Building wall surface wind-driven rain grabbing rate prediction method based on machine learning

Abstract

The invention relates to a building wall surface wind-driven rain capture rate prediction method based on machine learning. The method comprises the following steps: obtaining typical building feature information; calculating the wind environment of the typical building at different wind speeds by using a numerical simulation method, and simulating and calculating the wind-driven rain distribution condition of the wall surface of the building by using an Euler multiphase flow method; extracting characteristic parameters required by machine learning model training from a numerical simulation result; data preprocessing is carried out on the extracted parameters, different machine learning models are trained, the optimal machine learning model is compared and selected for raindrop grabbing rate prediction, geometric parameters, wind field parameters and raindrop parameters serve as input of the machine learning models, and the raindrop grabbing rate serves as output of the machine learning models; and by utilizing the BEST raindrop spectrum, converting the predicted raindrop grabbing rate of each size into the total wind-driven rain grabbing rate of the building wall surface according to the rainfall intensity. Compared with the prior art, the method has the advantages of high prediction precision, high speed and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of wind-driven rain capture rate prediction, and in particular to a method for predicting the wind-driven rain capture rate of a building wall based on machine learning. Background Art

[0002] The construction sector accounts for a large proportion of global energy consumption. The thermal performance and durability of the building envelope, as the "outer coat" of the building, directly determine the heat and cold loads during the building's operation phase. The heat and moisture distribution inside the wall is an important factor affecting these two properties. Currently, the source of moisture that has the greatest impact on the changes in heat and humidity inside the wall is outdoor rainwater, that is, wind-driven rain. Wind-driven rain refers to the phenomenon that raindrops are driven by the wind during the vertical descent process, thus falling obliquely. These obliquely falling raindrops hit the building wall and are absorbed by the wall, which in turn affects the internal heat and moisture distribution. However, due to the transient wind speed, wind direction and rainfall in the environment, coupled with the complex layout of the building, it is difficult to quantify the amount of wind-driven rain captured by the building wall.

[0003] There are currently three commonly used methods to determine the wind-driven rain grab rate of building walls: field measurement method, semi-empirical model method and computational fluid dynamics numerical simulation method.

[0004] The field measurement method uses a wind-driven rain capture device to directly measure wind-driven rain at designated locations on building walls. This method can be used to validate semi-empirical models and numerical simulation results. Scholars have compiled and cross-compared the results of recent field measurements and found that, regardless of the surrounding environment, the wind-driven rain capture rate on the windward side of buildings is consistently less than 0.5. However, field measurements are inevitably limited by site and economic conditions, and are primarily conducted on low-rise buildings. Furthermore, the building site and surrounding layout have certain limitations, making it difficult to transfer the results to buildings in other environments.

[0005] The semi-empirical model method uses meteorological data obtained from weather stations or field measurements to calculate the wind-driven rainfall on a wall according to empirical formulas. Currently, there are two most commonly used semi-empirical models: ISO 15927-3:2009 and ASHRAE standard 160:2016. ISO 15927-3:2009 quantifies the effects of terrain, obstacles, wall factors, and windward surface roughness of a building on wind-driven rainfall as empirical coefficients. It calculates the wind-driven rainfall on a building's walls based on hourly meteorological data of wind speed, wind direction, and rainfall intensity and empirical formulas. ASHRAE standard 160:2016 also calculates wind-driven rainfall on a building based on wind speed and rainfall from weather stations. This model takes into account more simplified factors such as the building's height, roof type, and surrounding terrain. However, multiple studies have shown that ASHRAE's semi-empirical model often overestimates wind-driven rainfall, while ISO's semi-empirical model often underestimates it. Moreover, since the commonly used semi-empirical method assumes that the spatial distribution of wind-driven rain on the building facade is the same along the height direction or only distinguishes the top and bottom of the exterior wall, the calculation of wind-driven rain intensity on building facades in complex high-density urban environments will cause large errors.

