Downburst average wind speed prediction method and system based on regression random forest
By combining the theory of impact jets and Haldane sequence sampling with a regression random forest approach, the hyperparameters of the model are optimized, a training dataset is constructed, and the regression random forest model is trained. This solves the problem of high cost and high time consumption in predicting the wind speed time history of moving downbursts, and achieves real-time and accurate wind speed prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- 中铁科学研究院集团有限公司
- Filing Date
- 2026-04-14
- Publication Date
- 2026-05-12
AI Technical Summary
Existing technologies struggle to predict the wind speed time history of moving downbursts quickly and cost-effectively while maintaining prediction accuracy, especially for accurate predictions of unsteady wind fields.
A computational fluid dynamics numerical simulation model was constructed using a regression random forest-based approach, combined with impingement jet theory and Haldane sequence sampling. The model hyperparameters were optimized through parameter sensitivity analysis, and a training dataset was built to train the regression random forest model, enabling rapid prediction of the average wind speed of downbursts.
It enables real-time and accurate prediction of wind speed time histories of moving downbursts, improving computational efficiency by tens of millions of times. It can capture complex nonlinear characteristics, especially bimodal characteristics, and provides efficient and accurate support for wind-resistant design of engineering structures.
Smart Images
Figure CN122021471A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of wind engineering and artificial intelligence technology, and in particular to a method and system for predicting the average wind speed of downbursts based on regression random forest. Background Technology
[0002] A downburst is a strong wind weather phenomenon formed when a powerful downdraft from the base of a thunderstorm cloud impacts the ground and spreads outwards. Its sudden onset, short duration, and immense destructive power make it a significant threat to the structural safety of power transmission lines, long-span bridges, and high-rise buildings. Moving downbursts, in particular, have a more complex near-surface wind field structure due to the added velocity of the storm center, and their impact on structures often exceeds that of stationary downbursts. Therefore, accurately predicting the wind speed time-history characteristics of moving downbursts is crucial for the wind-resistant design of engineering structures and for disaster prevention and mitigation.
[0003] Currently, the simulation and prediction of downburst wind fields mainly employ methods such as field measurements, wind tunnel tests, and computational fluid dynamics (CFD) numerical simulations. While field measurements offer the most realistic representation, the probability of capturing a complete downburst event is extremely low, resulting in a severe lack of data. Although wind tunnel tests and CFD numerical simulations can reproduce the wind field characteristics of downbursts, wind tunnel tests are time-consuming, costly, and difficult to accurately simulate storm movement effects. While CFD numerical simulations offer high accuracy, they often require enormous computational resources and time. For example, for unsteady-state simulations of moving downbursts, a single calculation can often take hours or even days, making it difficult to meet the needs of rapid assessment or real-time early warning of massive amounts of data in practical engineering projects.
[0004] In recent years, with the rapid development of artificial intelligence technology, machine learning algorithms have shown great potential in the field of wind speed prediction. However, most existing wind speed prediction models are designed for conventional atmospheric boundary layer wind fields or steady-state wind fields, while research on predictions for extreme wind fields with strong nonlinear and non-stationary characteristics, such as moving downbursts, is relatively limited. How to significantly reduce computational costs while ensuring prediction accuracy, and achieve rapid prediction of downburst wind speed time histories under arbitrary combinations of moving speeds and ambient wind speeds, is a pressing technical challenge that needs to be addressed. Summary of the Invention
[0005] This invention provides a method and system for predicting the average wind speed of downbursts based on regression random forests. It can replace time-consuming CFD numerical simulations and improves computational efficiency by tens of millions of times while ensuring prediction accuracy. It enables real-time and accurate prediction of the wind speed time history of moving downbursts under any working conditions.
[0006] This invention provides a method for predicting the average wind speed of downbursts based on regression random forest, characterized by comprising: S1. Based on the theory of impact jets, a computational fluid dynamics numerical simulation model of a moving downburst is established. S2. Determine the range of variation of ambient wind speed and storm center movement speed to construct a two-dimensional parameter space. Use Haldane sequence sampling method to perform uniform sampling in the two-dimensional parameter space to generate parameter combination samples containing different ambient wind speeds and storm center movement speeds. S3. Input the parameter combination sample into the computational fluid dynamics numerical simulation model for calculation, and construct a training dataset containing multiple sets of input feature variables and target output variables based on the wind speed time history data of the monitoring points obtained from the calculation; wherein, the input feature variables include ambient wind speed, storm center movement speed and time point, and the target output variable is the average wind speed of the downburst at the corresponding time point; S4. Construct an initial regression random forest model and train it using the training dataset. During the training process, use parameter sensitivity analysis to optimize the number of decision trees and the maximum depth of decision trees to obtain the optimal hyperparameter combination. The trained regression random forest model is then used as the downburst average wind speed prediction model. S5. Obtain the ambient wind speed, storm center movement speed, and time series to be predicted for the working condition to be predicted, input them into the downburst average wind speed prediction model, and obtain the downburst average wind speed prediction result for the corresponding time series under the working condition.
