Severe convection strong wind weather identification method based on machine learning

Abstract

The invention relates to a severe convection strong wind weather identification method based on machine learning, and solves the problems of poor accuracy and real-time performance of severe convection strong wind early warning in the prior art. The adopted technical scheme is as follows: obtaining radar echo combination reflectivity data, gridding maximum wind speed observation data and historical tropical cyclone path data; after quality control and data resampling, data influenced by the tropical cyclones are eliminated, and effective samples with space, regional statistics and time change characteristics of radar echoes are screened; and inputting the effective samples into a machine learning model for training, and outputting the probability and / or spatial distribution of the strong convection wind. The method has the advantages that by eliminating tropical cyclone influence data, the sample purity is improved, and the misjudgment basis is reduced; multi-dimensional features are extracted, strong convection strong wind attributes are completely captured, and the accuracy can be remarkably improved; and training the effective samples through a machine learning model, outputting the probability and / or spatial distribution of strong convection strong wind, and conforming to a real-time accurate early warning service scene.

Description

Technical Field

[0001] This invention relates to the field of strong wind weather warning, and in particular to a machine learning-based method for identifying strong convective wind weather for issuing early warnings. Background Technology

[0002] Severe convective winds are extreme weather phenomena triggered by convective activity, typically accompanied by short-duration strong winds, thunderstorms, and precipitation. This type of weather is not only sudden and localized, but also has a serious impact on public safety, transportation, agriculture, and construction facilities. Especially in urban and coastal areas, severe convective winds can cause catastrophic consequences such as power outages, fallen trees, and building damage. Therefore, accurate and timely early warning of severe convective winds is of significant practical importance.

[0003] With the deepening of meteorological research, especially the continuous advancement of meteorological data acquisition and analysis technologies, existing early warning methods have developed into various technical means, including comprehensive analysis methods based on meteorological station data, radar echoes, satellite remote sensing, and numerical weather prediction. These traditional methods provide important data support for the monitoring and early warning of severe convective winds, but they still face many challenges in dealing with the suddenness and localization of severe convective weather. Traditional weather early warning methods based on physical models are usually limited by the timeliness, accuracy, and computational power of the data, making it difficult to provide high-precision early warning results in real time.

[0004] In recent years, with the rapid development of machine learning and artificial intelligence technologies, more and more researchers are attempting to introduce these emerging technologies into early warning systems for severe convective winds. By applying deep learning and pattern recognition to historical data, machine learning models can discover the potential patterns of severe convective winds, thereby improving the accuracy and timeliness of early warning systems. For example, deep learning methods such as generative adversarial networks and long short-term memory networks have been applied to meteorological data analysis, demonstrating good predictive performance. Especially in high-resolution data processing and rapid response, machine learning methods provide new solutions for early warning of severe convective winds.

[0005] Although some progress has been made in the early warning of severe convective winds, bottlenecks still exist in terms of model accuracy and real-time performance. While traditional early warning methods combining radar echoes and satellite data are effective, the generalization ability of the model and data quality issues still affect the effectiveness of early warnings in complex meteorological environments. To address this, this invention proposes an early warning method for severe convective winds based on the LightGBM model. Summary of the Invention

[0006] The purpose of this invention is to address the aforementioned problems in existing technologies by providing a machine learning-based method for identifying severe convective winds. This method improves sample purity and reduces the basis for misjudgments by removing data influenced by tropical cyclones. It extracts multi-dimensional features including spatial characteristics, regional statistical characteristics, and temporal variation characteristics, enabling the complete capture of severe convective wind attributes (including locality and suddenness), significantly improving early warning accuracy. Furthermore, it trains effective samples using a machine learning model to output the probability and / or spatial distribution of severe convective winds, aligning with real-time and accurate early warning scenarios and adapting to business needs.

[0007] The above-mentioned technical objective of the present invention is mainly achieved through the following technical solution: a multifunctional leisure sofa, including a support frame and a seat cushion set on the support frame, characterized in that the support frame is a hollow frame formed by pressing and splicing steel sheets, with symmetrical openings on both sides of the support frame and connecting them, and the hollow part forming a receiving cavity.

