A long-period typhoon wind and rain field return period estimation method based on random forest

Abstract

This invention discloses a method for estimating the return period of long-period typhoon wind and rain fields based on random forests. It aims to address the shortcomings of existing typhoon wind and rain field simulation technologies, such as the difficulty in balancing computational complexity, data accuracy, and data dependency, and the lack of a lightweight prediction framework suitable for the massive typhoon scenario extrapolation requirements in long-term engineering risk assessments. The invention includes steps such as data collection, data preprocessing, model building, model training, long-period simulation, and wind and rain estimation during the return period. By constructing rainfall and wind speed models based on random forest regression models and combining them with extreme value statistical fitting methods, it achieves efficient estimation of the return period curves of wind speed and rainfall in the target area, meeting the needs of long-term engineering risk assessments and urban resilience analysis.

Description

Technical Field

[0001] This invention relates to the field of typhoon disaster prediction, and more specifically, to a method for estimating the return period of long-period typhoon wind and rain fields based on random forests. Background Technology

[0002] Currently, one of the mainstream methods for simulating typhoon wind and rain fields is numerical weather prediction (NWP) models based on fluid dynamics and physical mechanisms. This technology can accurately reproduce the three-dimensional structure of typhoons and their wind and rain evolution process, and has a clear physical basis. However, NWP models not only have extremely high requirements for the quality and resolution of various input data such as initial fields, boundary conditions, topography, and underlying surfaces, but also involve complex physical parameterization schemes and high-resolution spatial discretization, resulting in extremely large computational resource consumption. A single simulation usually takes several hours to several days. This extremely low time efficiency is difficult to meet the practical needs of rapid, batch disaster field simulation of a large number of typhoon scenarios in long-term engineering problems.

[0003] Another common approach is the parametric wind / rainfall model. While these models are computationally fast, they are often based on idealized symmetry or semi-symmetry assumptions, resulting in overly simplistic model structures. In practical applications, parametric models neglect the interference of complex terrain undulations and urban underlying surface roughness on the local characteristics of wind and rain fields, and fail to fully utilize the massive amounts of historical observational data from real meteorological stations for correction and calibration. This leads to significant errors in simulating hourly wind speeds and rainfall at specific stations, failing to accurately reproduce the complex and detailed local wind and rain characteristics during typhoons.

[0004] In recent years, machine learning technology has provided a new technical approach for typhoon wind and rain field simulation. It can automatically learn the complex nonlinear mapping relationships of wind and rain fields from massive historical observation and reanalysis data, constructing data-driven typhoon wind and rain field simulation models. These models maintain high prediction accuracy while improving computational efficiency by 2-3 orders of magnitude compared to traditional numerical models, achieving wind and rain field predictions at the second or even millisecond level, thus effectively meeting the timeliness requirements of batch simulation tasks in long-cycle engineering projects. Representative technologies include: typhoon wind and rain field simulation models based on Convolutional Neural Networks (CNNs), achieving high-precision wind and rain field reconstruction through end-to-end learning; typhoon wind and rain generation models based on Generative Adversarial Networks (GANs) and Diffusion Models, capable of learning the statistical distribution of historical typhoon wind and rain fields to generate diverse wind and rain scenarios that conform to physical laws; and typhoon wind and rain field simulation models based on Physics-Informed Neural Networks (PINNs), which, by embedding control equation constraints, ensure that the prediction results satisfy the conservation laws of mass, momentum, and other physical properties.

[0005] However, existing data-driven typhoon wind and rain simulation models still have significant limitations. From an application perspective, these models primarily serve for rapid forecasting of single typhoon events, with limited application for estimating wind and rain fields over long-term engineering projects. From a data perspective, existing models generally rely heavily on high-resolution reanalysis grid data, making it difficult to directly utilize limited historical meteorological station observation data within the target area for lightweight training. In summary, there is currently a lack of a lightweight prediction framework that can rapidly and in batches generate high-precision wind and rain fields by simply inputting basic typhoon parameters (such as typhoon path and central pressure), thus filling the technological gap for long-term engineering needs. Summary of the Invention

[0006] To address the challenge of achieving an effective balance between computational complexity, data accuracy, and data dependency in existing technologies, this invention provides a method for estimating the return period of long-period typhoon wind and rain fields based on random forests. This method achieves return period estimation accuracy that meets engineering application requirements with relatively low computational overhead and limited data input. This effectively supports the batch and continuous simulation of hundreds or thousands of typhoon events in long-term engineering risk assessments and urban resilience analyses, while also considering the timeliness and convenience requirements in practical applications.

