A 24-hour forecast method for heat wave in the vicinity of tropical cyclone triggering area

Abstract

This invention discloses a 24-hour forecasting method for heat waves triggered by approaching tropical cyclones, belonging to the field of meteorological disaster forecasting technology. The method first delineates the tropical cyclone approach area window and defines the approach sample time; then, it integrates tropical cyclone path, meteorological, and topographic data to construct a candidate factor pool for physical classification; next, it identifies regional heat wave events, generating 24-hour forecast labels and historical sample sets; subsequently, through grouped cross-validation and comprehensive score evaluation, it adaptively selects the optimal factor combination and trains the classification model; finally, it combines multi-objective optimization to determine the warning threshold and outputs a heat wave risk warning. This method solves the problems of weak scenario specificity and poor generalization performance of existing models, achieving accurate forecasting of heat waves under tropical cyclone approach scenarios, and providing reliable technical support for early prevention of disaster chains.

Description

Technical Field

[0001] This invention relates to the field of meteorological disaster risk forecasting and early warning technology, specifically to a 24-hour forecasting method for heat waves in areas triggered by approaching tropical cyclones. Background Technology

[0002] Heat waves are extreme high-temperature weather events that pose a serious threat to public health, agricultural production, and energy supply. Existing research indicates that under specific circulation patterns, the peripheral descending airflow of tropical cyclones, along with accompanying clear-sky radiation and orographic foehn effects, can lead to high-intensity, regional heat waves within their affected areas. Taking South China as an example, studies have shown that a significant proportion of heat wave events in this region are linked to tropical cyclone activity. These heat waves often precede the impact of the tropical cyclone's core wind and rain zone, forming a "high temperature-strong wind and heavy rain" disaster chain, significantly amplifying the overall risk of tropical cyclones. Therefore, conducting 24-hour advance forecasts and warnings for regional heat wave risks under tropical cyclone approach scenarios is of significant practical importance for early prevention of disaster chains.

[0003] Currently, heatwave forecasting under the influence of tropical cyclones mainly relies on two technical approaches: The first is dynamic forecasting and its correction methods based on numerical weather prediction models or reanalysis data. In operational practice, meteorological departments often refer to the 2-meter temperature forecasts output by the European Centre for Medium-Range Weather Forecasts (ECMWF) or regional numerical models when issuing high-temperature warnings. However, the forecasting effectiveness of this method is highly sensitive to the forecasting errors of tropical cyclone tracks, intensity, and related mesoscale processes. Furthermore, the evolution of tropical cyclones and their peripheral circulation structures has significant uncertainties, making it difficult for forecast results to consistently reflect heatwave risk in some cases. The second approach utilizes machine learning for statistical heatwave forecasting. Existing research and related patented technologies typically collect multi-source meteorological factors and construct machine learning models incorporating algorithms such as random forests and gradient boosting trees to predict high-temperature or heatwave events. This method, to some extent, improves the ability to characterize complex nonlinear relationships and has been applied in heatwave forecasting practices at different time and regional scales. However, these general-purpose heat wave prediction models primarily target general heat wave events and lack specific sample organization and feature engineering design for the significantly influential precursory weather scenario of "tropical cyclone approach." Specifically, tropical cyclones are mobile weather systems, and their association with heat waves in the target area has strict spatiotemporal thresholds. Directly using raw cyclone path data or global meteorological data fails to define this crucial precursory scenario of cyclone approach, resulting in the inclusion of a large amount of unrelated sample data. For example, meteorological data from when a tropical cyclone is far from the target area has no physical connection to the occurrence of a heat wave; directly inputting this data into the model severely dilutes the effective signal, leading to a significant decrease in the model's generalization ability. Therefore, general-purpose heat wave prediction models struggle to fully characterize the discriminative information directly related to the approaching cyclone, such as cyclone attributes, relative positional relationships, and possible topographic interactions. Furthermore, the dominant processes differ across geographical regions. For instance, mountainous areas may be more susceptible to topographically related warming processes, while flat areas may be more modulated by humidity and radiation conditions. General-purpose models typically employ relatively fixed factor structures, making it difficult to adaptively select key influencing factors in different regions and maintain optimal forecast performance.