[0008] As a further improvement and supplement to the above technical solution, the present invention adopts the following technical measures: The technical solution of the first technical subject of this invention is as follows: A method for identifying severe convective winds based on machine learning, characterized by the following steps: Data acquisition: Acquire radar echo combined reflectivity data with a preset time span, first spatiotemporal resolution, and coverage of a preset geographical area; acquire gridded maximum wind speed observation data with second spatiotemporal resolution; acquire historical tropical cyclone path data; Data quality control: Remove interference data from the radar echo combined reflectivity data; eliminate outliers from the gridded maximum wind speed observation data; Data resampling: Based on the target spatiotemporal resolution matched with the maximum wind speed observation data, the quality-controlled radar echo combined reflectivity data is resampled to obtain resampled data; Formation of severe convective wind weather samples: Based on the preset tropical cyclone path data, data within the period affected by the tropical cyclone in the resampled data are removed; a single grid in the gridded maximum wind speed observation data is used as the target grid, and data in the remaining data after removal that meet the conditions that the maximum wind speed of the target grid is greater than or equal to a preset wind speed threshold and that there are radar echoes greater than or equal to a preset echo intensity threshold within a preset radius centered on the target grid are judged as valid data samples of severe convective wind weather; Feature extraction: From the effective data samples, feature parameters of radar echo combined reflectivity are extracted, including spatial features, regional statistical features and temporal variation features; Model training: The feature parameters are input into a preset machine learning model for training to obtain the probability of strong convective winds occurring at the corresponding spatial location and / or the spatial distribution of strong convective winds.

[0009] Preferably, the preset time span is a time span that includes a complete seasonal cycle or a time span greater than or equal to one year. The first spatiotemporal resolution is the first temporal resolution corresponding to the first preset time interval and the first spatial resolution corresponding to the first preset spatial precision; The preset geographical area is a geographical range that includes at least one sub-region with a high incidence of severe convective weather. The second spatiotemporal resolution is the second time resolution corresponding to the second preset time interval and the second spatial resolution corresponding to the second preset spatial precision; The historical tropical cyclone track data is standardized data that includes information on tropical cyclone track, intensity, and duration of impact.

[0010] Preferably, in the data quality control step, the radar echo combined reflectivity data is processed based on a preset filtering algorithm to remove interference data; and outliers are removed from the gridded maximum wind speed observation data through extreme value analysis.

[0011] Preferably, the preset wind speed threshold is a critical wind speed value that can characterize severe convective wind weather. This critical value is determined based on the definition standard of severe convective wind in meteorology and the needs of disaster early warning. The preset spatial range is an effective range that can reflect the spatial correlation between radar echoes and severe convective winds. This range is determined based on the horizontal scale of the severe convective system and the influence radius of the radar echo. The preset echo intensity threshold is a critical radar echo intensity value that can indicate the existence of a severe convective system. This critical value is determined based on the statistical analysis of radar echo characteristics of severe convective weather and the need to determine the validity of samples.

[0012] Preferably, in the data resampling step, the target spatiotemporal resolution includes a target temporal resolution and a target spatial resolution, wherein the target temporal resolution is a resolution that matches the temporal resolution of the maximum wind speed observation data, and the target spatial resolution is a resolution that matches the spatial resolution of the maximum wind speed observation data. In the feature extraction step, the feature parameters are quantitative parameters that characterize the combined reflectivity of radar echoes within a time window corresponding to each target temporal resolution. These parameters include: the local echo intensity of the target grid, the statistical value of the echo intensity within a preset radius centered on the target grid, the proportion of grid areas within the preset radius whose echo intensity is greater than or equal to a preset threshold, and the change of each feature parameter within adjacent time windows. The statistical value includes the maximum value, the average value, and the variance.

[0013] As a preferred method, extreme value pooling is used when resampling the target spatial resolution of radar echo combined reflectivity data: Within the target grid, a sampling window with a preset latitude and longitude range matching the target resolution is constructed with the grid center as the reference. The maximum value of all original radar echo combined reflectivity data within the sampling window is calculated as the representative value of the target grid. For the sampling window portion that exceeds the coverage of the original data, boundary adaptive processing is performed.

[0014] Preferably, the preset machine learning model is a tree-based ensemble learning model suitable for identifying strong convective winds, including gradient boosting decision tree models; the model uses a classification performance evaluation index that reflects the balance between the accuracy and coverage of strong convective wind identification as the objective function; the optimal hyperparameter combination of the model is determined by an adaptive hyperparameter optimization algorithm combined with multi-fold cross-validation.