[0007] Furthermore, S1 specifically includes: S101. A three-dimensional impact jet model is adopted to construct a three-dimensional cylindrical computational domain, and a jet inlet is set at the top of the computational domain. For moving downbursts, the entire computational domain is divided into a stationary region and a moving nozzle region, and the interface of the regions is processed using sliding mesh technology. S102. The computational domain is discretized using a hexahedral structured mesh, and the mesh is refined in the near-ground region. A sliding mesh interface between the stationary region and the moving nozzle region is constructed using an arbitrary coupled mesh interface. S103. The unsteady Reynolds-averaged Navier-Stokes equations are adopted as the fluid control equations, and SST-k is used. Turbulence models are used to analyze flow separation phenomena, and the SIMPLIC algorithm is used to solve pressure-velocity coupling problems. S104. Set the velocity components and moving velocity of the jet inlet, set the ambient wind speed inlet, set the ground no-slip wall condition, and set the boundary conditions of the top and sides of the computational domain to obtain the computational fluid dynamics numerical simulation model of the moving downburst.
[0008] Furthermore, S2 specifically includes: S201, Select ambient wind speed and the speed of movement at the center of the storm As key control parameters, the variation range of ambient wind speed was determined. and the range of change in the speed of the storm center's movement ; S202. Select two distinct prime numbers as the bases for the environmental wind speed dimension and the storm center movement speed dimension, respectively. Use the Halton sequence generation algorithm to calculate the Halton sequence value of each sample point in the two dimensions, and generate N uniformly distributed dimensionless sample points. S203. Using the inverse normalization method, the dimensionless sample points are mapped back to the range of environmental wind speed change and the range of storm center movement speed change, and the corresponding actual physical parameter values are calculated. S204. Treat each set of actual physical parameter values obtained by calculation as an independent working condition sample, and summarize them to generate a parameter combination sample set for driving computational fluid dynamics numerical simulation.
[0009] Furthermore, in S202, the calculation process of the Halton sequence value is as follows: For the nth sample point, first expand its index value n in cardinality b: ; Then, the Haldane sequence value of this sample point at cardinality b is calculated. : ; in, The k-th digit of n in base b ; m is the highest digit; the base of the environmental wind speed dimension is 2, and the corresponding Halton sequence value is The dimension of the storm center's movement speed is selected with a base of 3, and the corresponding Halton sequence value is... ; In S203, the formula for calculating the actual physical parameter values is as follows: ; ; in, Let n be the ambient wind speed corresponding to the nth sample. The speed at which the storm center moves corresponds to the nth sample. and These are the minimum and maximum values of the ambient wind speed, respectively. and These are the minimum and maximum values of the storm's center movement speed, respectively.
[0010] Furthermore, S3 specifically includes: S301. The parameter combination sample set is used as input conditions in sequence and assigned to the motion properties of the environmental wind speed inlet boundary and the moving area of the computational fluid dynamics model respectively. The solver is started to perform unsteady-state calculations until the preset total simulation time is reached. S302. Pre-set monitoring points in the computational domain. During the simulation of each set of operating conditions, record the velocity vector at the monitoring points in real time, extract its horizontal radial component as the average wind speed of the downburst, and output the wind speed time history curve of the monitoring point over time for each set of operating conditions; wherein, the monitoring point is located on the axis of the storm's movement path, and the average wind speed of the downburst is the radial velocity in the downburst wind field. Curve of change with time t ; S303. Discretize the wind speed time history curve, divide the continuous time into multiple time points, extract the instantaneous wind speed value corresponding to each time point, associate each time point with the simulated environmental wind speed and storm center movement speed, construct an input feature vector, and use the corresponding instantaneous wind speed value as the target output scalar. S304. Iterate through all working conditions and their corresponding time points, summarize all generated input feature vectors and target output scalars, and construct a training dataset containing multiple sets of input feature variables and target output variables.
[0011] Further, in step S303, the input feature vector is: ; The target output scalar is: ; in, This is the input feature vector at the k-th time point under the n-th working condition. Let be the ambient wind speed corresponding to the first set of operating conditions. Let n be the storm center movement speed corresponding to the nth working condition. For the k-th discrete time point; Output a scalar for the target at the corresponding time point. Let be the radial wind speed of the downburst at the k-th discrete time point.
[0012] Furthermore, S4 specifically includes: S401. Based on ensemble learning theory, the regression random forest algorithm is used as the basic architecture of the prediction model, and the Bootstrap sampling method is used to generate multiple sub-datasets from the training dataset, which are used to train each decision tree respectively. S402, Number of decision trees used and the maximum depth of the decision tree As key hyperparameters, a two-dimensional parameter grid was constructed by setting the range of values for the number of decision trees and the maximum depth of the decision trees. S403. The two-dimensional parameter grid is traversed using the parameter sensitivity analysis method. For each parameter combination in the grid, a corresponding regression random forest model is constructed and trained using the training dataset, and the determination coefficient under that parameter combination is calculated. As an evaluation index for prediction accuracy; S404. Selecting the coefficient of determination The parameter combination that reaches the maximum value is taken as the optimal hyperparameter combination. The regression random forest model is retrained using the optimal hyperparameter combination and the training dataset. The trained regression random forest model is then used as the downburst average wind speed prediction model.
