Method, device and medium for building wind environment prediction under sparse measurement

Abstract

The application provides a building wind environment prediction method and device under sparse measurement and a medium, and has the characteristics that the method comprises the following steps: collecting flow field measurement and environmental parameter data of a research area, constructing a training data set based on a corresponding data set construction method according to the data completeness, automatically extracting building area features based on flow field characteristics by using a preset building influence factor extraction model, generating a building influence factor field as a soft geometric constraint, inputting sparse boundary conditions of a to-be-predicted working condition into a preset PINN wind environment prediction model, and obtaining a global high-resolution wind environment distribution. The application effectively solves the problems of high calculation cost of a traditional CFD and poor physical consistency of a pure data-driven method under sparse data, and significantly improves the prediction accuracy and engineering applicability.

Description

Technical Field

[0001] This document relates to the field of building wind environment prediction technology, and in particular to a method, equipment and medium for predicting building wind environment under sparse measurement. Background Technology

[0002] The flow field around a building typically involves complex phenomena such as flow separation, vortex shedding, and reattachment. Traditional computational fluid dynamics methods have advantages in simulation accuracy, but their solution process involves frequent iterations and complex mesh generation, making it difficult to achieve rapid scheme comparison or real-time environmental monitoring in the early stages of design. While wind tunnel tests and field measurements can provide reliable data, deploying high-density sensor networks across the entire area is costly and impractical.

[0003] Physical Information Neural Networks (PINNs) introduce physical constraints by embedding the governing equations into a loss function, theoretically enabling high-precision flow field predictions even with limited sample sizes. In the field of building wind engineering, Rui et al. first applied PINN to the reconstruction of 3D flow fields around buildings, achieving good results using only sparse wind speed data near the walls. They also proposed a dynamic priority adaptive loss balancing strategy to address the scale mismatch problem in multi-objective optimization. Wei et al. further applied PINN to the reconstruction of 2D street canyon flow fields and wind pressure distribution estimation, finding that even under extremely sparse conditions with only 16 measuring points, PINN's prediction error could still be reduced by approximately 30% compared to purely data-driven methods. These studies demonstrate that PINN exhibits good tolerance to sparse measuring points and can save approximately 42% of computation time compared to traditional CFD.

[0004] However, existing research still has the following limitations: First, it mainly focuses on single or a few working conditions and lacks a systematic evaluation of generalization performance under different meteorological conditions; Second, in terms of building geometric boundary treatment, traditional methods rely on complex boundary condition implementation or symbolic distance function, which increases model complexity; Third, existing PINN models have difficulty simultaneously ensuring prediction accuracy and physical consistency under sparse measurement data. Summary of the Invention

[0005] This invention provides a method, device, and medium for predicting building wind environment under sparse measurement, aiming to solve the above-mentioned problems.

[0006] According to an embodiment of the present invention, a method for predicting building wind environment under sparse measurement is provided, comprising: S1. Collect measured flow field and environmental parameter data of the study area, and construct a training dataset based on the data completeness using the corresponding dataset construction method; S2. Using a preset building impact factor extraction model, the building area features are automatically extracted based on the flow field characteristics to generate a building impact factor field as a soft geometric constraint. S3. Input the sparse boundary conditions of the working condition to be predicted into the preset PINN wind environment prediction model to obtain the high-resolution wind environment distribution across the entire domain. The PINN wind environment prediction model uses the building influence factor field as a soft geometric constraint input, constructs a physical constraint loss function based on the eddy transport equation, and is trained using a three-stage dynamic training strategy.

[0007] According to an embodiment of the present invention, an electronic device is provided, characterized in that it comprises: Processor; and, A memory is configured to store computer-executable instructions, which, when executed, cause the processor to implement the steps of the above-described method for predicting building wind environment under sparse measurements.

[0008] According to an embodiment of the present invention, a storage medium is provided, characterized in that it is used to store computer-executable instructions, which, when executed, implement the steps of the above-described method for predicting building wind environment under sparse measurement.

[0009] According to embodiments of the present invention, by introducing the eddy transport equation to eliminate the pressure term, a strong physical constraint is directly imposed on the measurable velocity field, effectively suppressing common non-physical artifacts and oscillations under sparse data, and ensuring the spatial continuity and physical rationality of the flow field. Combined with a velocity layered sampling strategy based on velocity amplitude, the training bias caused by the long-tail distribution of complex flow fields is overcome, achieving high-precision reconstruction under sparse measurement points or incomplete data conditions. At the same time, by using the building influence factor (BIF) as a soft geometric constraint, cumbersome grid generation and complex boundary processing are avoided, significantly improving modeling efficiency, and providing a solution with both high data efficiency and engineering applicability for urban microclimate monitoring and rapid assessment.