[0015] Preferably, before the model training step, a dataset partitioning step is included: dividing the effective samples of severe convective windy weather and the corresponding samples of windless weather into a training set, a validation set, and a test set, with the partitioning ratio determined based on the requirements of model training sufficiency, validation reliability, and test objectivity; during the partitioning process, sample equalization processing is performed to address sample class imbalance, so as to ensure that the effective samples of severe convective windy weather and the samples of windless weather in the training set, validation set, and test set maintain a consistent class distribution.

[0016] As a preferred embodiment, a machine learning-based method for identifying severe convective winds also includes an optimal discrimination threshold determination step: on the model validation set, multiple candidate probability thresholds within a preset probability threshold range are traversed, a preset classification performance evaluation index is calculated for the model identification result corresponding to each candidate probability threshold, and the candidate probability threshold that makes the classification performance evaluation index reach its optimal value is selected as the optimal discrimination threshold of the model.

[0017] As a preferred approach, during the operational phase, the real-time acquired radar echo combined reflectivity data undergoes quality control processing and data resampling processing sequentially. The processed data is then input into a pre-trained machine learning model, which outputs a probability distribution of severe convective wind weather that matches the target spatiotemporal resolution. Finally, based on the optimal discrimination threshold, a spatial distribution map of severe convective wind weather is generated.

[0018] The technical solution of the second technical subject matter involved in this invention is as follows: A computer-readable storage medium, characterized in that the computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to perform the aforementioned machine learning-based method for identifying severe convective wind weather.

[0019] The beneficial effects of this invention are as follows: 1. Improve sample purity and reduce misjudgment basis: By removing data affected by tropical cyclones, the confusion between typhoon winds and strong convective winds is avoided, ensuring that the training samples are "uncontaminated" and laying the foundation for the model to accurately identify the core features of strong convection.

[0020] 2. Enhance feature representation and improve recognition accuracy: Extract multi-dimensional features including "spatial features + regional statistical features + temporal variation features" to fully capture the "local and sudden" attributes of strong convective winds. Compared with single feature recognition, the accuracy is significantly improved. 3. Optimize model performance and reduce false negatives and missing reports: Solve the class imbalance problem by sample equalization, and combine adaptive hyperparameter optimization and system threshold determination to avoid biased predictions. The false negative and missing reports rates are reduced, while the generalization ability of the model in different regions is improved. 4. Adapt to business needs and achieve practical implementation: Unify data to the target spatiotemporal resolution, output a spatial distribution map of strong convective winds that can be directly applied, fit the business scenario of real-time early warning, solve the problem of "difficult to implement results" of existing technologies, and improve business practicality. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of a method flow of the present invention. Detailed Implementation

[0022] The technical solution of the present invention will be further described in detail below through embodiments and in conjunction with the accompanying drawings.

[0023] Example 1: As Figure 1 As shown, the technical solution of the first technical subject matter involved in this invention is as follows: A method for identifying severe convective winds based on machine learning, characterized by the following steps: Data acquisition: Acquire radar echo combined reflectivity data with a preset time span, first spatiotemporal resolution, and coverage of a preset geographical area; acquire gridded maximum wind speed observation data with second spatiotemporal resolution; acquire historical tropical cyclone path data; Data quality control: Remove interference data from the radar echo combined reflectivity data; eliminate outliers from the gridded maximum wind speed observation data; Data resampling: Based on the target spatiotemporal resolution matched with the maximum wind speed observation data, the quality-controlled radar echo combined reflectivity data is resampled to obtain resampled data; Formation of severe convective wind weather samples: Based on the preset tropical cyclone path data, data within the period affected by the tropical cyclone in the resampled data are removed; a single grid in the gridded maximum wind speed observation data is used as the target grid, and data in the remaining data after removal that meet the conditions that the maximum wind speed of the target grid is greater than or equal to a preset wind speed threshold and that there are radar echoes greater than or equal to a preset echo intensity threshold within a preset radius centered on the target grid are judged as valid data samples of severe convective wind weather; Feature extraction: From the effective data samples, feature parameters of radar echo combined reflectivity are extracted, including spatial features, regional statistical features and temporal variation features; Model training: The feature parameters are input into a preset machine learning model for training to obtain the probability of strong convective winds occurring at the corresponding spatial location and / or the spatial distribution of strong convective winds.