[0004] In summary, existing technologies still have the following shortcomings: numerical model-based forecasting methods are limited by the model's ability to characterize cyclones and their related coupling processes, and the forecast results are quite sensitive to critical path and intensity errors; existing machine learning forecasting methods lack a dedicated modeling framework for the specific problem of "tropical cyclone approach triggering heat waves," especially lacking adaptive factor selection mechanisms for regional differences and risk output rules that can be directly used for early warning. Therefore, it is necessary to propose a technical solution for tropical cyclone approach scenarios that can automatically adapt to key influencing factors in different regions and achieve 24-hour heat wave risk forecasting and early warning triggering. Summary of the Invention

[0005] To address the shortcomings of existing heatwave forecasting methods under the influence of tropical cyclones, such as weak scenario specificity, poor cross-regional generalization performance, and difficulty in directly supporting operational early warning, this invention proposes a 24-hour forecasting method for heatwaves triggered by approaching tropical cyclones. Targeting the precursory scenario of approaching tropical cyclones, this method constructs sample selection rules constrained by physical mechanisms, an adaptive factor combination optimization framework, and a multi-objective optimization scheme for determining early warning thresholds. This achieves accurate forecasting and efficient early warning of regional heatwave risks under tropical cyclone approach scenarios. It effectively overcomes the problem that existing high-temperature / heatwave forecasting methods based on numerical weather prediction model outputs are highly sensitive to forecasting errors in tropical cyclone paths, intensity, and peripheral circulation structures in such scenarios. This improves the stability and consistency of heatwave risk identification in individual approach cases, fulfilling the operational requirement for early prevention of regional heatwave disaster chains under tropical cyclone approach scenarios.

[0006] To achieve the above-mentioned technical objectives, the technical solution adopted by the present invention is as follows:

[0007] A 24-hour forecasting method for heat waves near the triggering zone of a tropical cyclone, the method comprising the following steps:

[0008] S1. Based on the spatial range of the impact of the tropical cyclone on the target area, a proximity window is set, and the moment when the latitude and longitude of the tropical cyclone center fall within the proximity window is defined as the proximity sample time.

[0009] S2, obtain the historical best path data of tropical cyclones and the corresponding meteorological data of the target area corresponding to each approaching sample time, and extract all variable factors that have an impact on the heat wave by combining the topographic data of the target area, and use them as candidate factors to generate a candidate factor pool; divide the variable factors in the candidate factor pool into several physical categories according to their physical meaning, and each physical category includes one or more variable factors;

[0010] S3. Based on meteorological data of the target area, identify regional heat wave events within the target area. For each approaching sample time, construct a heat wave forecast label for the next 24 hours based on the identified regional heat wave event. Combine the tropical cyclone number to obtain a set of sample data. Summarize all sample data to generate a historical sample dataset.

[0011] S4. Divide the historical sample dataset into W sample subsets according to the tropical cyclone number; set the range of the number of candidate factors; for each candidate factor number k, randomly select several groups of candidate factor combinations of size k with replacement from the candidate factor pool, and select only one variable factor for each physical category. Use the candidate factor combination as the independent variable input and the heat wave forecast label as the dependent variable. Use W-1 sample subsets as the training set to train the classification model independently, and use the remaining set as the validation set to validate the prediction model. Calculate the comprehensive score of each candidate factor combination.

[0012] S5. Select the optimal number of candidate factors based on the average comprehensive score of candidate factor combinations with different numbers of candidate factors. Enumerate all possible factor combinations from the candidate factor set according to the optimal number of candidate factors, train the classification model, select the factor combination with the highest comprehensive score as the optimal factor combination for the target area, and output the corresponding classification model as the prediction model to predict the probability of a heat wave in the target area in the next 24 hours.