[0010] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in one or more embodiments of this specification or in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1This is a flowchart of the building wind environment prediction method under sparse measurement according to an embodiment of the present invention; Figure 2 This is an overall flowchart illustrating a specific implementation of an embodiment of the present invention; Figure 3 This is a simplified diagram of the urban geometric model of the study area in this embodiment of the invention. Figure 4 This is a combined distribution heatmap of historical meteorological data according to an embodiment of the present invention. Figure 5 This is a partial dataset diagram of an embodiment of the present invention; Figure 6 This is a diagram of the PINN architecture according to an embodiment of the present invention; Figure 7 This is a building factor diagram of an embodiment of the present invention; Figure 8 This is a result analysis diagram of an embodiment of the present invention; Figure 9 This is a graph showing the variation of the loss function in an embodiment of the present invention; Figure 10 The above are error statistics and error box plots from embodiments of the present invention. Detailed Implementation

[0013] To enable those skilled in the art to better understand the technical solutions in one or more embodiments of this specification, the technical solutions in one or more embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this specification, and not all of the embodiments. Based on one or more embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of this document.

[0014] Method Implementation Examples According to embodiments of the present invention, a method for predicting building wind environment under sparse measurement is provided. Figure 1 This is a flowchart of the building wind environment prediction method under sparse measurement according to an embodiment of the present invention, such as... Figure 1 As shown, the building wind environment prediction method under sparse measurement according to an embodiment of the present invention specifically includes: S1. Collect measured flow field and environmental parameter data of the study area, and construct a training dataset based on the data completeness using the corresponding dataset construction method; Figure 2 This is a flowchart illustrating the overall implementation of the present invention. Figure 2 It can be seen that this method mainly includes four stages: data collection and dataset construction, building impact factor extraction, PINN model training and prediction output.

[0015] Specifically, a micro-scale urban densely built-up area is selected; in this embodiment of the invention, the selected micro-scale is less than 2 kilometers in horizontal dimension. A coordinate system for the study area is established to provide a spatial positioning basis for subsequent flow field calculations and model input.

[0016] Collect measured flow field data and environmental parameter data in the study area to determine the completeness of the flow field-related data. Data completeness refers to having a sufficient number of sensor measured data with adequate coverage. If the data is complete, a training dataset can be directly constructed using sparse measurement point data. Specifically, limited sensor measurement point data, such as location coordinates and corresponding wind speed and temperature values, can be directly used as data constraint points.

[0017] If the data is incomplete, reference flow field data is obtained through high-precision simulation, and a training dataset is constructed by velocity hierarchical sampling based on velocity amplitude after obtaining the reference flow field data.

[0018] The high-precision simulation specifically includes: using remote sensing technology to establish a three-dimensional geometric model of the study area based on OSM road network data and building geometry data by acquiring the target city location range; setting boundary conditions such as inlet velocity and temperature; and running computational fluid dynamics (CFD) simulation to obtain reference flow field data for the area under specific operating conditions. The physical boundary conditions are determined by using the daily average temperature over historical years a (a≥3). To build an efficient PINN algorithm development environment, a mature two-dimensional steady-state simplification strategy from urban wind engineering is further adopted, using a horizontal section at a pedestrian height of 1.5m as the computational domain. This strategy significantly reduces the computational dimensionality while preserving key geometric features, providing an ideal data foundation for deep learning models. Specific settings are shown in Table 1. Table 1 CFD simulation parameter settings

[0019] The velocity-level sampling based on velocity amplitude specifically includes: dividing the entire data space into different flow characteristic intervals based on the statistical distribution characteristics of the flow field velocity amplitude; performing balanced sampling in each interval according to the contribution of each flow characteristic interval to model training, so that the model can achieve balanced learning of samples from different flow characteristic regions and alleviate the training bias caused by the non-uniform distribution of flow field data.

[0020] S2. Using a preset building impact factor extraction model, the building area features are automatically extracted based on the flow field characteristics to generate a building impact factor field as a soft geometric constraint. The steps for constructing the Building Impact Factor (BIF) extraction model include: S21. Based on the physical characteristics of the flow field, calculate the characteristic indicators used to identify the building's stagnant and shear regions. S21 specifically includes: First, based on the CFD initial field data, two key physical indicators are defined to identify potential building areas: relative velocity ratio and velocity gradient norm.