[0007] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: A method for estimating the return period of long-period typhoon wind and rain fields based on random forests includes the following steps: Data collection includes collecting static geographic data from meteorological observation stations in the target area and dynamic historical disaster site data during typhoon disasters; Data preprocessing involves extracting the geographical features of meteorological observation stations, extracting the hourly typhoon center pressure features and the distance features from meteorological observation stations to the typhoon center for each typhoon event, and using the hourly maximum wind speed and rainfall of each meteorological observation station in each typhoon event as feature labels. Model construction: Based on the random forest regression model, a rainfall model and a wind speed model are constructed respectively. The geographical characteristics of the meteorological observation station, the hourly typhoon center pressure are used as input features and the distance from the meteorological observation station to the typhoon center are used as output features. The maximum wind speed and rainfall of the meteorological observation station are used as output features. Model training involves dividing the samples obtained from data preprocessing into training and testing sets, and training the rainfall model and wind speed model separately until the model overcomes the overfitting defect. Long-term simulation, based on the simulated typhoon event set, uses the trained wind speed model and rainfall model to generate typhoon wind and rain field for each grid cell in the target area, and obtains the wind and rain distribution of each grid cell in the target area; For wind and rain estimation during the return period, based on the wind and rain distribution of each grid cell, the cumulative probability density function of wind and rain intensity for each grid cell is constructed by fitting the extreme value statistical fitting method, and the return period curves of wind speed and rainfall for each grid cell are output.

[0008] As a preferred option, static geographic data includes digital elevation model (DEM) and coastline data; dynamic historical disaster sites mainly include historical typhoon trajectories and measured meteorological data from various meteorological stations within the target area during each typhoon.

[0009] Preferably, geographical features include the elevation, slope and aspect of the observation point, and the shortest straight-line distance from the observation point to the coastline.

[0010] Preferably, during model training, the samples in the training set are exponentially weighted, corresponding to either the wind speed model or the rainfall model, as expressed by:

[0011] in, Let i be the training weights for the i-th sample. Adjusting hyperparameters that focus on extreme values. Indicates the first The actual wind speed or rainfall of each sample and This represents the minimum and maximum values ​​of the actual wind speed or rainfall in the training set.

[0012] As a preferred approach, the indicators for a model to overcome overfitting are that the mean squared error (MSE) and mean absolute error (MAE) of the training and test sets are less than a preset threshold.

[0013] As a preferred approach, feature importance assessment can be used to improve the interpretability of wind speed and rainfall models.

[0014] As a preferred option, a spatiotemporal distribution test of the disaster site is conducted: using the test set samples as input, the rainfall and wind speed data output by the rainfall model and wind speed model are obtained; the degree of agreement between the rainfall and wind speed data and the actual disaster-affected areas and occurrence times is compared; when the degree of agreement is higher than the threshold, it proves that the rainfall model and wind speed model have overcome the overfitting defect.

[0015] As a preferred option, the expression for the cumulative probability density function is:

[0016] Indicates the recurrence period. This indicates the average annual frequency of typhoons affecting the target area. This indicates that the wind speed or rainfall exceeded the threshold during the typhoon period. The probability of.

[0017] As a preferred method, the extreme value statistical fitting method includes one of the empirical distribution, the Gumbel distribution, and the Weibull distribution.

[0018] As a preferred option, based on a given return period By obtaining the wind speed and rainfall corresponding to the return period of the grid cell through the return period curve, a wind and rain field distribution map of the target area is generated based on the wind speed and rainfall of the grid cell.

[0019] Compared with the prior art, the beneficial effects of the present invention are: (1) Lowering the data threshold and improving adaptability: Get rid of the heavy dependence on massive, multi-dimensional reanalysis grid data, realize lightweight training by directly using limited historical meteorological station observation data in the target area, and reduce the localization migration cost of the model.

[0020] (2) Balancing regional prediction accuracy and computational efficiency: Overcoming the shortcomings of overly idealistic parametric empirical models, using real observation data to restore local wind and rain characteristics; and significantly improving computational efficiency compared to traditional numerical models.

[0021] (3) Enables convenient generation of massive disaster scenarios: Simplifies model input elements, addresses the limitation of existing models that only serve a single short-term forecast, and meets the timeliness and convenience requirements of batch and continuous simulation of hundreds or thousands of typhoon events in long-term engineering risk assessment and urban resilience analysis. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the method of the present invention; Figure 2 This is a schematic diagram of the random forest model structure of this invention; Figure 3 This is a graph showing the results of the feature importance assessment of the present invention; Figure 4 This is a graph showing the maximum wind speed and rainfall of a certain cell grid and their recurrence year obtained in the verification example of this invention; Figure 5 This is a schematic diagram of the device of the present invention. Detailed Implementation

[0023] The present disclosure will be further described below with reference to the accompanying drawings and embodiments.