[0024] This technical solution effectively addresses the shortcomings of existing technologies: 1. Insufficient sample purity: Data affected by tropical cyclones (typhoons) were not specifically removed, and samples of typhoon strong winds were confused with those of severe convective winds, resulting in "contamination" of the training samples. The model is prone to misjudgment and cannot accurately identify the core characteristics of severe convective winds. 2. Limited feature dimensions: Most of them rely solely on basic features such as single-point radar echo intensity, without combining "regional statistical features (such as echo mean / variance within the radius)" and "temporal variation features (such as echo intensity trend)", which makes it impossible to capture the essential attributes of strong convective winds as "local and sudden", resulting in low identification accuracy. 3. Insufficient model optimization: The problem of "class imbalance" in strong convection samples (few positive samples) has not been addressed, and the model is prone to biased prediction of "no strong winds", resulting in a high false negative rate; moreover, hyperparameters and discrimination thresholds are mostly set based on experience and have not been determined through adaptive optimization + multi-fold cross-validation system, resulting in poor model generalization. 1. Weak business adaptability: Data processing (such as resampling) is not unified to the target spatiotemporal resolution that matches business needs, and the output results (such as probability values) are not converted into a "spatial distribution map" that can be directly applied, which cannot meet the business implementation needs of real-time early warning.

[0025] Therefore, the technical effects achieved by this technology are as follows: 1. Improve sample purity and reduce misjudgment basis: By removing data affected by tropical cyclones, the confusion between typhoon winds and strong convective winds is avoided, ensuring that the training samples are "uncontaminated" and laying the foundation for the model to accurately identify the core features of strong convection.

[0026] 2. Enhance feature representation and improve recognition accuracy: Extract multi-dimensional features including "spatial features + regional statistical features + temporal variation features" to fully capture the "local and sudden" attributes of strong convective winds. Compared with single feature recognition, the accuracy is significantly improved. 3. Optimize model performance and reduce false negatives and missing reports: Solve the class imbalance problem by sample equalization, and combine adaptive hyperparameter optimization and system threshold determination to avoid biased predictions. The false negative and missing reports rates are reduced, while the generalization ability of the model in different regions is improved. 4. Adapt to business needs and achieve practical implementation: Unify data to the target spatiotemporal resolution, output a spatial distribution map of strong convective winds that can be directly applied, fit the business scenario of real-time early warning, solve the problem of "difficult to implement results" of existing technologies, and improve business practicality.

[0027] Next, the above technical solution will be explained in detail: In practical applications, the preset time span is a time span that includes a complete seasonal cycle or a time span that is greater than or equal to one year. The first spatiotemporal resolution is the first temporal resolution corresponding to the first preset time interval and the first spatial resolution corresponding to the first preset spatial precision; The preset geographical area is a geographical range that includes at least one sub-region with a high incidence of severe convective weather. The second spatiotemporal resolution is the second time resolution corresponding to the second preset time interval and the second spatial resolution corresponding to the second preset spatial precision; The historical tropical cyclone track data is standardized data that includes information on tropical cyclone track, intensity, and duration of impact.

[0028] In practical applications, during the data quality control step, the radar echo combined reflectivity data is processed based on a preset filtering algorithm to remove interference data; and outliers are eliminated from the gridded maximum wind speed observation data through extreme value analysis.

[0029] In practical applications, the preset wind speed threshold is a critical wind speed value that can characterize severe convective wind weather. This critical value is determined based on the definition standard of severe convective wind in meteorology and the needs of disaster early warning. The preset spatial range is an effective range that can reflect the spatial correlation between radar echoes and severe convective winds. This range is determined based on the horizontal scale of the severe convective system and the influence radius of the radar echo. The preset echo intensity threshold is a critical radar echo intensity value that can indicate the existence of a severe convective system. This critical value is determined based on the statistical analysis of radar echo characteristics of severe convective weather and the need to determine the validity of samples.