[0021] Relative speed ratio: To identify extremely low-speed zones inside the building and near the walls, relative velocity is defined. : (1); in, For the local velocity magnitude, For reference velocity, the 90th percentile of the total velocity magnitude is taken to represent the free flow velocity, thus avoiding distortion of the reference velocity due to local extrema.

[0022] Velocity gradient norm: To capture the strong shear layers and vortex structures at the edges of the building silhouette, the norm of the velocity gradient tensor is calculated. : (2); The central difference method is used to calculate the directional derivatives and screen out high gradient regions.

[0023] S22. Construct a building area identification scoring field based on the aforementioned feature indicators; Specifically, by integrating information from low-speed and high-gradient regions, an original building scoring field is constructed. An indicator function is then introduced for logical judgment. (3); in, The low-speed threshold is set to 0.05. The gradient threshold is the 85th percentile of the gradient. and These are the weighting coefficients for the extremely low-speed region and the high-gradient region, respectively, used to balance the contributions of the building's internal resistance effect and the edge shear effect.

[0024] S23. Perform spatial smoothing and morphological optimization processing on the building area identification scoring field to generate a continuous and noise-free building influence factor field.

[0025] The original scoring field is typically discontinuous and noisy. Considering that the Parametric Physical Information Neural Network (PINN) calculates derivatives through automatic differentiation during training, good spatial continuity of the input features is required. Therefore, this embodiment of the invention employs Gaussian kernel convolution to smooth the scoring field. Furthermore, to eliminate isolated noise points and fill in voids within the building, morphological opening and closing operations are applied before generating the final mask. This ensures that the building influence factor accurately and smoothly conforms to the real building contour, providing reliable soft geometric constraints for the model.

[0026] To verify the accuracy of the extraction algorithm, typical working conditions were selected, and the scoring regions were binarized with a threshold b (0 ≤ b ≤ 1). The accuracy of the extracted building outlines was verified to ensure a high degree of consistency with the actual geometric boundaries. This indicates whether the building influence factor can serve as a soft geometric constraint, effectively embedding boundary conditions into the PINN input space.

[0027] S3. Input the sparse boundary conditions of the working condition to be predicted into the preset PINN wind environment prediction model to obtain the high-resolution wind environment distribution across the entire domain. The PINN wind environment prediction model uses the building influence factor field as a soft geometric constraint input, constructs a physical constraint loss function based on the eddy transport equation, and is trained using a three-stage dynamic training strategy.

[0028] The PINN wind environment prediction model construction method includes: Constructing the network input layer: The network input is spatial steady-state coordinates, specifically including the following six channels: (1) x-coordinate and y-coordinate: These two channels constitute the reference spatial coordinate system of the model, enabling the model to perceive and learn the absolute positional relationship of the wind field in space.

[0029] (2) Ambient temperature: describes the air temperature field at the inlet boundary and is an important thermodynamic parameter that affects air density and flow.

[0030] (3) Environmental inlet wind speed: Provides the inlet driving conditions for wind field simulation and defines the basic intensity of the incoming wind.

[0031] (4) Building surface temperature: reflects the thermal condition of the building surface and affects the air flow near the wall.

[0032] (5) Building Impact Factor: This is a key feature constructed in the embodiments of the present invention, used to quantify the physical interference of buildings on the wind field. This factor explicitly identifies areas where buildings may cause a sudden drop in wind speed or severe shear.

[0033] Construct a fully connected neural network with 4 hidden layers. The hidden layers use the Tanh activation function, specifically including: A fully connected neural network with four hidden layers and 50 neurons in each layer is constructed. This depth and width setting ensures that the model has sufficient fitting capacity while effectively avoiding the risk of overfitting caused by overparameterization. All hidden layers use the hyperbolic tangent function (Tanh) as the activation function to ensure effective gradient propagation and smoothness, satisfying the continuity requirements of higher-order derivatives in physical equations.

[0034] The network output layer settings include: setting the output layer to flow field velocity components u and v, directly outputting the wind speed field that is of most concern to the project, without the need for additional wind pressure conversion, which is more suitable for sparse measurement scenarios.