[0024] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0025] Example 1: A method for estimating the return period of long-period typhoon wind and rain fields based on random forests, such as... Figure 1 As shown, it includes the following steps: Data collection includes collecting static geographic data and dynamic historical disaster site data; Data preprocessing involves extracting geographical features, typhoon characteristics, and labeling specific locations with maximum wind speed and rainfall at corresponding times. Model building: Constructing a random forest regression model; Model training and evaluation, through the partitioning of the dataset, the weighting of samples, and the evaluation of the model, achieve a model without fitting defects; Long-term simulation, through artificially synthesized typhoon samples, enables the generation of typhoon winds and rain in each grid cell within the target area; The return period assessment of wind and rain quantifies the return period of wind and rain by using the cumulative probability density function and the wind speed-rainfall return period curve.

[0026] The following is a detailed description of each step and stage: Data collection involves gathering static geographic data and dynamic historical disaster field data related to typhoon disasters in the target area. In some embodiments, static geographic data includes digital elevation models and coastline data; dynamic historical disaster fields include historical typhoon trajectories and measured meteorological data from various meteorological stations within the target area during each typhoon.

[0027] Based on static geographic data, geographic data such as elevation, slope, and aspect of a specific location within the target area can be obtained. Based on coastline data, the shortest straight-line distance between the specific location and the coastline can be obtained. Based on dynamic historical disaster site data, the hourly typhoon central pressure and the shortest straight-line distance from the typhoon center to the target location can be obtained.

[0028] In the data preprocessing stage, the geographical features of meteorological observation stations are extracted, and the hourly central pressure and distance from the meteorological observation station to the typhoon center for each typhoon event are extracted. The hourly wind speed and rainfall data of each meteorological observation station in each typhoon event are used as feature labels.

[0029] In the data preprocessing stage, data cleaning and preprocessing are performed to extract specific data from the aforementioned static geographic data and dynamic historical disaster site data. Specifically, geographic data such as elevation, slope, and aspect of specific locations in the target area are extracted from the static geographic data, and the shortest straight-line distance between the specific location and the coastline is extracted from the coastline data. Hourly typhoon center pressure and the shortest straight-line distance from the typhoon center to the specific location are extracted from the dynamic historical disaster site data.

[0030] In this phase, this embodiment takes data from each observation point during each typhoon at a specific time interval resolution: Rainfall and wind speed data are extracted as feature labels. Geographic information of the observation points, including elevation, slope, and aspect, as well as the shortest straight-line distance from the meteorological observation station to the coastline, is extracted. Hourly central pressure and the distance from the meteorological observation station to the typhoon eye for each typhoon event are extracted. The data are then processed using normalization techniques to construct feature vectors. The specific normalization method is common knowledge in the field and will not be elaborated here.

[0031] In some embodiments, the geographical features of the observation point are also extracted, specifically, the elevation, slope and aspect features of the observation point, as well as the shortest straight-line distance from the observation point to the coastline are extracted.

[0032] The resolution is set to 1 hour, meaning that rainfall and wind speed data, typhoon center pressure data, and straight-line distance from the typhoon center observation point are collected at hourly intervals.

[0033] Through the data preprocessing stage, this embodiment obtains the input features and labels of each meteorological observation station during each typhoon, thereby constructing a sample set.

[0034] Model construction: Based on the random forest regression model, a rainfall model and a wind speed model were constructed respectively. The geographical information of the observation points and the central pressure characteristics were used as input features, and the hourly wind speed and rainfall were used as output features.

[0035] like Figure 2 As shown, the random forest regression model is a machine learning algorithm that integrates multiple decision trees. It constructs a training set through Bootstrap random sampling, randomly selects a subset of features to train each tree, and finally obtains the result by averaging the prediction values ​​of each tree. It uses the diversity of the "forest" to reduce variance and achieve high-precision regression prediction. By combining the prediction results of each tree, it has the characteristics of strong anti-overfitting ability and good robustness, and is suitable for complex nonlinear regression problems.

[0036] The sample set is divided into a training set and a test set. In this embodiment, the ratio is 9:1, that is, the training set accounts for 90% and the test set accounts for 10%. A rainfall model and a wind speed model are trained separately, with the same input. The output of the rainfall model is the rainfall amount of the grid cells in the target area at the current moment, and the output of the wind speed model is the wind speed of the grid cells in the target area at the current moment.

[0037] Considering that the purpose of this embodiment is mainly to focus on extreme wind and rain loads during typhoons, and that extreme wind and rain loads often occur under extreme rainfall and wind speeds, this embodiment adjusts the weights of the sample set during the model training phase, exponentially amplifying the weights of samples with high rainfall and wind speeds to prevent them from being submerged in the regular samples and causing the model to overfit.

[0038] The expression is:

[0039] in, Let i be the training weights for the i-th sample. The hyperparameter for adjusting the focus on extreme values ​​should be a number greater than a certain threshold. Indicates the first The actual wind speed or rainfall of each sample and This represents the minimum and maximum values ​​of the actual wind speed or rainfall in the training set.