[0006] Computational fluid dynamics (CFD) numerical simulation utilizes numerical methods to calculate wind and rain fields. This method can determine the spatiotemporal distribution of wind-driven rain capture rates on building walls for any building shape and under various rainfall conditions. Currently, two main models are commonly used for rain field calculation: 1) the Euler-Lagrangian model, which solves the equation of motion for each raindrop to determine the motion trajectories of all raindrops; and 2) the Euler-Euler model, which treats the rain phase as a continuous phase like the gas phase and solves the continuity and momentum equations for different raindrop sizes. The Euler-Euler model significantly reduces computational time and operational complexity compared to the Euler-Lagrangian model. While numerical simulation methods can provide detailed spatial distribution of raindrops, their modeling and computational processes are complex and time-consuming. Furthermore, they only perform steady-state calculations for specific building layouts. For transient rainfall events, global results must be interpolated from multiple simulated conditions.

[0007] In general, for wind-driven rainfall distribution on building walls, field measurements offer high reliability but are time-consuming and immutable; semi-empirical methods offer simplicity but low accuracy; and numerical simulations offer high accuracy but are complex and time-consuming. Consequently, a highly accurate and rapid method for calculating wind-driven rainfall distribution is currently lacking. Summary of the Invention

[0008] The purpose of the present invention is to provide a method for predicting the wind-driven rain capture rate of building walls based on machine learning, which solves the problems of complex calculation, long time consumption and low prediction accuracy in traditional methods when calculating the wind-driven rain distribution on building walls in urban residential areas.

[0009] The purpose of the present invention can be achieved by the following technical solutions:

[0010] A method for predicting wind-driven rain capture rate of building walls based on machine learning includes the following steps:

[0011] S1, obtain typical building feature information;

[0012] S2, using numerical simulation methods to calculate the wind environment of a typical building under different wind speeds, and using the Euler multiphase flow method to simulate the wind-driven rain distribution on the building wall;

[0013] S3, extracting characteristic parameters required for machine learning model training from the numerical simulation results, wherein the characteristic parameters include geometric parameters, wind field parameters, raindrop parameters, and raindrop capture rate;

[0014] S4, performing data preprocessing on the extracted parameters and using them to train different machine learning models, comparing and selecting the optimal machine learning model to predict the raindrop capture rate, wherein the machine learning model takes geometric parameters, wind field parameters, and raindrop parameters as inputs and uses the raindrop capture rate as output;

[0015] S5, using the BEST raindrop spectrum, converts the predicted raindrop capture rates of various sizes into the total wind-driven rain capture rate of the building wall according to the rainfall intensity.

[0016] The step S1 specifically includes: determining the scale information of typical urban residential buildings based on relevant standards and specifications for cities and residential areas, classifying typical residential building types based on this, and determining corresponding building characteristics.

[0017] The scale information includes building width, depth, height and distance between buildings.

[0018] The step S2 specifically includes: using the Reynolds-averaged Navier-Stokes equations or the large eddy simulation method to calculate the internal wind field of the residential building, with the top as the free slip boundary and the wall set to a standard wall function; using a multiphase flow solver to calculate wind-driven rain, with the rain phase inlet being the top and the wind field inlet boundary, and when raindrops hit the building surface, the ground or the outlet boundary, the raindrop component and the raindrop velocity normal gradient are set to zero to cause them to leave the simulation domain.

[0019] The geometric parameters include residential building width, height, building location, and calculation point coordinates on the target facade.

[0020] The wind field parameters include the inlet wind speed at the building roof height and the wind speed component corresponding to the calculation point on the target facade.

[0021] The raindrop parameters include raindrop size.

[0022] The data preprocessing includes edge line value removal and feature data normalization processing.

[0023] The different machine learning models include artificial neural networks, random forests, support vector machines, K-nearest neighbor algorithms, or multiple networks in extreme gradient boosting.