[0035] Specifically, the physical constraint loss function constructed based on the eddy transport equation includes: The pressure term in the control equation is eliminated by introducing vorticity variables, as follows: In engineering practice of building wind environment, the assessment of key issues such as pedestrian wind comfort and pollutant dispersion directly depends on wind speed distribution, and obtaining wind speed data through anemometers is relatively convenient and economical. To construct a predictive model with wind speed as the core variable, the influence of the pressure term *p* in modeling and solving is avoided, and an eddy current transport model is adopted to describe the flow. This model, by introducing the vortex variable, eliminates the pressure term from the governing equations, allowing physical constraints to directly act on the wind speed field. This solves the pain points of wind pressure being difficult to measure and constrain.

[0036] The mass conservation equation is: (4); The momentum conservation equation is: (5); (6); In a two-dimensional flow field, vorticity ω is defined as the plane component of the curl of the velocity vector: (7); To eliminate the unobservable pressure term p, we use the cross-differentiation method to take the partial derivative of equation (5) with respect to x, and the partial derivative of equation (6) with respect to y: (8); (9); The subtracted convection-diffusion term can be simplified to the vorticity convection-diffusion term by expansion and using the continuity equation: Diffusion = (10); Convection = (11); The PDE physical residual loss is constructed based on the two-dimensional steady-state eddy transport equation, including: The two-dimensional steady-state eddy transport equation can be expressed as: (12); The total loss is a weighted average of data loss and physical residual loss.

[0037] In the Parametric Physical Information Neural Network (PINN) model of this invention, we use the residual defined by the final two-dimensional steady-state vortex transport equation as the momentum equation part of the physical loss function LPDE. Specifically, the model calculates the higher-order derivatives of the velocity field with respect to spatial coordinates by automatic differentiation, and then directly solves the residual of the equation, forcing the neural network output to satisfy the wind field distribution that satisfies the law of conservation of hydrodynamics.

[0038] The output impairment function constraint of the neural network is jointly determined by the data loss (Ldata) and the physical loss (LPDE). The data loss arises from the difference between the predicted value and the selected baseline value, while the physical loss arises from the difference between the dynamic flow field predicted by the neural network and the governing equations.

[0039] In view of the characteristics of large-scale urban blocks and complex environment in the embodiments of the present invention, the model specifically removes the explicit initial condition (IC) and boundary condition (BC) constraints in the design of the loss function, mainly for the following two reasons: (1) In predicting the average wind field distribution in urban microenvironment, the physical object of interest is the steady-state flow field. According to the Reynolds-averaged (RANS) assumption, the flow field variables do not change significantly with time. Therefore, the governing equations degenerate into elliptic partial differential equations, and the time dimension is compressed, thus mathematically eliminating the need for initial conditions.

[0040] (2) In real-world urban-scale simulations, accurately obtaining real-time velocity profiles at regional boundaries is often extremely challenging and involves significant uncertainty. Furthermore, grid areas outside of building areas are not considered part of this region, and adding boundary condition loss would greatly increase the computational cost of the model.

[0041] Finally, the model's total loss function Ltotal is defined as the weighted sum of the data-driven terms and the physical residual terms, i.e.: (13); in and To balance the weighting coefficients of different loss terms, this approach not only avoids the systematic errors that may be introduced by manually setting boundary conditions, but also greatly improves the model's adaptability to different wind speeds and building layouts.

[0042] To ensure that the flow field predicted by the neural network is consistent with the CFD data at the observation point, while also considering the convergence speed of training and robustness to abrupt changes in the flow field, this embodiment of the invention employs a composite data-driven loss function. This function is composed of a weighted average of mean squared error and mean absolute error. The data loss term is expressed as follows: (14); in: (15); (16); In the formula, and These are the weighting coefficients for the two losses, respectively. For network speed prediction, This is the actual velocity vector.

[0043] Since the loss function of PINN is composed of both data residuals and complex physical partial differential equation (PDE) residuals, there is a significant gradient competition between the two in the optimization space. Direct joint training can easily lead to the optimization process getting trapped in local minima or exhibiting gradient pathology. Therefore, this embodiment of the invention employs a three-stage dynamic training strategy. The AdamW optimizer is used throughout, utilizing its decoupled weight decay mechanism to provide better regularization. The specific settings for each stage are as follows: The first stage is the basic fitting stage, where the physical loss weight is set to 0, and a first learning rate is used to allow the network to learn the flow field topology. The first stage aims to guide the neural network to quickly learn the approximate topology of the flow field and establish a preliminary data mapping. In a specific implementation of this invention, this stage is set to 100 iterations, and the initial learning rate is relatively high (2×10⁻⁶). 3 Because the initial physical constraint weights are relatively low, the network can avoid failing to converge due to complex PDE gradient noise, thus quickly approximating the data manifold.