[0040] According to this expression, the higher the value of rainfall or wind speed, the greater the weight of the sample, and thus the greater the contribution of these sample individuals during the model training process.

[0041] After training the model, it is evaluated using a test set. Evaluation criteria include a mean squared error (MSE) and mean absolute error (MAE) of both the training and test sets being less than preset thresholds, and the set Pearson correlation coefficient meeting preset benchmarks. In this embodiment, feature importance assessment is used to improve the interpretability of the wind speed and rainfall models, determining the contribution of each input feature to wind and rain prediction. Feature importance assessment identifies the key factors with the greatest impact on typhoon wind and rain, such as typhoon center pressure, observation point location, and historical wind speed, helping to understand the model's decision-making logic, optimize feature selection, and improve model reliability and prediction accuracy.

[0042] In some other possible embodiments, the method also includes performing a disaster site spatiotemporal distribution check: The target area is divided into grid cells, and geographical features and hourly typhoon center pressure and distance to the typhoon center are extracted as inputs to obtain rainfall and wind speed data output by the rainfall model and wind speed model. To verify the model output from a meteorological consistency perspective: the wind speed field should evolve continuously with the movement of the typhoon, with high-speed areas mainly distributed near the typhoon center and gradually weakening with increasing distance; the rainfall field should correspond to the typhoon's path, with areas of heavy rainfall continuously migrating with the typhoon. By analyzing the temporal continuity, spatial smoothness, and distance attenuation patterns of the wind and rain fields, and detecting the existence of isolated anomalous high values, the rationality and accuracy of the model results in the spatiotemporal dimensions can be verified. The accuracy of the model results is quantified by comparing the degree of agreement between rainfall and wind speed data and the actual affected areas and time periods.

[0043] Since the sample set originates from various meteorological observation stations, the implementation plan needs to extend the model's application scope to the entire target area during subsequent use. Therefore, using the trained rainfall and wind speed models, and taking one or more historical typhoon events as input, typhoon wind and rain data are generated for the target area to obtain wind and rain information for each grid cell. Then, the rainfall and wind speed data from the grid cells are filtered to obtain extreme values, and the geographical locations of these extreme values ​​are compared with the actual affected areas in the typhoon event to determine their overlap. When the overlap reaches a preset threshold, it indicates that the model can be extended to the entire target area and has good generalization ability.

[0044] The grid cell refers to the smallest resolution unit capable of resolving the required return period of wind and rain. Specifically, within the same target unit, the errors in rainfall and wind speed values ​​can be ignored. In some embodiments, the grid cell value is... To simplify calculations, the center point of each grid cell is used as a representative point to calculate the rainfall and wind speed at that point. The data from the aforementioned observation points can represent the data of their respective grid cells; it is not necessary to divide the target area into grids based on the specific location of the observation points to ensure they fall at the grid's center point.

[0045] During the long-term simulation phase, based on the simulated typhoon event set, the trained rainfall and wind speed models are used to generate typhoon wind and rain information for the target area, thereby obtaining wind and rain information for each grid cell in the target area.

[0046] The typhoon event set is a sample collection of multiple typhoon events generated based on existing typhoon simulation models. The samples include the typhoon's size, direction, intensity, and duration. Based on this typhoon sample collection as dynamic historical disaster data and the geographic information of the target area, rainfall and wind speed data under different typhoon events in the target area are simulated, thereby obtaining wind and rain information for each grid cell. The typhoon simulation model is existing technology, using a typhoon full-path simulation method based on conditional generative adversarial networks disclosed in CN117195724A to simulate and generate typhoons, establish a typhoon model, obtain its intensity evolution and typhoon path based on the typhoon model, and construct the sample set through multiple generation processes. By simulating typhoon samples over many years, a rich set of typhoon samples closely resembling the real environment can be constructed, providing a foundation for subsequent return-time wind and rain estimation. The generation method is a conventional technique in the field; the aforementioned patent only indicates the existence of this method and does not limit its specific use in this application.

[0047] Using typhoon event sets and geographic information as input, rainfall and wind speed data for each grid cell are obtained. Then, based on the rainfall and wind speed data from each grid cell, a rainfall and wind speed field for the target area is generated. At this point, this embodiment already possesses the rainfall and wind speed fields for the target area under different typhoon events, providing a foundation for subsequent return-time wind and rain estimation.

[0048] During the wind and rain estimation phase of the recurrence period, based on the typhoon event set and the wind and rain information of each grid unit, the cumulative probability density function of wind and rain intensity for each grid unit is constructed by using the extreme value statistical fitting method, and the recurrence period curves of wind speed and rainfall for each grid unit are output.