[0030] In practical applications, in the data resampling step, the target spatiotemporal resolution includes a target temporal resolution and a target spatial resolution. The target temporal resolution is a resolution that matches the temporal resolution of the maximum wind speed observation data, and the target spatial resolution is a resolution that matches the spatial resolution of the maximum wind speed observation data. In the feature extraction step, the feature parameters are quantitative parameters that characterize the combined reflectivity of radar echoes within a time window corresponding to each target temporal resolution. These parameters include: the local echo intensity of the target grid, the statistical value of the echo intensity within a preset radius centered on the target grid, the proportion of grid areas within the preset radius whose echo intensity is greater than or equal to a preset threshold, and the change of each feature parameter within adjacent time windows. The statistical value includes the maximum value, the average value, and the variance.

[0031] In practical applications, extreme value pooling is used when resampling radar echo combined reflectivity data for target spatial resolution. Within the target grid, a sampling window with a preset latitude and longitude range matching the target resolution is constructed with the grid center as the reference. The maximum value of all original radar echo combined reflectivity data within the sampling window is calculated as the representative value of the target grid. For the sampling window portion that exceeds the coverage of the original data, boundary adaptive processing is performed.

[0032] In practical applications, the preset machine learning model is a tree-based ensemble learning model suitable for identifying strong convective winds, including gradient boosting decision tree models; the model uses a classification performance evaluation index that reflects the balance between accuracy and coverage in identifying strong convective winds as the objective function; the optimal hyperparameter combination of the model is determined by an adaptive hyperparameter optimization algorithm combined with multi-fold cross-validation.

[0033] In practical applications, before the model training step, a dataset partitioning step is also included: dividing the effective samples of severe convective windy weather and the corresponding samples of windless weather into training set, validation set and test set, with the partitioning ratio determined based on the requirements of model training sufficiency, validation reliability and test objectivity; during the partitioning process, sample equalization processing is performed to address sample class imbalance, so as to ensure that the effective samples of severe convective windy weather and the samples of windless weather in the training set, validation set and test set maintain a consistent class distribution.

[0034] In practical applications, a machine learning-based method for identifying severe convective winds also includes an optimal discrimination threshold determination step: on the model validation set, multiple candidate probability thresholds within a preset probability threshold range are traversed, a preset classification performance evaluation index is calculated for the model recognition result corresponding to each candidate probability threshold, and the candidate probability threshold that makes the classification performance evaluation index reach its optimal value is selected as the optimal discrimination threshold of the model.

[0035] In practical applications, during the operational phase, the real-time acquired radar echo combined reflectivity data undergoes quality control processing and data resampling processing sequentially. The processed data is then input into a pre-trained machine learning model, which outputs a probability distribution of severe convective wind weather that matches the target spatiotemporal resolution. Finally, based on the optimal discrimination threshold, a spatial distribution map of severe convective wind weather is generated.

[0036] Finally, the above technical solutions are applied to practical cases for explanation and illustration, with appropriate supplements provided: 1. Data collection and quality control: (1) The radar echo combined reflectivity data (referred to as radar echo data) covering a certain area (such as Beijing, Zhejiang Province, Shanghai, etc.), gridded maximum wind speed observation data (referred to as maximum wind speed data, which is usually long-term data), and historical tropical cyclone track data (such as tropical cyclone track data or optimal tropical cyclone track data obtained by the China Meteorological Administration, which are used to screen out typhoon-related gale weather) are used. The radar echo data covers the area and surrounding areas, with a spatial resolution of 250m and a temporal resolution of 6 minutes, and a data time span of more than one year. The maximum wind speed data covers a similar spatial range, with a spatial resolution of 1km and a temporal resolution of 10 minutes, and the time span of the maximum wind speed data is the same as that of the radar echo data; the historical tropical cyclone track data includes statistics on typhoon occurrence records and tracks in the Northwest Pacific.

[0037] (2) The radar echo data is processed by Gabella filter. By utilizing its discrimination mechanism based on spatial consistency and echo characteristics, non-meteorological echoes caused by ground object reflection, clutter, antenna sidelobes, etc. are effectively identified and eliminated, thereby improving the authenticity and usability of the echo field and ensuring the accuracy of subsequent strong convection identification.