[0024] In step S4, the comparison and selection of the optimal machine learning model are specifically as follows: the prediction effects of different machine learning models are evaluated according to the set evaluation indicators, and the machine learning model with the best indicators is selected as the preliminary model; based on the test data, the raindrop capture rate is predicted using the preliminary model, and the difference between the prediction results and the numerical simulation results is compared. When the difference is less than the preset threshold, the preliminary model is used as the optimal model for subsequent predictions; when the difference is greater than or equal to the preset threshold, it is determined whether the evaluation indicator of the preliminary model is greater than the preset indicator threshold. If so, the feature parameters of the model input are optimized to obtain important features, and the preliminary model is retrained based on the important features. Otherwise, a new machine learning model is selected as the preliminary model.

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] (1) The current CFD professional software does not have the solver required for wind-driven rain simulation. The present invention can avoid the extremely complex compilation of wind-driven rain multiphase flow solvers.

[0027] (2) The present invention skips the wind-driven rain multiphase flow simulation step, which can save a lot of computing resources and time.

[0028] (3) In current urban and building performance evaluation, wind environment simulation is the basis of many evaluation analyses. Based on the results of the present invention, the wind-driven rain distribution on the wall of urban residential areas can be quickly obtained based on the wind environment.

[0029] (4) Urban environmental research requires a multi-physics coupling approach. Previous studies considering wind-driven rain required solving different physical fields in sequence during each iteration of the multi-physics coupling process. Since wind-driven rain simulations are steady-state solutions, each time step requires waiting for the solution to be re-solved, making the process extremely time-consuming. However, the present invention uses a machine learning model to directly obtain steady-state results, making the computational time of the process almost negligible and significantly improving computational efficiency.

[0030] (5) In terms of actual measurement, existing measurement schemes use wind-driven rain collectors and rain gauges to measure wind-driven rainfall on the wall. The present invention can calculate the wind-driven rainfall at any location on the wall by monitoring the wind speed at a certain distance outside the wall, the site flow wind speed, and the horizontal rainfall, which can greatly save money and time. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 is a flow chart of the method of the present invention;

[0032] Figure 2 Schematic diagram of the machine learning prediction process of the present invention;

[0033] Figure 3 Figures 1 and 2 show the numerical simulation results and artificial neural network (ANN) prediction results in one embodiment, where (a) is a schematic diagram of the residential building structure, (b) is the numerical simulation result, (c) is the ANN prediction result, and (d) is the difference between the numerical simulation result and the ANN prediction result. DETAILED DESCRIPTION

[0034] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0035] This embodiment provides a method for predicting the wind-driven rain capture rate of a building wall based on machine learning. Figure 1 As shown, the following steps are included:

[0036] S1, obtain typical building feature information.

[0037] Specifically, based on relevant urban and residential standards and specifications, we determined the dimensions of typical urban residential buildings, including building width, depth, height, and distance between buildings. This information was then used to classify typical residential building types and identify their corresponding building characteristics. This data collection provided fundamental support for numerical simulations and machine learning models.

[0038] To ensure excellent model generalization performance, this example summarizes four scales of residential building layouts based on current Chinese standards to provide machine learning training data: low-rise residential buildings (10 meters), mid-rise residential buildings (30 meters), high-rise residential buildings (50 meters), and super-high-rise residential buildings (100 meters). The residential layout consists of nine identical building units.

[0039] S2 uses numerical simulation methods to calculate the wind environment of a typical building under different wind speeds, and adopts the Euler multiphase flow method to simulate the wind-driven rain distribution on the building wall.

[0040] This process will generate the wind speed, wind direction and wind-driven rain capture rate of building walls at any location, providing basic data for subsequent parameter extraction and model training.