[0049] The expression for the cumulative probability density function is:

[0050] Indicates the recurrence period. This indicates the average annual frequency of typhoons affecting the target area. This indicates that the wind speed or rainfall exceeded the threshold during the typhoon period. The probability of.

[0051] Extreme value statistical fitting methods include one of the following: empirical distribution, Gumbel distribution, and Weibull distribution. Empirical distributions are directly constructed based on the frequency distribution of sample data, intuitively reflecting actual observation results; the Gumbel distribution is often used to describe the probability distribution of maxima and is suitable for extreme event analysis; the Weibull distribution can flexibly fit different types of data distributions and can be used to describe the statistical characteristics of natural phenomena such as wind speed. The core principle of these methods is to fit the probability distribution of extreme events using mathematical models. The basic steps include: collecting historical typhoon wind and rain data for the target area; selecting extreme value samples (such as annual maximum values); choosing a suitable distribution model; determining model parameters through methods such as maximum likelihood estimation; verifying the model's fitting effect; and calculating wind and rain intensity values ​​for different return periods based on the fitted probability distribution.

[0052] After obtaining the return period curves of each grid cell, in order to indicate the long-term engineering design of the region, the return period T can be fixed according to the actual disaster resistance needs of the local area. Based on the aforementioned cumulative probability density function, the rainfall and wind speed of each grid cell under the return period can be calculated. Then, each grid cell can be stretched into a wind and rain field distribution map of the target area, which can indicate the areas that need to be strengthened for disaster prevention, and realize disaster prevention projects or regional disaster prevention systems that can withstand a 100-year or 500-year return period.

[0053] Verification example: Using Province Z as the target area, the following static geographic data and dynamic historical disaster site data were collected: Static geographic data: Acquire digital elevation model (DEM) data of the target area with a resolution of 30m×30m, as well as the latest coastline vector data provided by the National Geomatics Center of China.

[0054] Dynamic historical disaster data: We collected a dataset of historical typhoon best paths for the target area from 2012 to 2022 (including typhoon center latitude and longitude and minimum central pressure at 3 or 6-hour time resolution, covering a total of 22 typhoon events, ensuring that the typhoon paths were within 200 kilometers of the target area boundary). Simultaneously, we collected hourly meteorological data from 2000 national standard surface meteorological stations in the area during the aforementioned typhoon events.

[0055] Using GIS spatial analysis tools and Python data processing libraries, features and labels are extracted: Geographical environmental feature extraction: Using ArcGIS software, slope and aspect rasters were generated based on the DEM data. The elevation, slope, and aspect of the 2000 weather stations were extracted. The Euclidean distance tool was used to calculate the straight-line distance from each weather station to the nearest coastline.

[0056] Typhoon dynamic feature extraction: Since the original typhoon track time resolution typically varies from 1 to 6 hours, linear interpolation is used to uniformly interpolate and refine it to a 1-hour time resolution. The shortest straight-line distance from the typhoon center to each meteorological observation station at each time point is calculated based on the spherical distance (Haversine formula).

[0057] Actual wind and rain label values ​​assigned: The observation data from 2000 meteorological stations were cleaned, and typhoon event samples with a missing rate of more than 10% were removed. For the remaining valid samples, the "maximum 10-minute average wind speed at 10 meters above ground (hereinafter referred to as wind speed)" and "hourly rainfall (hereinafter referred to as rainfall)" of each meteorological station during each typhoon were extracted as real labels for supervised training of machine learning models.

[0058] This embodiment uses the scikit-learn machine learning library in Python to build a Random Forest regression model. Since the physical formation mechanisms of wind speed and rainfall differ, this embodiment constructs a Random Forest wind speed model (hereinafter referred to as the wind speed model) and a Random Forest rainfall model (hereinafter referred to as the rainfall model) separately.

[0059] Input features (X): There are 6 items in total, including dynamic factors (shortest straight-line distance from the station to the typhoon center, current typhoon center pressure) and static factors (elevation, slope, aspect, and shortest straight-line distance from the coastline) of the target location (meteorological observation station or grid unit center point).

[0060] Output (Y): Wind speed (or rainfall) at the target point (meteorological observation station or grid cell center point) at the current time.

[0061] Model parameters: To balance the model's generalization ability and prevent overfitting during training, in some embodiments, the main hyperparameters of the random forest regression model are configured as follows: The number of decision trees (n_estimators) is set to 200 to ensure the model's fitting stability when the feature dimensions are small; The maximum depth of the tree (max_depth) is limited to 10 to control model complexity; Set the minimum number of samples required for internal node re-split (min_samples_split) to 15, and the minimum number of samples required for leaf node (min_samples_leaf) to 10; To make full use of existing features, the maximum feature ratio (max_features) considered when finding the best segmentation is set to 0.5; Simultaneously enable sampling with replacement (bootstrap=True) and set the maximum sampling ratio (max_samples) to 0.8.