[0038] (3) Combining long-term historical observation data (i.e. gridded maximum wind speed observation data), when a certain observed wind speed deviates significantly from the long-term extreme value distribution of the region, it is judged as an outlier and removed, thereby avoiding the impact of unreasonable maximum wind speed values ​​caused by occasional erroneous observations or instrument failures on the reliability of the identification results.

[0039] 2. Sample labeling for severe convective wind weather (1) Resample the radar echo data to a 1km resolution to match the gridded maximum wind observation data. Max pooling was used to resample the radar data to a 1km resolution to match the gridded high wind data. Specifically, within the target grid, on the target latitude and longitude regular grid (grid step size 0.01°), a sampling window (latitude ±0.005°, longitude ±0.005°) consistent with the target resolution (e.g., spatial resolution of 1 km, temporal resolution of 10 minutes) was constructed with the center of each target grid point (i.e., the grid center) as the reference. The pixel index interval covered by the sampling window was located in the original regular grid, and the maximum value within the window was calculated as the representative value of the target grid. The portion of the sampling window that exceeded the boundary was automatically clipped (i.e., boundary adaptive processing). The final output target field (i.e., the radar echo combined reflectivity data after quality control, obtained by resampling the radar echo data after matching the target spatiotemporal resolution with the gridded maximum wind speed observation data) was strictly aligned with the gridded high wind data in terms of spatial resolution for subsequent grid-by-grid comparison and statistical analysis.

[0040] (2) For each full 10-minute interval, the radar echo data of the current time and at most two past times are used, and the maximum value at each location is taken to sample the time resolution to 10 minutes. For example, for 1 hour 20 minutes, the maximum value of the radar echo observed at 1 hour 12 minutes and 1 hour 18 minutes is taken.

[0041] (3) Based on historical tropical cyclone path data, data within the typhoon impact period are removed and not considered as valid samples.

[0042] (4) In this invention, a grid with a maximum wind speed of level 6 (10.8 ms-1) or above and a radar echo of greater than or equal to 35 dBZ within a radius of 20 km is recorded as a valid sample of strong convective wind weather.

[0043] 3. Training the LightGBM machine learning model (i.e., the LightGBM model) (1) Based on radar echo data and maximum wind speed data, considering multiple factors (i.e., multidimensional features), the characteristics of single-point radar echoes are statistically analyzed. Specifically, this includes: local single-point combined reflectivity intensity (i.e., the local echo intensity of the target grid); statistical values ​​of combined reflectivity within a radius of 20km (including maximum, average, and variance); the area proportion within a radius of 20km with combined reflectivity ≥35dBZ; the trend of the maximum combined reflectivity within a radius of 20km over the past 10 minutes; and the trend of the area proportion within a radius of 20km with combined reflectivity ≥35dBZ over the past 10 minutes.

[0044] (2) In this invention, 80% of the data is used as the training set, 10% of the data is used as the validation set, and 10% of the data is used as the test set. Since the number of samples of strong convective windy weather is much smaller than the number of samples of no windy weather, when dividing the dataset, the imbalance of the samples is considered and the samples are randomly resampled. The proportion of strong convective windy weather samples and no windy weather samples is consistent in different datasets.

[0045] (3) In this invention, the input to the LightGBM model is the radar echo features extracted from radar echo data at a resolution of 1km and every 10 minutes, as well as the results of identifying strong convective wind areas. To fully consider the accuracy of wind weather identification and avoid prediction bias due to an excessive number of samples without strong winds, this invention uses the Critical Success Index (CSI) as the objective function. The CSI is determined by the number of hits (TruePositive, TP), false alarms (FalsePositive, FP), and false negatives (FalseNegative, FN). This index comprehensively measures the model's false alarm and false alarm rates, making it more suitable for identifying strong convective winds under imbalanced sample conditions. Its calculation formula is: .

[0046] Meanwhile, to improve model performance, this invention employs the Optuna hyperparameter optimization framework to automatically search for hyperparameters in the LightGBM model. Optuna, through a Bayesian optimization-based sampling algorithm, can dynamically adjust the parameter search direction using previous experimental results, thereby quickly approximating the global optimum with a limited number of iterations. Furthermore, during the hyperparameter optimization process, this invention uses a five-fold cross-validation method to evaluate the performance of each set of candidate parameters. This involves dividing the training data into five subsets, which are used alternately as the validation set, with the remaining subset serving as the training set. The average CSI from the five validations is calculated as the final performance index for that set of parameters, ensuring the stability and generalization ability of the parameter selection.