[0013] Step S1 further includes:

[0014] Target areas where heat waves may occur Define it as a rectangular window with latitude and longitude coordinates:

[0015] ;

[0016] In the formula, and These are the minimum and maximum latitudes of the target region R, respectively. and These are the minimum and maximum longitudes of the target area R, respectively;

[0017] Expand the target area outwards by preset latitude and longitude ranges in the south, north, west, and east directions to obtain the proximity area window. :

[0018] ;

[0019] Among them, Δlat and Δlon are the latitudinal expansion parameter and the longitude expansion parameter, respectively, which are related to the spatial range of the impact of tropical cyclones on the target area;

[0020] The moment when the latitude and longitude of the tropical cyclone center fall within the approach area window R′ is defined as the approach sample moment.

[0021] Further, in step S2, the historical optimal path data of tropical cyclones and the corresponding meteorological data of the target area are obtained for each approaching sample time. Combined with the topographic data of the target area, all variable factors that have an impact on the heat wave are extracted and used as candidate factors to generate a candidate factor pool. The variable factors in the candidate factor pool are divided into several physical categories according to their physical meaning. Each physical category includes one or more variable factors.

[0022] The optimal track data for tropical cyclones shall include at least the latitude and longitude information of the tropical cyclone center and the corresponding time; the meteorological data shall include at least 2m air temperature, total precipitation, total cloud cover, net solar radiation at the surface, net thermal radiation at the surface, latent heat flux at the surface, sensible heat flux at the surface, soil temperature, soil moisture content, planetary boundary layer height, 10m wind field data, 850hPa wind field data, and 850hPa vertical motion.

[0023] Furthermore, the physical categories include at least time-related categories, preheating condition categories, radiation condition categories, large-scale subsidence categories, soil moisture categories, pre-precipitation categories, foehn effect categories, atmospheric wind field categories, tropical cyclone attribute categories, and land surface process categories.

[0024] Furthermore, in step S3, the process of identifying regional heat wave events within the target area based on meteorological data of the target area includes the following steps:

[0025] Set a high-temperature threshold Tthr, and determine the daily maximum temperature Tmax,i(d) of each grid point or station in the target area on any date d, defining an indicator variable. :

[0026] ;

[0027] According to the indicator variable Calculate the proportion of high-temperature grid points within the target area and use it as the high-temperature coverage rate within the target area. :

[0028] ;

[0029] Where M is the number of grid points within the target area;

[0030] When the high-temperature coverage rate C(d) in a region is not lower than the set percentage threshold P% for N consecutive days, the continuous process is defined as a regional heat wave event.

[0031] Furthermore, in step S3, the process of constructing heatwave forecast labels for the next 24 hours based on the identified regional heatwave events includes the following steps:

[0032] For each approach time t, if a regional heat wave event occurs in the target area R within the time interval [t, t+24h], then record the heat wave forecast label corresponding to that approach time t. The value is 1; otherwise, record the heat wave forecast label corresponding to the time t when the sample approaches. The value of is 0.

[0033] Furthermore, in step S4, the process of calculating the comprehensive score for each candidate factor combination includes:

[0034] After training the classification model for any candidate factor combination S, cross-validation by grouping by tropical cyclone number is used to obtain the cross-validation mean indices: cv_AUC_mean, cv_Accuracy_mean, cv_Brier_mean, cv_R2_mean, and the training set indices: train_AUC, train_Accuracy, train_R2.