[0062] Model training and evaluation include sample partitioning, exponentially weighted training, and model evaluation.

[0063] Sample splitting: The effective sample set of 22 typhoon events (number of typhoon events × number of meteorological stations × duration of impact) was randomly divided into training set and test set in a ratio of 9:1.

[0064] Exponentially weighted training: To enhance the model's ability to capture extreme wind and rain loads, a sample weight matrix is ​​introduced during the model training phase. The weight of each sample is calculated using the weighting formula provided by this invention:

[0065] In this embodiment, for the wind speed model, an extreme value attention hyperparameter is set. If the true wind speed of a certain sample Close to the maximum value Its weight Approaching (Approximately 1024); if close to the minimum value, the weight is close to 1. For the rainfall model, considering the greater uncertainty of rainfall extreme values, the hyperparameter of extreme value attention is set to alpha=6. If the actual rainfall y_i of a certain sample is close to the maximum value... Its weight Approaching (Approximately 64).

[0066] The calculated weight array was used as the sample weight parameters and fed into the random forest algorithm for model training. The training hardware environment consisted of a computing system equipped with an Intel Core i7-13700K CPU and an NVIDIA GeForce RTX 4070 Ti GPU. The wind speed model and the rainfall model had similar training times, averaging 3 minutes, demonstrating the model's advantage in deployment efficiency.

[0067] Model Evaluation: The trained wind speed and rainfall models are used to predict on the training and test sets to evaluate the model training effect. The evaluation results are as follows: Figure 2 As shown. The evaluation method includes calculating the mean squared error (MSE) and the mean absolute error (MAE). In this embodiment, the error rate difference between the MSE and MAE of the wind speed model and the test set is within 5%, indicating that overfitting has not occurred.

[0068] The model performance was also evaluated by calculating the Pearson correlation coefficient (R) on the test set. In this embodiment, the R values ​​for the wind speed model and the precipitation model on the test set were 0.66 and 0.40, respectively. The wind speed model's R value was greater than 0.6, demonstrating good predictive correlation. Given the inherent strong physical uncertainties and spatial heterogeneity of precipitation processes, the precipitation model's R value was greater than 0.3, indicating that its predictions had a moderate degree of correlation, thus verifying the effectiveness and engineering practical value of the model established in this embodiment.

[0069] The importance of features output by the model, such as Figure 3 The display shows that, Figure 3 The subplot 'a' shows the characteristic importance of the wind speed model. The cumulative importance of "elevation" and "distance from the typhoon center" in the wind speed model exceeds 60%, indicating that topographic conditions and relative distance are the main controlling factors in wind field evolution. Figure 3 Subplot b shows the feature importance of the rainfall model. The cumulative importance of "distance from the typhoon center" and "typhoon central pressure" in the rainfall model exceeds 60%, reflecting that rainfall is directly driven by the typhoon's spatial location and the intensity of its central pressure system. This analysis is consistent with objective meteorological laws and effectively verifies the rationality of the model's feature construction.

[0070] Spatiotemporal distribution: In this embodiment, a coastal area of ​​Province Z is selected and divided into a 1km × 1km grid (approximately 12,000 cells), and the static geographic features corresponding to the center points are extracted. For a specific typhoon event, hourly dynamic features of the typhoon are calculated and fed into the wind speed or rainfall model along with the static features.

[0071] The results show that the spatiotemporal evolution of the wind and rain field is consistent with objective meteorological understanding at all stages before, during, and after the typhoon makes landfall, with no significant spatial anomalies. This effectively verifies the continuity and accuracy of the model's prediction results in the spatiotemporal dimensions.

[0072] Long-term simulation grid division: The target area, Province Z, is divided into 1km × 1km grid cells (a total of approximately 102,000 grid cells), and static features such as elevation, slope, aspect, and distance from the coastline of the center point of each grid are extracted.

[0073] Synthetic Typhoon Generation: Based on ERA5 reanalysis data (1979–2021) under current climatic conditions, and using the statistical-dynamic downscaling method proposed by Emanuel et al. (2006; 2008), a total of 5,031 synthetic typhoon tracks were generated within a 200 km radius of the Z province boundary. This track set corresponds to typhoon events that may have affected Z province on a timescale of approximately 1735 years.

[0074] Large-scale wind and rain field simulation: For these 5,031 synthetic typhoon events, the distance field from the typhoon center to the center point of each grid is calculated hourly. This distance field, pressure field, and grid static features are input into a trained and validated random forest wind speed and precipitation model. The model quickly simulates (each grid cell simulates a typhoon event for less than 0.1 seconds, comparable to the speed of traditional parametric wind and rain models), outputting the time-series wind and rain field generated by each synthetic typhoon event in each grid cell. Furthermore, the maximum wind speed extreme value sequence and the maximum hourly rainfall extreme value sequence for each grid cell in these 5,031 typhoons are extracted.