[0013] Furthermore, S5 specifically includes: S501. Obtain the wind speed of the environment to be predicted for the predicted operating condition. and the predicted speed of the storm center's movement And determine the time series that need to be predicted. ; S502. Combine the wind speed to be predicted and the speed of the storm center to be predicted with each time point in the time series to construct an input feature matrix for the predicted operating condition; wherein, the formula for the input feature matrix is: ; in, The input feature matrix is the working condition to be predicted. For the predicted ambient wind speed, Let P be the speed at which the storm center is to be predicted, and P be the total number of prediction time steps. The time series The various points in time; S503. Input the input feature matrix into the downburst average wind speed prediction model. The model calculates the input feature matrix according to the mapping rules learned internally and outputs the predicted value of the downburst average wind speed at the corresponding time. S504. Arrange the multiple predicted values output by the downburst average wind speed prediction model in the order of the time series to reconstruct the downburst average wind speed time history curve under the predicted working condition.
[0014] This invention also provides a downburst average wind speed prediction system based on regression random forest. Based on the downburst average wind speed prediction method based on regression random forest described above, the system includes: A module was established to create a computational fluid dynamics numerical simulation model of a moving downburst flow based on the theory of impact jets. The generation module is used to determine the range of variation of ambient wind speed and storm center movement speed to construct a two-dimensional parameter space. The Haldane sequence sampling method is used to uniformly sample within the two-dimensional parameter space to generate parameter combination samples containing different ambient wind speeds and storm center movement speeds. The calculation module is used to input the parameter combination sample into the computational fluid dynamics numerical simulation model for calculation, and construct a training dataset containing multiple sets of input feature variables and target output variables based on the wind speed time history data of the monitoring points obtained from the calculation; wherein, the input feature variables include ambient wind speed, storm center movement speed and time point, and the target output variable is the downburst average wind speed at the corresponding time point; The training module is used to construct an initial regression random forest model and train it using the training dataset. During the training process, the parameter sensitivity analysis method is used to optimize the number of decision trees and the maximum depth of decision trees to obtain the optimal combination of hyperparameters. The trained regression random forest model is then used as the downburst average wind speed prediction model. The prediction module is used to obtain the ambient wind speed, storm center movement speed and time series to be predicted for the working condition to be predicted, input the downburst average wind speed prediction model, and obtain the downburst average wind speed prediction result for the corresponding time series under the working condition.
[0015] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described method.
[0016] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described method.
[0017] The beneficial effects of this invention are as follows: This invention addresses the challenges of high computational cost and long computation time in predicting moving downburst wind fields by constructing a surrogate model based on regression random forest to replace traditional computational fluid dynamics (CFD) numerical simulation. Uniform sampling within a two-dimensional parameter space using Halton sequences ensures the completeness and representativeness of the training samples. Furthermore, parameter sensitivity analysis is employed to optimize key hyperparameters, ensuring the model possesses optimal generalization ability. Experiments demonstrate that this method accurately captures the complex nonlinear characteristics of the wind speed time history of moving downbursts, especially the bimodal characteristic, exhibiting high prediction accuracy. Compared to traditional CFD simulation, computational efficiency is significantly improved, enabling real-time and rapid prediction of downburst wind fields under arbitrary combinations of environmental wind speed and moving speed. This provides efficient and accurate data support for the wind-resistant design of engineering structures such as power transmission lines. Attached Figure Description
[0018] Figure 1 This is a schematic diagram of a method flow according to an embodiment of the present invention.
[0019] Figure 2 This is a cross-sectional view of the computational domain for a static downburst in this invention.
[0020] Figure 3 This is a cross-sectional view of the center section of the grid division for a static downburst in this invention.
[0021] Figure 4 This is a schematic diagram of the device structure according to an embodiment of the present invention.
[0022] Figure 5 This is a schematic diagram of the internal structure of a computer device according to an embodiment of the present invention.
[0023] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation
[0024] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
[0025] like Figure 1 As shown, this invention provides a method for predicting the average wind speed of downbursts based on regression random forest, including: S1. Based on the theory of impact jets, a computational fluid dynamics (CFD) numerical simulation model of a moving downburst is established.
[0026] In one embodiment, step S1 mainly includes the following sub-steps: S101. Based on the theory of impact jets, a three-dimensional impact jet model is used as the basic wind field model to construct a three-dimensional cylindrical computational domain, such as... Figure 2The diagram shows a schematic cross-section of the computational domain. It can be seen that the descending airflow rushes directly from the inlet to the ground, and after contact with the ground, disperses radially, forming the maximum wind speed near the ground. The diameter of the computational domain is set to the jet diameter. 16 times ( The height is the jet diameter. 4 times ( A circular downward airflow inlet (jet inlet) is placed at the top center of the computational domain, with the inlet height above the ground being [missing information]. This setting is based on the initial height of the downburst observed by Doppler radar. For moving downbursts, the calculation domain is divided into a stationary region and a moving nozzle region, and the nozzle diameter is set accordingly. The length is 1000 m, and numerical simulations are performed using this value. A sliding mesh technique is used to handle the region interface, allowing the velocity inlet to follow a specific trajectory (e.g., a length of...). The area is moved to simulate the translational movement of the storm center.
[0027] S102. Discretize the computational domain using a hexahedral structured mesh, such as... Figure 3 As shown, to accurately capture wind speed changes in the near-surface region (the area where downburst maximum wind speeds typically occur), the near-surface region is densified vertically. The height of the first near-surface grid layer is set to [value missing]. To meet computational accuracy requirements, in moving simulations, an arbitrary coupled mesh interface (ACMI) technique is used to construct a sliding mesh interface, ensuring data transfer and flux interpolation between the stationary and moving regions at the interface.