[0035] Define Brier's skill items as:

[0036] ;

[0037] The overfitting penalty term is defined as follows:

[0038] ;

[0039] ;

[0040] ;

[0041] ;

[0042] The comprehensive score of candidate factor combination S is calculated using the following formula:

[0043] ;

[0044] In the formula, cv_AUC_mean represents the mean AUC under cross-validation, which reflects the classification model's ability to distinguish heat wave events; cv_Accuracy_mean represents the mean accuracy under cross-validation; cv_Brier_mean represents the mean Brier score under cross-validation, which measures the error of probability prediction; and cv_R2_mean represents the R-squared score under cross-validation. 2 The mean value reflects how well the classification model fits the data; train_AUC represents the AUC value on the training set; train_Accuracy represents the accuracy on the training set; and train_R² represents the R-squared value on the training set. 2value; gap_auc represents the overfitting gap of the AUC value; gap_r2 represents R value. 2 The overfitting gap is represented by the accuracy gap; gap_acc represents the overfitting gap in accuracy; and overfit_penalty represents the overall overfitting penalty term.

[0045] Step S5 further includes:

[0046] Based on different candidate factor numbers The average comprehensive score of candidate factor combinations Select the optimal number of candidate factors :

[0047] ;

[0048] In the formula, b = 1, 2, ..., B, where B is the number of randomly selected factor combinations; This represents the b-th group of factor combinations of size k randomly selected from the candidate factor pool. Representing factor combinations The overall score;

[0049] Based on the optimal number of candidate factors, enumerate all factors in the candidate factor pool F with the following numbers: Factor combinations That is, satisfying All possible combinations; train a classification model for each factor combination separately and calculate its comprehensive score. The factor combination with the highest comprehensive score is selected as the optimal factor combination for the classification model.

[0050] Furthermore, the method also includes:

[0051] S6, using a predictive model to assess the probability of a heatwave occurring in the target area within the next 24 hours. Make predictions and set early warning thresholds. When the predicted risk probability at a certain close sample time satisfies:

[0052] ;

[0053] If the target area is deemed to have a significant risk of heatwave in the next 24 hours, an early warning trigger signal will be issued; otherwise, no early warning will be triggered.

[0054] Furthermore, the warning threshold The determination process includes the following steps:

[0055] S61, Extract the predicted heatwave risk probability for all samples in the validation set. and real labels;

[0056] S62, generate a series of candidate warning thresholds in the probability interval [0,1] according to a preset step size;

[0057] S63, for each candidate warning threshold, calculate the sensitivity (to represent the proportion correctly predicted when a heat wave occurs), the specificity (to represent the proportion correctly predicted when a heat wave does not occur), the accuracy (to represent the proportion correctly predicted for all samples), and the Brier score (to represent the probability prediction error value), and calculate the Youden index for each candidate warning threshold based on the sensitivity and specificity.

[0058] S64. Based on the minimum value constraints corresponding to sensitivity, specificity and accuracy, candidate warning thresholds that do not meet the requirements are eliminated; for two candidate warning thresholds whose Youden index differs from the preset Youden index, the candidate warning threshold with the larger Brier score is eliminated.

[0059] S65, calculate the cost of missed alarms and the cost of false alarms corresponding to different candidate warning thresholds, and calculate the total cost; with the optimization objectives of maximizing the Youden index and accuracy and minimizing the total cost, construct a weighted model of Youden index, accuracy and total cost, and select the optimal candidate warning threshold from the candidate warning thresholds;

[0060] S66, put the optimal candidate warning threshold back into the historical sample data to verify whether the warning effect meets the warning requirements. If it does, directly output the optimal candidate warning threshold as the final warning threshold. If it does not, adjust the step size of the candidate threshold and return to step S62 to reselect the candidate warning threshold until the optimal candidate warning threshold that meets the warning requirements is selected as the final warning threshold.

[0061] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0062] First, the 24-hour forecasting method for heat waves triggered by approaching tropical cyclones in this invention takes "heat waves triggered by approaching tropical cyclones" as the explicit forecasting object. By delineating the approaching area window and defining the approaching sample time, it achieves precise spatiotemporal matching of sample data with the tropical cyclone-heat wave coupling mechanism. This solves the problem that existing machine learning heat wave prediction methods mainly target general heat wave events and lack scenario-specific modeling frameworks, enabling the model to fully utilize discriminative information directly related to approaching cyclones, such as cyclone attributes, relative location, and terrain interactions. Addressing the significant differences in dominant warming processes across different regions, this invention overcomes the shortcomings of generalized models with unstable generalization performance when applied across regions and the low efficiency of factor category and time scale selection relying on expert experience by classifying candidate factors according to physical categories and adaptively selecting the optimal factor combination. This achieves automatic adaptation of key influencing factors for different geographical regions.