[0047] (4) In the validation set, this invention further performs threshold sensitivity analysis on the probability output by the model under the optimal hyperparameter combination obtained by five-fold cross-validation. Specifically, for each sample in the validation set, the LightGBM model first outputs the probability value of its belonging to the strong convective wind category. Subsequently, within the range of 0.01 to 0.99, a discrimination threshold is set one by one with a step size of 0.01, and the sample is classified based on each threshold. That is, when the predicted probability is greater than or equal to the threshold, it is judged as a strong wind event; otherwise, it is judged as a non-strong wind event. Further, this invention calculates the model's CSI under each candidate threshold to comprehensively evaluate the discrimination effect under different probability threshold conditions. By comparing the CSI index under all candidate thresholds, the threshold that maximizes the CSI value is selected as the final optimal discrimination probability threshold. This threshold is fixed in subsequent independent sample prediction and actual business applications to achieve optimal identification of strong convective wind areas.

[0048] (5) During the testing phase, the LightGBM model, after hyperparameter optimization and determination of the optimal discrimination threshold, was applied to an independent test set to evaluate its discrimination skills and generalization ability for severe convective winds. Specifically, indicators such as CSI, Probability of Detection (POD), False Alarm Ratio (FAR), and Missing Rate (MR) were selected to systematically evaluate the model performance. The calculation formulas for POD, FAR, and MR are as follows: , , .

[0049] Among them, POD is used to measure the model's ability to detect actual strong wind events, FAR is used to reflect the model's false alarm situation, MR reflects the degree of model missed detection, and CSI comprehensively considers hits, false alarms and missed detections, and can fully reflect the model's practical application value in the identification of strong convective winds.

[0050] 4. Model Business Practice During operation, the real-time radar echo reflectivity data from the two most recent moments are subjected to quality control and resampled to a spatial resolution of 1km and a temporal resolution of 10 minutes. The 10-minute radar echo data of a single frame is used as the input to the business model, and the probability distribution map of gale-force winds of level 6 or above at a spatial resolution of 1km is output. The spatial distribution of gale-force winds is obtained by using the optimal probability discrimination threshold.

[0051] Example 2: The technical solution of the second technical subject matter involved in this invention is as follows: A computer-readable storage medium, characterized in that it includes a stored computer program, wherein, when the computer program is executed, it controls the device containing the computer-readable storage medium to perform the aforementioned machine learning-based method for identifying severe convective winds.

[0052] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Various modifications and variations can be made to the above embodiments. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for identifying severe convective winds based on machine learning, characterized in that, The steps include: Data acquisition: Acquire radar echo combined reflectivity data with a preset time span, first spatiotemporal resolution, and coverage of a preset geographical area; acquire gridded maximum wind speed observation data with second spatiotemporal resolution; acquire historical tropical cyclone path data; Data quality control: Remove interference data from the radar echo combined reflectivity data; remove outliers from the gridded maximum wind speed observation data; Data resampling: Based on the target spatiotemporal resolution matched with the maximum wind speed observation data, the quality-controlled radar echo combined reflectivity data is resampled to obtain resampled data; Formation of severe convective wind weather samples: Based on the preset tropical cyclone path data, data within the period affected by the tropical cyclone in the resampled data are removed; a single grid in the gridded maximum wind speed observation data is used as the target grid, and data in the remaining data after removal that meet the conditions that the maximum wind speed of the target grid is greater than or equal to a preset wind speed threshold and that there are radar echoes greater than or equal to a preset echo intensity threshold within a preset radius centered on the target grid are judged as valid data samples of severe convective wind weather; Feature extraction: From the effective data samples, feature parameters of radar echo combined reflectivity are extracted, including spatial features, regional statistical features and temporal variation features; Model training: The feature parameters are input into a preset machine learning model for training to obtain the probability of strong convective winds occurring at the corresponding spatial location and / or the spatial distribution of strong convective winds.