[0028] S103. Based on the laws of fluid mass conservation and momentum conservation, the unsteady Reynolds-averaged Navier-Stokes (RANS) equations are selected as the fluid governing equations, specifically: (1) Mass conservation equation The law of conservation of fluid mass mainly describes the mathematical relationship between the increase in mass of a fluid element per unit time and the net inflow mass of the fluid element. Its mathematical expression is: ; in, t represents air density; t represents time. Let be the average velocity component in the i-direction (x, y, z directions); Represents a coordinate vector.
[0029] (2) Momentum conservation equation The law of conservation of momentum in fluids states that the rate of change of momentum of a fluid element is equal to the total external force acting on the element, essentially reflecting the application of Newton's second law of motion in fluid dynamics. Its mathematical expression is: ; Where P represents the average pressure. This represents the body force component acting on the infinitesimal element. This represents the viscous stress component.
[0030] (3) Reynolds-averaged Navier-Stokes (RANS) equations By introducing the average and fluctuating velocity components into the governing equations, the Reynolds equations can be obtained through ensemble averaging, as shown below: ; ; in, Let represent the Reynolds stress tensor. Due to the presence of Reynolds stress, the equations are not closed, requiring a turbulence model to close them. Based on the Boussinesq assumption, the Reynolds stress tensor in the turbulence model can be expressed as: ; in, Let k represent turbulent viscosity and k represent turbulent kinetic energy. This represents the mean strain rate tensor. The SST-k turbulence model is selected. A turbulence model is used to accurately analyze flow separation phenomena and near-wall flow characteristics. The solver and algorithm employ the finite volume method-based CFD software OpenFOAM, using the SIMPLIC algorithm to solve the pressure-velocity coupling problem. Unsteady terms in the discretization scheme are discretized using a second-order implicit scheme, while convection and diffusion terms are discretized using second-order precision. An adjustable time step is used to ensure stability during the computation.
[0031] S104. Based on the physical characteristics of the moving downburst, the boundary conditions of the computational fluid dynamics numerical simulation model are set as follows: Velocity inlet: The vertically downward velocity component is given as the jet inlet velocity. And according to the set storm center movement speed Control the translational movement of the entrance within the moving area.
[0032] Ambient wind field: Set velocity inlet boundary conditions in the direction of incoming flow in the computational domain, and input the ambient wind speed. The simulation simulates the effect of ambient wind on downbursts; the simulation is initialized using the logarithmic wind profile of the ambient boundary layer, and the movement of the downburst is driven by the ambient boundary layer wind profile.
[0033] Ground boundary: set to no-slip wall condition; the cylindrical surface at the top and entrance of the computational domain adopts slip wall boundary condition to ignore wall shear stress; the side of the computational domain is set as pressure outlet and the relative pressure is set to 0.
[0034] S2. Determine the range of variation of ambient wind speed and storm center movement speed to construct a two-dimensional parameter space. Use Haldane sequence sampling method to perform uniform sampling in the two-dimensional parameter space to generate parameter combination samples containing different ambient wind speeds and storm center movement speeds.
[0035] In one embodiment, step S2 mainly includes the following sub-steps: S201. Analyze the key factors affecting the wind field characteristics of moving downbursts and select ambient wind speeds. and the speed of movement at the center of the storm These are two independent key control parameters. Based on historical thunderstorm observation data and actual engineering application requirements, the numerical variation range of the above two parameters is determined as follows: Determine the ambient wind speed The range of change is Determine the speed of the storm center's movement. The range of change is The variation ranges of these two parameters together constitute the two-dimensional parameter space for downburst wind speed prediction.
[0036] S202. To obtain uniformly distributed and representative samples within the parameter space, a low-discrepancy sequence, namely the Halton sequence, is used for sampling. The Halton sequence, as an effective low-discrepancy sequence sampling method, can generate more uniformly distributed points in a multidimensional parameter space. The specific process is as follows: 1) For the nth sample point, its Halton sequence value under cardinality b The calculation is performed using the following steps: First, expand the index value n in radix b to obtain: ; in, It is the k-th digit of n in the base b. , where m is the highest-order digit. Then, based on the expansion coefficients... Given a cardinality b, calculate the value of the Halton sequence using the inverse function of free radicals. : ; This formula mirrors the numbers of the expansion of the base b of the integer n to the right of the decimal point, thus generating a uniformly distributed value in the interval [0,1).
[0037] 2) Regarding ambient wind speed and the speed of movement at the center of the storm For each of these two parameter dimensions, a different prime number is chosen as the base. In this embodiment, the prime number 2 is chosen as the ambient wind speed. cardinality of dimensions ( ); The prime number 3 is chosen as the speed at which the storm center moves. cardinality of dimensions ( Subsequently, according to the above formula (i.e., the expansion of n and...), The formula calculates the Halton sequence value for each parameter dimension, generating a two-dimensional dimensionless sample set containing N sample points. The generated series of points are uniformly distributed in the interval [0, 1].