[0044] The second stage is the physics enhancement stage, which enables physics constraint loss and uses a second learning rate (lower than the first learning rate) to correct predictions that do not conform to physical laws. Building upon the initial fit, this second stage introduces enhanced physics constraints. The iteration count is set to 500 steps, and the initial learning rate is lowered to 8×10⁻⁶. 4 The focus at this stage is to correct predictions that, while conforming to the data distribution, violate principles of fluid mechanics (such as mass conservation), and to initially conform to physical laws.

[0045] The third stage is the fine-tuning stage, where the weighting of data fitting is reduced, and a third learning rate, lower than the second, is adopted to allow the model to achieve parameter convergence within the solution space that satisfies the physical governing equations. This is the most time-consuming and crucial stage, aiming to solve for high-precision physical fields. The number of iterations is set to 60,000, and the initial learning rate is significantly reduced to 2×10⁻⁶. 4 This is combined with an early stopping mechanism. By reducing the relative weight of data fitting, the model is forced to fine-tune within a very small solution space, ensuring that the final model can accurately fit the observed data while strictly adhering to the physical control equations.

[0046] The sparse boundary conditions include inlet flow parameters, environmental thermodynamic parameters, and target region spatial coordinates. The model outputs the global wind field distribution within seconds after input. The trained PINN model is then deployed. For new operating conditions to be predicted, only the sparse boundary conditions (such as inlet wind speed and temperature) and the spatial coordinates of the target region at that moment need to be input. The model can then output a high-resolution global wind environment (velocity field, temperature field) distribution within seconds, achieving rapid prediction.

[0047] To further illustrate the above method, the following example is used for specific explanation: The example subject is a typical high-density residential block in Chaoyang District, Beijing (approximately 0.33 km²). This area has a complex building layout and typical street canyon features, making it a typical sample for urban wind environment research. Figure 3 This is a simplified diagram of the urban geometric model of the study area in this embodiment of the invention, calculated according to the above method. Historical meteorological data of Beijing from January 1, 2019 to April 1, 2025 were collected, and a total of 5,911 days of valid data were screened. Based on the joint distribution probability of wind speed and temperature, Latin hypercube sampling (LHS) was used to extract 80 representative independent working conditions. 64 of these working conditions were divided into a training set and 16 into a test set. A CFD simulation was established to obtain the reference flow field. A velocity stratification sampling strategy based on velocity amplitude was adopted, and eight different sampling densities were set from extremely sparse (N=210) to sparse (N=700).

[0048] A parameterized PINN network was constructed with four hidden layers, each containing 50 neurons, and the Tanh activation function was selected. Training was performed using a three-stage dynamic training strategy. The basic fitting stage involved 100 iterations (learning rate...). The physics enhancement phase iterates 500 times (learning rate). The fine-tuning phase iterates 60,000 times (learning rate). The results show that, under extremely sparse data (N=210), the coefficient of determination (R²) of this method reaches 0.50, which is about 14.4% higher than that of a purely data-driven neural network; when N=700, R² increases to 0.81. Furthermore, the flow field generated by this method exhibits good spatial continuity and effectively suppresses artifacts and non-physical oscillations. Figure 4 This is a joint distribution heatmap of historical meteorological data according to an embodiment of the present invention. Figure 5 This is a partial dataset diagram of an embodiment of the present invention. Figure 6 This is a diagram of the PINN architecture according to an embodiment of the present invention. Figure 7 This is a building factor diagram according to an embodiment of the present invention. Figure 8 This is a result analysis diagram of an embodiment of the present invention. Figure 9 This is a graph showing the variation of the loss function according to an embodiment of the present invention. Figure 10 The above are error statistics and error box plots from embodiments of the present invention.

[0049] According to embodiments of the present invention, by introducing the eddy transport equation to eliminate the pressure term, a strong physical constraint is directly imposed on the measurable velocity field, effectively suppressing common non-physical artifacts and oscillations under sparse data, and ensuring the spatial continuity and physical rationality of the flow field. Combined with a velocity layered sampling strategy based on velocity amplitude, the training bias caused by the long-tail distribution of complex flow fields is overcome, achieving high-precision reconstruction under sparse measurement points or incomplete data conditions. At the same time, by using the building influence factor (BIF) as a soft geometric constraint, cumbersome grid generation and complex boundary processing are avoided, significantly improving modeling efficiency, and providing a solution with both high data efficiency and engineering applicability for urban microclimate monitoring and rapid assessment.