[0041] Specifically, the target building and its surrounding environment are modeled, and the internal wind field of the residential building is calculated based on numerical simulation software (such as OpenFOAM, Fluent, COMSOL, etc.) using methods such as the Reynolds-averaged Navier-Stokes (RANS) equations or large eddy simulation (LES). The aerodynamic roughness length at the inlet is 1 meter, the top is a free slip boundary, and the wall is set with a standard wall function. A multiphase flow solver is used to calculate wind-driven rain, and the rain phase inlet is the top and the wind field inlet boundary. When raindrops hit the building surface, the ground or the outlet boundary, the raindrop component and the raindrop velocity normal gradient are set to zero to make them leave the simulation domain. This step must meet the numerical simulation requirements of the wind environment simulation. The inlet wind speed setting is not limited, and it can be subsequently scaled down as needed to obtain the target wind speed.

[0042] S3, extracting characteristic parameters required for machine learning model training from the numerical simulation results, wherein the characteristic parameters include geometric parameters, wind field parameters, raindrop parameters, and raindrop capture rate.

[0043] In this embodiment, geometric parameters include residential building width, height, building location, and the coordinates of the calculation points on the target facade; wind field parameters include the inlet wind speed at building roof height and the wind speed components corresponding to the calculation points on the target facade; and raindrop parameters include raindrop size. To obtain the complete distribution characteristics of the wall surface, a discrete grid of points can be selected on the wall surface and the predicted values ​​of the grid can be interpolated to obtain the wind-driven rain capture rate distribution characteristics for the entire wall surface.

[0044] S4, performing data preprocessing on the extracted parameters and using them to train different machine learning models, comparing and selecting the optimal machine learning model to predict the raindrop capture rate, wherein the machine learning model uses geometric parameters, wind field parameters, and raindrop parameters as inputs and uses the raindrop capture rate as output.

[0045] In this embodiment, data preprocessing includes edge line value removal and feature data normalization processing:

[0046] 1) Edge line value removal: In the simulation, the data of the facade edge is obtained by interpolating the center values ​​of two intersecting grids. Since one of the grids has almost no capture volume, the edge value is much smaller than the actual value, so this value needs to be removed before training.

[0047] 2) Data normalization: All feature values ​​are normalized to between 0 and 1. When training the model, methods such as 5-fold cross-validation should be used to fully utilize the data and reduce the risk of overfitting.

[0048] Different machine learning models can be used, including but not limited to artificial neural networks, random forests, support vector machines, K-nearest neighbors, or various networks from extreme gradient boosting. Once a model is selected, it is trained and validated using extensive numerical simulation data to ensure high predictive accuracy and generalization capabilities.

[0049] Compare the evaluation indicators of different machine learning models, such as mean square error (MSE), coefficient of determination (R 2 ) and evaluate the performance of the model on the verification model. By comparison, the best machine learning model is selected to ensure that it has high prediction performance under different wind environments, rainfall intensities and residential layouts. Specifically: the prediction effects of different machine learning models are evaluated according to the set evaluation indicators, and the machine learning model with the best indicators is selected as the preliminary model; based on the test data, the raindrop capture rate is predicted using the preliminary model, and the difference between the prediction results and the numerical simulation results is compared. When the difference is less than 5%, it indicates that the prediction accuracy of the alternative model is higher, and the preliminary model is used as the optimal model for subsequent predictions; when the difference is greater than or equal to 5%, it is determined whether the evaluation index of the preliminary model is greater than the preset index threshold. If so, the feature parameters of the model input are optimized, and the factors affecting the wind-driven rain capture rate are fully considered to enrich the feature selection. At the same time, random forests, principal component analysis or unsupervised clustering algorithms are used to simplify features and obtain important features, and the preliminary model is retrained based on the important features. Otherwise, a new machine learning model is selected as the preliminary model.

[0050] S5, using the BEST raindrop spectrum, converts the predicted raindrop capture rates of various sizes into the total wind-driven rain capture rate of the building wall according to the rainfall intensity.

[0051] The specific implementation steps are as follows: Based on the rainfall intensity, the optimal machine learning model is used to calculate the specific capture rate of raindrops of each size; based on the distribution function of raindrops of each size in the raindrop spectrum, the capture rate of each raindrop is calculated as the global rainfall capture rate.