[0075] Return Period Wind and Rain Estimation Extreme Value Distribution Fitting: Based on long-period (e.g., approximately 1735 years) extreme disaster field data generated in this embodiment, this embodiment takes a target grid cell in the coastal area of ​​Province Z as an example. An empirical distribution function is used to fit the extreme value sequences of maximum wind speed and maximum rainfall of this cell, respectively, thereby constructing the cumulative probability density function (CDF) for wind speed and rainfall. Figure 4 As shown in subplots a and c, the CDF curve for maximum wind speed rises sharply in the low-value area and then flattens out, indicating that the frequency of medium and low-intensity wind speeds is extremely high, while the probability of extreme strong wind speeds (>40 m / s) is less than 5%. The CDF curve for maximum hourly rainfall continues to rise in the high-value area, indicating that the risk of extreme heavy rainfall in this area is higher than that of extreme strong winds.

[0076] Return period calculation: In this embodiment, the average annual frequency of typhoon impact on Province Z is [missing information]. (Unit: times / year). Based on the above cumulative probability density function, the quantitative relationship between return period and extreme wind and rain loads is established using the following formula:

[0077] Based on this formula, the return-time load curves of wind speed or rainfall for the target grid cell are derived and output, such as... Figure 4As shown in subplots b and d, the wind speed return period curve shows a gradual slowdown in growth after 50 years, while the rainfall return period curve reaches saturation after 25 years, indicating that the extreme rainfall intensity reaches its upper limit relatively early, and extending the return period has limited effect on its improvement. By extending this calculation process in parallel to all grid cells within Province Z, spatial distribution data of wind and rain fields at different return periods (e.g., 50-year, 100-year return periods) can be obtained for the entire region. When conducting urban disaster prevention spatial planning and long-term engineering structural design, wind and rain field distribution maps at the corresponding return period levels can be directly retrieved based on local disaster resistance standards and needs, thus providing a high-precision quantitative design basis for regional long-term engineering design.

[0078] Example 2: Based on Example 1, a long-period typhoon wind and rain field return period estimation system based on random forest is also disclosed, including: The data collection module is used to collect static geographic data of meteorological observation stations in the target area and dynamic historical disaster site data during typhoon disasters; The data preprocessing module is used to extract the geographical features of meteorological observation stations, extract the hourly typhoon center pressure features and the distance features from meteorological observation stations to the typhoon center for each typhoon event, and use the hourly maximum wind speed and rainfall of each meteorological observation station in each typhoon event as feature labels. The model building module is used to build rainfall and wind speed models based on the random forest regression model. The geographical features of the meteorological observation station, the hourly typhoon center pressure and the distance from the meteorological observation station to the typhoon center are used as input features and the maximum wind speed and rainfall of the meteorological observation station are used as output features. The model training module is used to divide the samples obtained from data preprocessing into training and testing sets, and to train the rainfall model and wind speed model separately until the model overcomes the overfitting defect. The long-period simulation module is used to generate typhoon wind and rain fields for each grid cell in the target area based on the simulated typhoon event set, using the trained wind speed model and rainfall model, and obtain the wind and rain distribution of each grid cell in the target area. The wind and rain estimation module during the return period is used to construct the cumulative probability density function of wind and rain intensity for each grid cell based on the wind and rain distribution of each grid cell using extreme value statistical fitting method, and outputs the return period curve of wind speed and rainfall for each grid cell.

[0079] Example 3: like Figure 5As shown, the present invention also provides an electronic device, which may be a computer device, a microcontroller device, a smart mobile device, etc. The electronic device in this embodiment may include a processor, a memory, a transceiver component, etc. The memory, processor, and transceiver component are connected via a bus; the memory can be used to store executable programs, and an exemplary executable program may include instructions; the processor is used to execute the instructions stored in the memory. The memory can also be used to store data, which can be accessed and / or modified when instructions are executed.

[0080] The processor may be a Central Processing Unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, and it is suitable for implementing one or more instructions. Specifically, it is suitable for loading and executing one or more instructions in the storage medium to implement the corresponding method flow or corresponding function, so as to implement the steps of the long-period typhoon wind and rain field return period estimation method based on random forest in the above embodiment.