[0050] Device Example 1 According to an embodiment of the present invention, an electronic device is provided, characterized in that it comprises: Processor; and, A memory is configured to store computer-executable instructions, which, when executed, cause the processor to perform the steps of the above-described method embodiments.

[0051] Device Example 2 According to an embodiment of the present invention, a storage medium is provided, characterized in that it is used to store computer-executable instructions, which, when executed, implement the steps of the above-described method embodiments.

[0052] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for building wind environment prediction under sparse measurements, characterized in that, include: S1. Collect measured flow field and environmental parameter data of the study area, and construct a training dataset based on the data completeness using the corresponding dataset construction method; S2. Using a preset building impact factor extraction model, the building area features are automatically extracted based on the flow field characteristics to generate a building impact factor field as a soft geometric constraint. S3. Input the sparse boundary conditions of the working condition to be predicted into the preset PINN wind environment prediction model to obtain the high-resolution wind environment distribution across the entire domain. The PINN wind environment prediction model uses the building influence factor field as a soft geometric constraint input, constructs a physical constraint loss function based on the eddy transport equation, and is trained using a three-stage dynamic training strategy.

2. The method according to claim 1, characterized in that, S1 specifically includes: Assess the completeness of flow field data in the study area; If the data is complete, the training dataset can be constructed directly using sparse measurement point data; If the data is incomplete, reference flow field data is obtained through high-precision simulation, and a training dataset is constructed using velocity stratification sampling based on velocity amplitude.

3. The method according to claim 2, characterized in that, The velocity-level sampling based on velocity amplitude specifically includes: Based on the statistical distribution characteristics of the velocity amplitude in the flow field, the entire data space is divided into intervals with different flow characteristics; Based on the contribution of each flow characteristic interval to model training, equal sampling is performed within each interval to enable the model to learn samples from different flow characteristic regions in a balanced manner, thereby mitigating training bias caused by non-uniform distribution of flow field data.

4. The method according to claim 1, characterized in that, The steps for constructing the building impact factor extraction model include: Based on the physical characteristics of the flow field, characteristic indices are calculated to identify the building's stagnant and shear regions. Construct a building area identification scoring field based on the aforementioned feature indicators; Spatial smoothing and morphological optimization are performed on the building area identification scoring field to generate a continuous and noise-free building influence factor field.

5. The method according to claim 1, characterized in that, The PINN wind environment prediction model construction method includes: A fully connected neural network with 4 hidden layers is constructed, and the hidden layers use the Tanh activation function. The input layer is set to include spatial coordinates, environmental inlet parameters, building surface parameters, and building impact factors; Set the output layer to the flow field velocity component.

6. The method according to claim 1, characterized in that, The physical constraint loss function constructed based on the vortex transport equation specifically includes: Introducing vorticity variables eliminates the pressure term in the control equations; Constructing PDE physical residual loss based on two-dimensional steady-state eddy transport equation; The total loss is a weighted average of data loss and physical residual loss.

7. The method according to claim 1, characterized in that, The three-stage dynamic training strategy specifically includes: The first stage is the basic fitting stage, in which the physical loss weight is set to 0 and the first learning rate is used to enable the network to learn the flow field topology. The second stage is the physical enhancement stage, which enables physical constraint loss and uses a second learning rate lower than the first learning rate to correct prediction results that do not conform to physical laws. The third stage is the fine-tuning stage, which reduces the weight of data fitting and adopts a third learning rate that is lower than the second learning rate, so that the model can complete parameter convergence within the solution space that satisfies the physical control equations.

8. The method according to claim 1, characterized in that, The sparse boundary conditions include inlet flow parameters, environmental thermodynamic parameters, and target region spatial coordinates. The model outputs the global wind field distribution within seconds after input.

9. An electronic device, characterized in that, include: processor; as well as, A memory configured to store computer-executable instructions, which, when executed, cause the processor to perform the steps of the building wind environment prediction method under sparse measurements as described in any one of claims 1 to 8.

10. A storage medium, characterized in that, Used to store computer-executable instructions, which, when executed, implement the steps of the building wind environment prediction method under sparse measurement as described in any one of claims 1 to 8.

Images