[0052] Ultimately, the proposed method requires only wind environment simulation results and residential area scale information to calculate the wind-driven rainfall distribution on building walls in urban residential areas using a trained machine learning model. Compared to traditional methods, this method significantly reduces computational complexity, improves prediction accuracy, and possesses greater adaptability and practicality.

[0053] Figure 2 This is a schematic diagram of the machine learning prediction process of the present invention. Figure 2 The following prediction steps are explained in detail.

[0054] In step 411, the most common multi-story residential area scale in China ( Figure 3Taking (a)) as an example, physical modeling and CFD wind field simulation are performed. The inlet wind speed at a height of 10m is 4m / s, and the target surface for grasping rate calculation is the windward side of the center of the residential area.

[0055] In step 412, parameters are extracted based on the features used in the training model. In this embodiment, 20*20 calculation points are uniformly extracted on the target, excluding edge points. The feature parameters in this embodiment are as follows:

[0056] BN: Building number (-), represents the location of the building, this parameter will affect the wind field characteristics around the building. Figure 3 As shown in (a), 5 represents that the building is in the middle of the building complex, and other values ​​represent that the building is at the edge of the building complex. The specific value is related to the wind direction.

[0057] W: Width of the target building facade (m). The facade width affects the vortex distribution in the wind field and the horizontal distribution of the wind-driven rain capture rate.

[0058] H: Height of the target building facade (m), which affects the wind field distribution and vertical distribution.

[0059] w: The horizontal coordinate of the calculation point on the target facade (m). Depending on the building width, this value needs to be normalized to -1 to 1 and increases in the positive direction of the x-axis.

[0060] h: The vertical coordinate of the calculation point on the target facade (m). Depending on the building height, this value needs to be normalized to 0 to 1 and increases in the positive direction of the z axis.

[0061] Ur: Inlet wind speed at the roof height of the target building (m / s).

[0062] Ux: The x-component of the wind speed (m / s), which is the wind speed 1m outside the target facade calculation point along the facade normal direction.

[0063] Uy: Y-direction component of wind speed (m / s), same position as Ux.

[0064] Uz: z-direction component of wind speed (m / s), same position as Ux.

[0065] RS: diameter of raindrops (m). The 17 raindrop diameters are: 0.0003, 0.0004, 0.0005, 0.0006, 0.0007, 0.0008, 0.0009, 0.001, 0.0012, 0.0014, 0.0016, 0.0018, 0.002, 0.003, 0.004, 0.005, and 0.006 m.

[0066] In step 413, the parameters extracted in the previous step are normalized according to the normalization parameters saved when training the model. x is the current eigenvalue, xnew is the normalized value, x' max The upper limit of the parameter range saved when training the model, x' min The lower limit of the parameter range saved when training the model.

[0067] In step 414, the embodiment is predicted using the optimal machine learning model. In this embodiment, the optimal model is an artificial neural network model (ANN). This step can obtain the capture rate of raindrops of various sizes.

[0068] In step 415, the raindrop capture rate is calculated as the global rainfall capture rate according to the rainfall intensity. In this embodiment, the rainfall intensity is 10 mm / h. Based on the raindrop size distribution of this rainfall intensity and the capture rates of raindrops of various sizes obtained in the previous step, the wind-driven rain capture rate of the building wall under this rainfall intensity can be obtained.

[0069] The final result is visualized in step 416 . Figure 3 Figures (b)-(d) show the numerical simulation results of this embodiment, the ANN prediction results, and the differences between the two models, demonstrating the high accuracy of this method. Furthermore, for this embodiment model, the ANN prediction only takes 7 seconds, while the wind-driven rain numerical simulation takes 2691 seconds on a multi-core workstation with 32 cores. The machine learning model prediction is 384 times faster.

[0070] The above describes in detail the preferred embodiments of the present invention. It should be understood that those skilled in the art can make numerous modifications and variations based on the concepts of the present invention without inventive effort. Therefore, any technical solutions that can be derived by those skilled in the art through logical analysis, reasoning, or limited experimentation based on the concepts of the present invention and the prior art should be within the scope of protection defined by the claims.