[0081] Example 4: Based on the same inventive concept, this invention also provides a readable storage medium, specifically an electronic device readable storage medium (Memory). This readable storage medium is a memory device within an electronic device used to store programs and data. It is understood that the storage medium here can include both built-in storage media within the electronic device and extended storage media supported by the electronic device. The storage medium provides storage space, which stores the terminal's operating system. Furthermore, this storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more executable programs (including program code). It should be noted that the storage medium here can be high-speed RAM or non-volatile memory, such as at least one disk storage device. Loading and executing one or more instructions stored in the storage medium by the processor can implement the steps of the long-period typhoon wind and rain field return period estimation method based on random forest in the above embodiments.

[0082] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0083] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0084] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0085] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0086] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any way. Other variations and modifications may be made without departing from the technical solutions described in the claims.

Claims

1. A method for estimating the return period of long-period typhoon wind and rain fields based on random forest, characterized in that... Includes the following steps: Data collection includes collecting static geographic data from meteorological observation stations in the target area and dynamic historical disaster site data during typhoon disasters; Data preprocessing involves extracting the geographical features of meteorological observation stations, extracting the hourly typhoon center pressure features and the distance features from meteorological observation stations to the typhoon center for each typhoon event, and using the hourly maximum wind speed and rainfall of each meteorological observation station in each typhoon event as feature labels. Model construction: Based on the random forest regression model, a rainfall model and a wind speed model are constructed respectively. The geographical characteristics of the meteorological observation station, the hourly typhoon center pressure are used as input features and the distance from the meteorological observation station to the typhoon center are used as output features. The maximum wind speed and rainfall of the meteorological observation station are used as output features. Model training involves dividing the samples obtained from data preprocessing into training and testing sets, and training the rainfall model and wind speed model separately until the model overcomes the overfitting defect. Long-term simulation, based on the simulated typhoon event set, uses the trained wind speed model and rainfall model to generate typhoon wind and rain field for each grid cell in the target area, and obtains the wind and rain distribution of each grid cell in the target area; For wind and rain estimation during the return period, based on the wind and rain distribution of each grid cell, the cumulative probability density function of wind and rain intensity for each grid cell is constructed by fitting the extreme value statistical fitting method, and the return period curves of wind speed and rainfall for each grid cell are output.

2. The method for estimating the return period of long-period typhoon wind and rain fields based on random forest according to claim 1, characterized in that, Static geographic data includes digital elevation model (DEM) and coastline data; dynamic historical disaster sites mainly include historical typhoon trajectories and measured meteorological data from various meteorological stations within the target area during each typhoon.

3. The method for estimating the return period of long-period typhoon wind and rain fields based on random forest according to claim 2, characterized in that, In data collection, geographic features include the elevation, slope and aspect of the observation point, as well as the shortest straight-line distance from the observation point to the coastline.

4. The method for estimating the return period of long-period typhoon wind and rain fields based on random forest as described in claim 1, characterized in that... During model training, for the wind speed model or model, the samples in its training set are exponentially weighted, as expressed by: in, Let i be the training weights for the i-th sample. Adjusting hyperparameters that focus on extreme values. Indicates the first The actual wind speed or rainfall of each sample and This represents the minimum and maximum values ​​of the actual wind speed or rainfall in the training set.

5. The method for estimating the return period of long-period typhoon wind and rain fields based on random forest according to claim 1, characterized in that, The indicators for a model to overcome overfitting are that the mean squared error (MSE) and mean absolute error (MAE) of the training and test sets are less than a preset threshold.

6. The method for estimating the return period of long-period typhoon wind and rain fields based on random forest as described in claim 5, is characterized in that... Improve the interpretability of wind speed and rainfall models through feature importance assessment.

7. The method for estimating the return period of long-period typhoon wind and rain fields based on random forest according to claim 6, characterized in that, To verify the spatiotemporal distribution of the disaster site: using the test set samples as input, the rainfall and wind speed data output by the rainfall model and wind speed model are obtained; the degree of agreement between the rainfall and wind speed data and the actual disaster-affected areas and time periods is compared; when the degree of agreement is higher than the threshold, it proves that the rainfall model and wind speed model have overcome the overfitting defect.

8. The method for estimating the return period of long-period typhoon wind and rain fields based on random forest according to claim 1, characterized in that, The expression for the cumulative probability density function is: Indicates the recurrence period. This indicates the average annual frequency of typhoons affecting the target area. This indicates that the wind speed or rainfall exceeded the threshold during the typhoon period. The probability of.

9. The method for estimating the return period of long-period typhoon wind and rain fields based on random forest as described in claim 1, characterized in that, extreme values... Statistical fitting methods include one of the following: empirical distribution, Gumbel distribution, and Weibull distribution.

10. A method for estimating the return period of long-period typhoon wind and rain fields based on random forest according to any one of claims 1 to 9, characterized in that, Based on a given recurrence period By obtaining the wind speed and rainfall of the grid cell corresponding to the return period of the network cell through the return period curve, a wind and rain field distribution map of the target area is generated based on the wind speed and rainfall of the grid cell.

Images