2. The method for identifying severe convective winds based on machine learning according to claim 1, characterized in that, The preset time span is a time span that includes a complete seasonal cycle or a time span greater than or equal to one year; The first spatiotemporal resolution is the first temporal resolution corresponding to the first preset time interval and the first spatial resolution corresponding to the first preset spatial precision; The preset geographical area is a geographical range that includes at least one sub-region with a high incidence of severe convective weather. The second spatiotemporal resolution is the second time resolution corresponding to the second preset time interval and the second spatial resolution corresponding to the second preset spatial precision; The historical tropical cyclone track data is standardized data that includes information on tropical cyclone track, intensity, and duration of impact.

3. The method for identifying severe convective winds based on machine learning according to claim 1 or 2, characterized in that, In the data quality control step, the radar echo combined reflectivity data is processed based on a preset filtering algorithm to remove interference data; and outliers are removed from the gridded maximum wind speed observation data through extreme value analysis.

4. The method for identifying severe convective winds based on machine learning according to claim 1, characterized in that, The preset wind speed threshold is a critical wind speed value that can characterize severe convective wind weather. This critical value is determined based on the definition standard of severe convective wind in meteorology and the needs of disaster early warning. The preset spatial range is an effective range that can reflect the spatial correlation between radar echo and severe convective wind. This range is determined based on the horizontal scale of the severe convective system and the influence radius of the radar echo. The preset echo intensity threshold is a critical radar echo intensity value that can indicate the existence of a severe convective system. This critical value is determined based on the statistical analysis of radar echo characteristics of severe convective weather and the need to determine the validity of samples.

5. The method for identifying severe convective winds based on machine learning according to claim 4, characterized in that, In the data resampling step, the target spatiotemporal resolution includes a target temporal resolution and a target spatial resolution. The target temporal resolution is a resolution that matches the temporal resolution of the maximum wind speed observation data, and the target spatial resolution is a resolution that matches the spatial resolution of the maximum wind speed observation data. In the feature extraction step, the feature parameters are quantitative parameters that characterize the combined reflectivity of radar echoes within a time window corresponding to each target temporal resolution. These parameters include: the local echo intensity of the target grid, the statistical value of the echo intensity within a preset radius centered on the target grid, the proportion of grid areas within the preset radius whose echo intensity is greater than or equal to a preset threshold, and the change of each feature parameter within adjacent time windows. The statistical value includes the maximum value, the average value, and the variance.

6. The method for identifying severe convective winds based on machine learning according to claim 1 or 2, characterized in that, Extreme pooling is used when resampling target spatial resolution data of combined radar echo reflectivity data. Within the target grid, a sampling window with a preset latitude and longitude range matching the target resolution is constructed with the grid center as the reference. The maximum value of all original radar echo combined reflectivity data within the sampling window is calculated as the representative value of the target grid. For the sampling window portion that exceeds the coverage of the original data, boundary adaptive processing is performed.

7. The method for identifying severe convective winds based on machine learning according to claim 1, characterized in that, Before the model training step, a dataset partitioning step is also included: the effective samples of severe convective windy weather and the corresponding samples of windless weather are divided into training set, validation set and test set, and the partitioning ratio is determined based on the requirements of model training sufficiency, validation reliability and test objectivity; during the partitioning process, sample equalization processing is performed to address the sample class imbalance, so as to ensure that the effective samples of severe convective windy weather and the samples of windless weather in the training set, validation set and test set maintain a consistent class distribution.

8. The method for identifying severe convective winds based on machine learning according to claim 1, characterized in that, It also includes the optimal discrimination threshold determination step: on the model validation set, traverse multiple candidate probability thresholds within the preset probability threshold range, calculate the preset classification performance evaluation index for the model recognition result corresponding to each candidate probability threshold, and select the candidate probability threshold that makes the classification performance evaluation index reach the optimal value as the optimal discrimination threshold of the model.

9. The method for identifying severe convective winds based on machine learning according to claim 1, characterized in that, During the operational phase, the real-time acquired radar echo combined reflectivity data undergoes quality control processing and data resampling processing sequentially. The processed data is then input into a pre-trained machine learning model, which outputs a probability distribution of severe convective wind weather that matches the target spatiotemporal resolution. Finally, based on the optimal discrimination threshold, a spatial distribution map of severe convective wind weather is generated.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to perform the machine learning-based method for identifying severe convective winds as described in any one of claims 1 to 9.

Images