[0038] S203. Using the inverse normalization (linear mapping) method, the dimensionless sample set H generated in S202 is mapped back to the actual physical parameter range determined in S201. For each sample point n (n=1, 2, ..., N), the specific calculation formula is as follows: ; ; in, The actual physical value of the ambient wind speed corresponding to the nth sample; This represents the actual physical value of the storm center's movement speed corresponding to the nth sample. The Halton sequence value with a base of 2 (corresponding to the environmental wind speed dimension) is calculated in step S202. The Halton sequence value with a base of 3 (corresponding to the storm movement speed dimension) is calculated in step S202. These represent the lower and upper limits of environmental wind speed variation, respectively. These represent the lower and upper limits of the change in the storm's center movement speed, respectively.
[0039] S204, calculate each pair of actual physical parameters obtained in S203. As a single working condition sample, a parameter combination sample set containing N sets of data is generated: ; Each sample in this sample set contains a unique combination of ambient wind speed and storm center movement speed, and the whole is uniformly distributed within a preset two-dimensional parameter space, which is used as the boundary condition input for subsequent CFD numerical simulation models.
[0040] S3. Input the parameter combination sample into the computational fluid dynamics numerical simulation model for calculation, and construct a training dataset containing multiple sets of input feature variables and target output variables based on the wind speed time history data of the monitoring points obtained from the calculation; wherein, the input feature variables include ambient wind speed, storm center movement speed and time point, and the target output variable is the downburst average wind speed at the corresponding time point.
[0041] In one embodiment, step S3 mainly includes the following sub-steps: S301. Using the batch processing script provided by OpenFOAM, the sample set S containing N sets of parameter combinations generated in step S2 is used as input conditions sequentially. For each set of samples in the sample set... ,in, These values are then assigned to the motion properties of the environmental wind speed inlet boundary and the moving region of the CFD model, respectively. The solver is then started to perform unsteady-state calculations, simulating the entire process of a downburst from its generation and ground contact to its spread and movement across the ground, until the calculation time reaches the preset total simulation duration.
[0042] S302. Monitoring points are pre-set in the CFD computational domain. In this embodiment, a location on the storm's movement path axis and at a certain height above the ground (e.g., the height of the first layer of mesh or the height of engineering concern) is selected as the monitoring point. During the simulation of each working condition, the velocity vector at this monitoring point is recorded in real time, and its horizontal radial component is extracted as the average wind speed of the downburst. After the calculation is completed, the complete wind speed time history curve of the monitoring point over time under each set of operating conditions will be output. .
[0043] S303. Discretize each wind speed time history curve extracted in S302 and set a uniform time sampling interval. (For example, 0.1s or the same as the CFD output step size), the continuous time t is discretized into M time points. .
[0044] Extract each time point Corresponding instantaneous wind speed value . Each time point The global control parameters (ambient wind speed) of this simulation (group n) and the speed of movement at the center of the storm By associating these features, we can construct the input feature vector at that moment. and the corresponding wind speed value As the target output scalar .
[0045] S304. Iterate through all N sets of working conditions and their corresponding M time points, summarizing all generated input feature vectors and target output scalars. Construct a training dataset D with a total sample size of N×M: ; The input feature variables include ambient wind speed, storm center movement speed, and time point; the target output variable is the average downburst wind speed at the corresponding time. This dataset covers the complete pattern of wind speed development over time under different ambient wind speeds and different movement speeds, and is used for training the subsequent regression random forest model.
[0046] S4. Construct an initial regression random forest model and train it using the training dataset. During training, parameter sensitivity analysis is used to optimize the number of decision trees and the maximum depth of the decision trees to obtain the optimal hyperparameter combination. The trained regression random forest model is then used as the downburst average wind speed prediction model. In one embodiment, step S4 mainly includes the following sub-steps: S401. Based on ensemble learning theory, a regression random forest algorithm is used as the basic architecture of the prediction model. This model consists of a large number of decision trees. Multiple subsets are generated from the training dataset D constructed in step S3 using the Bootstrap sampling method (sampling with replacement), and each subset is used to train a different decision tree. During node splitting, a feature subset is randomly selected for optimal splitting. Finally, the predicted value of the downburst's average wind speed is output by averaging the prediction results of all decision trees.
[0047] Before training the model using the training dataset D constructed in step S3, the dataset is first randomly divided. Dataset D is randomly divided into a training set and a test set in a 7:3 ratio. The training set is used for model building and parameter updates, while the test set is used to evaluate the model's generalization ability on unseen data.
[0048] S402. To improve the prediction accuracy of the model, the two key hyperparameters that have the greatest impact on the performance of the RFR model are selected for optimization, namely the number of decision trees. and the maximum depth of the decision tree And determine the optimal value range for these two parameters, i.e., the number of decision trees. The value range is set to [10, 200] (e.g., incrementing in steps of 10); the maximum depth of the decision tree. The value range is set to [1, 20] (for example, incrementing by 1 step). These two parameters constitute the two-dimensional parameter grid for model optimization.