[0063] Secondly, the 24-hour forecasting method for heat waves triggered by approaching tropical cyclones in this invention constructs a regional heat wave risk forecasting framework oriented towards approaching cyclone scenarios. By combining a comprehensive scoring evaluation system and an early warning threshold optimization method, it effectively connects model output with operational early warning needs, enabling forecast results to directly form risk assessment criteria and triggering rules that can be used for early warning issuance. Compared to the shortcomings of traditional numerical model forecasting methods, which are greatly affected by errors in tropical cyclone track and intensity forecasts, the forecast results of this invention are more stable, thus providing more reliable technical support for regional heat wave prevention under the influence of tropical cyclones. This has significant practical implications for mitigating the combined impact of the high-temperature-storm-rainstorm disaster chain. Attached Figure Description

[0064] Figure 1 This is a flowchart of the 24-hour forecasting method for heat waves in the triggering area of ​​a tropical cyclone according to the present invention;

[0065] Figure 2 This is a schematic diagram showing the temporal distribution of heat waves and tropical cyclones in a certain region of South China from 1981 to 2024.

[0066] Figure 3 This is a diagram showing the average score of each factor, with the error bars representing the standard deviation.

[0067] Figure 4 This is a schematic diagram illustrating the prediction effect of the prediction model of the present invention. Detailed Implementation

[0068] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0069] This invention discloses a 24-hour forecasting method for heat waves in areas approaching tropical cyclones, the method comprising the following steps:

[0070] S1. Based on the spatial range of the impact of the tropical cyclone on the target area, a proximity window is set, and the moment when the latitude and longitude of the tropical cyclone center fall within the proximity window is defined as the proximity sample time.

[0071] S2, obtain the historical best path data of tropical cyclones and the corresponding meteorological data of the target area corresponding to each approaching sample time, and extract all variable factors that have an impact on the heat wave by combining the topographic data of the target area, and use them as candidate factors to generate a candidate factor pool; divide the variable factors in the candidate factor pool into several physical categories according to their physical meaning, and each physical category includes one or more variable factors;

[0072] S3. Based on meteorological data of the target area, identify regional heat wave events within the target area. For each approaching sample time, construct a heat wave forecast label for the next 24 hours based on the identified regional heat wave event. Combine the tropical cyclone number to obtain a set of sample data. Summarize all sample data to generate a historical sample dataset.

[0073] S4. Divide the historical sample dataset into W sample subsets according to the tropical cyclone number; set the range of the number of candidate factors; for each candidate factor number k, randomly select several groups of candidate factor combinations of size k with replacement from the candidate factor pool, and select only one variable factor for each physical category. Use the candidate factor combination as the independent variable input and the heat wave forecast label as the dependent variable. Use W-1 sample subsets as the training set to train the classification model independently, and use the remaining set as the validation set to validate the prediction model. Calculate the comprehensive score of each candidate factor combination.

[0074] S5. Select the optimal number of candidate factors based on the average comprehensive score of candidate factor combinations with different numbers of candidate factors. Enumerate all possible factor combinations from the candidate factor set according to the optimal number of candidate factors, train the classification model, select the factor combination with the highest comprehensive score as the optimal factor combination for the target area, and output the corresponding classification model as the prediction model to predict the probability of a heat wave in the target area in the next 24 hours.