Claims

1. A method for predicting wind-driven rain capture rate of building wall based on machine learning, characterized in that: The following steps are involved: S1, obtain typical building feature information; S2, using numerical simulation methods to calculate the wind environment of a typical building under different wind speeds, and using the Euler multiphase flow method to simulate the wind-driven rain distribution on the building wall; S3, extracting characteristic parameters required for machine learning model training from the numerical simulation results, wherein the characteristic parameters include geometric parameters, wind field parameters, raindrop parameters, and raindrop capture rate; S4, performing data preprocessing on the extracted parameters and using them to train different machine learning models, comparing and selecting the optimal machine learning model to predict the raindrop capture rate, wherein the machine learning model takes geometric parameters, wind field parameters, and raindrop parameters as inputs and uses the raindrop capture rate as output; S5, using the BEST raindrop spectrum, converts the predicted raindrop capture rates of various sizes into the total wind-driven rain capture rate of the building wall.

2. The method for predicting wind-driven rain capture rate of building wall based on machine learning according to claim 1 is characterized in that: The step S1 specifically includes: determining the scale information of typical urban residential buildings based on relevant standards and specifications for cities and residential areas, classifying typical residential building types based on this, and determining corresponding building characteristics.

3. The method for predicting wind-driven rain capture rate of building wall based on machine learning according to claim 2 is characterized in that: The scale information includes building width, depth, height and distance between buildings.

4. The method for predicting wind-driven rain capture rate of building wall based on machine learning according to claim 1 is characterized in that: The step S2 specifically includes: using the Reynolds-averaged Navier-Stokes equations or the large eddy simulation method to calculate the internal wind field of the residential building, with the top as the free slip boundary and the wall set to a standard wall function; using a multiphase flow solver to calculate wind-driven rain, with the rain phase inlet being the top and the wind field inlet boundary, and when raindrops hit the building surface, the ground or the outlet boundary, the raindrop component and the raindrop velocity normal gradient are set to zero to cause them to leave the simulation domain.

5. The method for predicting wind-driven rain capture rate of building wall based on machine learning according to claim 1 is characterized in that: The geometric parameters include residential building width, height, building location, and calculation point coordinates on the target facade.

6. The method for predicting wind-driven rain capture rate of building wall based on machine learning according to claim 1 is characterized in that: The wind field parameters include the inlet wind speed at the height of the building roof and the wind speed component at the external specific position corresponding to the calculation point on the target facade.

7. The method for predicting wind-driven rain capture rate of building wall based on machine learning according to claim 1 is characterized in that: The raindrop parameters include raindrop size.

8. The method for predicting wind-driven rain capture rate of building wall based on machine learning according to claim 1 is characterized in that: The data preprocessing includes edge line value removal and feature data normalization processing.

9. The method for predicting wind-driven rain capture rate of building wall based on machine learning according to claim 1, characterized in that: The different machine learning models include artificial neural networks, random forests, support vector machines, K-nearest neighbor algorithms, or multiple networks in extreme gradient boosting.

10. The method for predicting wind-driven rain capture rate of building wall based on machine learning according to claim 1, characterized in that: In step S4, the comparison and selection of the optimal machine learning model are specifically as follows: the prediction effects of different machine learning models are evaluated according to the set evaluation indicators, and the machine learning model with the best indicators is selected as the preliminary model; based on the test data, the raindrop capture rate is predicted using the preliminary model, and the difference between the prediction results and the numerical simulation results is compared. When the difference is less than the preset threshold, the preliminary model is used as the optimal model for subsequent predictions; when the difference is greater than or equal to the preset threshold, it is determined whether the evaluation indicator of the preliminary model is greater than the preset indicator threshold. If so, the feature parameters of the model input are optimized to obtain important features, and the preliminary model is retrained based on the important features. Otherwise, a new machine learning model is selected as the preliminary model.

Images