[0049] S403. Use the parameter sensitivity analysis method to traverse the parameter grid set in S402, and for each set of parameter combinations in the grid... A corresponding regression random forest model is constructed and trained using the training dataset D, and the performance of the trained model is evaluated using a cross-validation set. To quantitatively evaluate the model performance under different parameter combinations, this embodiment uses the coefficient of determination. The root mean square error (RMSE) and mean absolute error (MAE) are used as evaluation criteria, with the coefficient of determination being... The model is used to characterize its ability to explain wind speed variability; the root mean square error (RMSE) measures the degree of deviation between predicted and actual wind speeds; and the mean absolute error (MAE) reflects the actual magnitude of the prediction error. During the traversal, each set of parameters is calculated. For the three metrics mentioned above on the test set, select the coefficient of determination. As an evaluation metric for model prediction accuracy, the calculation of this parameter combination... Value. Through iterative calculations, a value is generated that reflects... Value follows and Sensitivity analysis heatmaps or 3D surface plots of the changes. S404. Analyze the evaluation index results obtained from S403 and select the coefficient of determination. The parameter combination that reaches its maximum value and tends to stabilize is considered the optimal hyperparameter combination. That is, as the number of decision trees increases... As the value increases, the model prediction errors (RMSE and MAE) decrease rapidly. When the value exceeds 100, the error tends to converge to a stable level, increasing with the maximum depth of the decision tree. The increase of the coefficient of determination First rise rapidly, when When it reaches 12 o'clock, The value remained stable above 0.99, and both RMSE and MAE reached their minimum values. Based on these validation results, the optimal number of decision trees was determined to be 100. The maximum depth of the optimal decision tree is 12. Using a determined optimal combination of hyperparameters With the complete training dataset D, retrain the regression random forest model, and use the trained model as the final downburst average wind speed prediction model. S5. Obtain the environmental wind speed, storm center movement speed, and time series to be predicted for the working condition, input them into the downburst average wind speed prediction model, and obtain the downburst average wind speed prediction result for the corresponding time series under the working condition.
[0050] In one embodiment, step S5 mainly includes the following sub-steps: S501, based on actual engineering needs or a specific thunderstorm weather background, determine the values of two key control parameters for the predicted working condition, namely, the wind speed of the predicted environment. Predicted speed of storm center movement Simultaneously, the time span and time resolution to be predicted are determined, and the time series to be predicted is generated. , where P is the total number of prediction time steps. S502. Since the trained regression random forest model is trained based on a single-point input, the parameters obtained in S501 need to be combined into a feature matrix that conforms to the model input format. A fixed environmental wind speed... and the speed of movement at the center of the storm respectively with time series Each point in time Combine the samples to construct an input feature matrix containing P groups of samples. : ; Each row of this matrix corresponds to an input feature vector at a given time step.
[0051] S503, The input feature matrix constructed in S502 The input is fed into the final downburst average wind speed prediction model obtained in step S4 (i.e., the regression random forest model after parameter optimization). Based on the nonlinear mapping rules learned internally, the model processes each row of the input vector in the matrix in parallel or calculates row by row, and outputs the predicted value of the downburst average wind speed at the corresponding time.
[0052] S504. Arrange the P discrete predicted values output by the model in the order of the time series to reconstruct the time history curve of the mean wind speed of the downburst under the predicted condition. The prediction results can be directly used to evaluate the characteristics of downburst wind fields under a specific combination of wind speed and moving speed in the environment, or compared and verified with computational fluid dynamics (CFD) simulation results. To verify the actual prediction effect of the final model, a typical working condition in a test set (the data of this working condition was not used in model training) was selected for comparative verification, specifically: Set ambient wind speed Storm center movement speed The wind speed time history curves of the downburst under this working condition were calculated using both the prediction model of this invention and the CFD numerical simulation method, and the two curves were plotted in the same coordinate system for comparison. In terms of accuracy, the wind speed time history curves obtained by the prediction model and the CFD simulation curves are highly consistent in shape, especially in the peak region where the downburst wind speed rises sharply (bimodal characteristic). The predicted values and simulated values show extremely high agreement, indicating that the model has successfully learned the complex nonlinear time-varying characteristics of the moving downburst. Regarding efficiency, the average computation time for the CFD numerical simulation under this working condition is approximately 6.5 hours (based on a high-performance workstation), while the average computation time of the prediction model of this invention is only 0.00069 seconds. Experimental results show that the method proposed in this invention ensures prediction accuracy. Under the premise of improving computational efficiency by tens of millions of times, it enables real-time and rapid prediction of downburst wind fields.
[0053] like Figure 4 As shown, the present invention also provides a downburst average wind speed prediction system based on regression random forest. Based on the downburst average wind speed prediction method based on regression random forest described above, the system includes: Module 1 is established to create a computational fluid dynamics numerical simulation model of a moving downburst flow based on the impingement jet theory. Module 2 is used to determine the range of variation of ambient wind speed and storm center movement speed to construct a two-dimensional parameter space. The Haldane sequence sampling method is used to uniformly sample within the two-dimensional parameter space to generate parameter combination samples containing different ambient wind speeds and storm center movement speeds. Calculation module 3 is used to input the parameter combination sample into the computational fluid dynamics numerical simulation model for calculation, and construct a training dataset containing multiple sets of input feature variables and target output variables based on the wind speed time history data of the monitoring points obtained from the calculation; wherein, the input feature variables include ambient wind speed, storm center movement speed and time point, and the target output variable is the downburst average wind speed at the corresponding time point; Training module 4 is used to construct an initial regression random forest model and train it using the training dataset. During the training process, the parameter sensitivity analysis method is used to optimize the number of decision trees and the maximum depth of decision trees to obtain the optimal hyperparameter combination, and the trained regression random forest model is used as the downburst average wind speed prediction model. Prediction module 5 is used to obtain the environmental wind speed, storm center movement speed and time series to be predicted for the working condition to be predicted, input the downburst average wind speed prediction model, and obtain the downburst average wind speed prediction result for the corresponding time series under the working condition.
