular windows in buildings based on a machine learning agent model according to claim 1, characterized in that, Based on the CatBoost regression prediction surrogate model, a multi-objective fitness evaluation function with an embedded positive maximum penalty is constructed to find the optimal layout scheme for irregular windows, including: Weighted calculation of dimensionless normalized performance values: Among them, DF norm UDI norm and ASE norm These are the normalized daylight factor, effective natural light intensity, and annual solar exposure index, respectively. ω 1. ω 2. ω 3 represents the weighting coefficients for the normalized daylight factor, effective natural illuminance, and annual solar exposure, respectively. Configure a multi-objective weighted normalized evaluation function; Set the number of windows facing each direction as the outer discrete independent variable, and perform a combined traversal within the integer domain range already learned by the daylight performance prediction proxy model to obtain the current traversal condition. Under the current traversal conditions, the differential evolution algorithm is used to perform mutation and crossover operations in the continuous coordinate domain to generate candidate decision vectors. Perform a geometrical interferometry test on the candidate decision vectors to obtain the interferometry test results; When the interference test result indicates a violation of the physical displacement constraint, a preset maximum penalty constant is added to the original fitness value, and the current fitness value is updated and changed to the fitness value. Based on the changed fitness value, a selection operation is performed to eliminate individuals that violate the rules and trigger a maximum penalty constant. Finally, the optimal layout scheme of the irregular window with the highest comprehensive score is output iteratively.

Images