[0054] Each of the above modules is used to perform the corresponding steps in the downburst average wind speed prediction method based on regression random forest. The specific implementation method is as described in the above method embodiment, and will not be repeated here.
[0055] like Figure 5 As shown, the present invention also provides a computer device, which may be a server, and its internal structure may be as follows: Figure 5As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides the environment for the operation of the operating system and computer programs in the non-volatile storage media. The database stores all the data required for the process of the downburst average wind speed prediction method based on regressive random forest. The network interface is used for communication with external terminals via a network connection. The computer program is executed by the processor to implement the downburst average wind speed prediction method based on regressive random forest.
[0056] Those skilled in the art will understand that Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer equipment on which the present application is applied.
[0057] An embodiment of this application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements any of the above-described methods for predicting average wind speed of downbursts based on regressive random forests.
[0058] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by hardware related to computer program instructions. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the above method embodiments. Any references to memory, storage, databases, or other media provided in this application and used in the embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM), such as dynamic RAM (used as main storage) or static RAM (commonly used as cache memory). By way of illustration and not limitation, RAM has various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), and Rambus DRAM (RDRAM).
[0059] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.
[0060] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.
Claims
1. A method for predicting the average wind speed of downbursts based on regression random forest, characterized in that, include: S1. Based on the theory of impact jets, a computational fluid dynamics numerical simulation model of a moving downburst is established. S2. Determine the range of variation of ambient wind speed and storm center movement speed to construct a two-dimensional parameter space. Use Haldane sequence sampling method to perform uniform sampling in the two-dimensional parameter space to generate parameter combination samples containing different ambient wind speeds and storm center movement speeds. S3. Input the parameter combination sample into the computational fluid dynamics numerical simulation model for calculation, and construct a training dataset containing multiple sets of input feature variables and target output variables based on the wind speed time history data of the monitoring points obtained from the calculation; wherein, the input feature variables include ambient wind speed, storm center movement speed and time point, and the target output variable is the average wind speed of the downburst at the corresponding time point; S4. Construct an initial regression random forest model and train it using the training dataset. During the training process, use parameter sensitivity analysis to optimize the number of decision trees and the maximum depth of decision trees to obtain the optimal hyperparameter combination. The trained regression random forest model is then used as the downburst average wind speed prediction model. S5. Obtain the ambient wind speed, storm center movement speed, and time series to be predicted for the working condition to be predicted, input them into the downburst average wind speed prediction model, and obtain the downburst average wind speed prediction result for the corresponding time series under the working condition.
2. The method for predicting the average wind speed of downbursts based on regression random forest according to claim 1, characterized in that, S1 specifically includes: S101. A three-dimensional impact jet model is adopted to construct a three-dimensional cylindrical computational domain, and a jet inlet is set at the top of the computational domain. For moving downbursts, the entire computational domain is divided into a stationary region and a moving nozzle region, and the interface of the regions is processed using sliding mesh technology. S102. The computational domain is discretized using a hexahedral structured mesh, and the mesh is refined in the near-ground region. A sliding mesh interface between the stationary region and the moving nozzle region is constructed using an arbitrary coupled mesh interface. S103. The unsteady Reynolds-averaged Navier-Stokes equations are adopted as the fluid control equations, and SST-k is used. Turbulence models are used to analyze flow separation phenomena, and the SIMPLIC algorithm is used to solve pressure-velocity coupling problems. S104. Set the velocity components and moving velocity of the jet inlet, set the ambient wind speed inlet, set the ground no-slip wall condition, and set the boundary conditions of the top and sides of the computational domain to obtain the computational fluid dynamics numerical simulation model of the moving downburst.
3. The method for predicting the average wind speed of downbursts based on regression random forest according to claim 1, characterized in that, S2 specifically includes: S201, Select ambient wind speed and the speed of movement at the center of the storm As key control parameters, the variation range of ambient wind speed was determined. and the range of change in the speed of the storm center's movement ; S202. Select two distinct prime numbers as the bases for the environmental wind speed dimension and the storm center movement speed dimension, respectively. Use the Halton sequence generation algorithm to calculate the Halton sequence value of each sample point in the two dimensions, and generate N uniformly distributed dimensionless sample points. S203. Using the inverse normalization method, the dimensionless sample points are mapped back to the range of environmental wind speed change and the range of storm center movement speed change, and the corresponding actual physical parameter values are calculated. S204. Treat each set of actual physical parameter values obtained by calculation as an independent working condition sample, and summarize them to generate a parameter combination sample set for driving computational fluid dynamics numerical simulation.
4. The method for predicting the average wind speed of downbursts based on regression random forest according to claim 3, characterized in that, In S202, the calculation process of the Halton sequence value is as follows: For the nth sample point, first expand its index value n in cardinality b: ； Then, the Haldane sequence value of this sample point at cardinality b is calculated. : ； in, The k-th digit of n in base b ; m is the highest digit; the base of the environmental wind speed dimension is 2, and the corresponding Halton sequence value is The dimension of the storm center's movement speed is selected with a base of 3, and the corresponding Halton sequence value is... ; In S203, the formula for calculating the actual physical parameter values is as follows: ； ； in, Let n be the ambient wind speed corresponding to the nth sample. The speed at which the storm center moves corresponds to the nth sample. and These are the minimum and maximum values of the ambient wind speed, respectively. and These are the minimum and maximum values of the storm's center movement speed, respectively.
5. The method for predicting average wind speed of downbursts based on regression random forest according to claim 3, characterized in that, S3 specifically includes: S301. The parameter combination sample set is used as input conditions in sequence and assigned to the motion properties of the environmental wind speed inlet boundary and the moving region of the computational fluid dynamics model respectively. The solver is started to perform unsteady-state calculations until the preset total simulation time is reached. S302. Pre-set monitoring points in the computational domain. During the simulation of each set of operating conditions, record the velocity vector at the monitoring points in real time, extract its horizontal radial component as the average wind speed of the downburst, and output the wind speed time history curve of the monitoring point over time for each set of operating conditions; wherein, the monitoring point is located on the axis of the storm's movement path, and the average wind speed of the downburst is the radial velocity in the downburst wind field. Curve of change with time t ; S303. Discretize the wind speed time history curve, divide the continuous time into multiple time points, extract the instantaneous wind speed value corresponding to each time point, associate each time point with the simulated environmental wind speed and storm center movement speed, construct an input feature vector, and use the corresponding instantaneous wind speed value as the target output scalar. S304. Iterate through all working conditions and their corresponding time points, summarize all generated input feature vectors and target output scalars, and construct a training dataset containing multiple sets of input feature variables and target output variables.
6. The method for predicting the average wind speed of downbursts based on regression random forest according to claim 5, characterized in that, In step S303, the input feature vector is: ； The target output scalar is: ; in, This is the input feature vector at the k-th time point under the n-th working condition. Let $\frac{ ... Let n be the storm center movement speed corresponding to the nth working condition. For the k-th discrete time point; Output a scalar for the target at the corresponding time point. Let be the radial wind speed of the downburst at the k-th discrete time point.
7. The method for predicting average wind speed of downbursts based on regression random forest according to claim 1, characterized in that, S4 specifically includes: S401. Based on ensemble learning theory, the regression random forest algorithm is used as the basic architecture of the prediction model, and the Bootstrap sampling method is used to generate multiple sub-datasets from the training dataset, which are used to train each decision tree respectively. S402, Number of decision trees used and the maximum depth of the decision tree As key hyperparameters, a two-dimensional parameter grid was constructed by setting the range of values for the number of decision trees and the maximum depth of the decision trees. S403. The two-dimensional parameter grid is traversed using the parameter sensitivity analysis method. For each parameter combination in the grid, a corresponding regression random forest model is constructed and trained using the training dataset, and the determination coefficient under that parameter combination is calculated. As an evaluation index for prediction accuracy; S404. Selecting the coefficient of determination The parameter combination that reaches the maximum value is taken as the optimal hyperparameter combination. The regression random forest model is retrained using the optimal hyperparameter combination and the training dataset. The trained regression random forest model is then used as the downburst average wind speed prediction model.
8. The method for predicting the average wind speed of downbursts based on regression random forest according to claim 1, characterized in that, S5 specifically includes: S501. Obtain the wind speed of the environment to be predicted for the predicted operating condition. and the predicted speed of the storm center's movement And determine the time series that need to be predicted. ; S502. Combine the wind speed to be predicted and the speed of the storm center to be predicted with each time point in the time series to construct an input feature matrix for the predicted operating condition; wherein, the formula for the input feature matrix is: ； in, The input feature matrix is the working condition to be predicted. For the predicted ambient wind speed, Let P be the speed at which the storm center is to be predicted, and P be the total number of prediction time steps. The time series The various points in time; S503. Input the input feature matrix into the downburst average wind speed prediction model. The model calculates the input feature matrix according to the mapping rules learned internally and outputs the predicted value of the downburst average wind speed at the corresponding time. S504. Arrange the multiple predicted values output by the downburst average wind speed prediction model in the order of the time series to reconstruct the downburst average wind speed time history curve under the predicted working condition.
9. A downburst average wind speed prediction system based on regression random forest, based on the downburst average wind speed prediction method based on regression random forest as described in any one of claims 1 to 8, characterized in that, The system includes: A module was established to create a computational fluid dynamics numerical simulation model of a moving downburst flow based on the theory of impact jets. The generation module is used to determine the range of variation of ambient wind speed and storm center movement speed to construct a two-dimensional parameter space. The Haldane sequence sampling method is used to uniformly sample within the two-dimensional parameter space to generate parameter combination samples containing different ambient wind speeds and storm center movement speeds. The calculation module is used to input the parameter combination sample into the computational fluid dynamics numerical simulation model for calculation, and construct a training dataset containing multiple sets of input feature variables and target output variables based on the wind speed time history data of the monitoring points obtained from the calculation; wherein, the input feature variables include ambient wind speed, storm center movement speed and time point, and the target output variable is the downburst average wind speed at the corresponding time point; The training module is used to construct an initial regression random forest model and train it using the training dataset. During the training process, the parameter sensitivity analysis method is used to optimize the number of decision trees and the maximum depth of decision trees to obtain the optimal combination of hyperparameters. The trained regression random forest model is then used as the downburst average wind speed prediction model. The prediction module is used to obtain the ambient wind speed, storm center movement speed and time series to be predicted for the working condition to be predicted, input the downburst average wind speed prediction model, and obtain the downburst average wind speed prediction result for the corresponding time series under the working condition.
10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 8.