[0075] (a) Definition of target area and proximity area

[0076] Based on the spatial extent of the tropical cyclone's impact on the target area, a proximity area window is defined. The moment when the latitude and longitude of the tropical cyclone's center fall within this window is defined as the proximity sample time. The target area can be circular, rectangular, or other shapes. For ease of description, this embodiment refers to the target area where a heat wave may occur. Define it as a rectangular window with latitude and longitude coordinates:

[0077] ;

[0078] Expand the target area outwards by a certain range of latitude and longitude in the south, north, west, and east directions to obtain the proximity area window. :

[0079] ;

[0080] Among them, Δlat and Δlon are the latitudinal expansion parameter and the longitude expansion parameter, respectively. Both are configurable parameters used to represent the spatial range of the impact of a tropical cyclone on the target area.

[0081] When the latitude and longitude of the center of a tropical cyclone fall within the approach area window R′, that moment is defined as an approach sample moment.

[0082] This invention introduces a parameterized definition of a target area's latitude and longitude rectangular window and its outward-extending approach area to uniformly determine whether a tropical cyclone is approaching the target area. This avoids the applicability limitations imposed by relying on a single distance threshold or administrative division, allowing the method to flexibly adapt to different geographical regions. While maintaining computational simplicity, this approach improves the consistency and portability of characterizing the impact range of tropical cyclones.

[0083] (ii) Data acquisition and spatiotemporal alignment

[0084] To obtain the statistical patterns of heat waves triggered in target areas under the scenario of a tropical cyclone approaching, this invention first obtains historical sample data for training the prediction model.

[0085] Historical sample data includes the best historical track data of tropical cyclones at each approach time and the corresponding meteorological data of the target area, used to characterize the spatiotemporal characteristics and physical conditions of regional heat waves during the approach of tropical cyclones. When the target area has significant topographic influence, topographic data can be added.

[0086] The optimal track data for tropical cyclones must include at least the latitude and longitude of the cyclone center and the corresponding time. The meteorological data within the target area R can be sourced from station observations or reanalysis data and must be time-aligned with the tropical cyclone track data. The meteorological data must include at least: 2m air temperature, total precipitation, total cloud cover, net solar radiation at the surface, net thermal radiation at the surface, latent heat flux at the surface, sensible heat flux at the surface, soil temperature, soil moisture content, planetary boundary layer height, 10m wind field (u and v components), 850hPa wind field (u and v components), and 850hPa vertical motion. The above data must be unified to the same time step (e.g., 6 hours) and spatially clipped or aggregated to the target area R.

[0087] All variables influencing the heatwave are extracted from historical optimal tropical cyclone track data, meteorological data of the target area, and topographic data of the target area. Statistics for each variable are calculated within different time windows (e.g., 6 hours, 24 hours, 48 ​​hours), forming a candidate factor pool F. The variables in the candidate factor pool are then categorized into several physical categories based on their physical meaning. A feature vector is constructed for each approach sample time. These physical categories include at least time-related categories, preheating conditions (e.g., temperature background characteristics before approach), radiation conditions, large-scale subsidence, soil moisture, previous precipitation, foehn effect, atmospheric wind field, tropical cyclone attributes (e.g., relative position, movement characteristics, intensity), and land surface processes.

[0088] (III) Identification of Regional Heat Wave Events

[0089] Within the target area R, regional heat wave events are identified according to the following rules:

[0090] Set a high temperature threshold Tthr, which can be an absolute threshold, a percentile threshold, or a combination of both.

[0091] For each grid point (or station) within the region, determine the daily maximum temperature Tmax,i(d) on a given date d, and define an indicator variable. :

[0092] ;

[0093] The proportion of high-temperature grid points within the target area is calculated and used as the high-temperature coverage rate of the target area. :

[0094] ;

[0095] Where M is the number of grid points (or equivalent area units) within the region.

[0096] When the high-temperature coverage rate C(d) in a region is not lower than the set threshold P% for N consecutive days, the continuous process is defined as a regional heat wave event.

[0097] (iv) Construction of 24-hour forecast labels

[0098] For each time t when approaching a sample, construct a heat wave forecast label for the next 24 hours.

[0099] If a regional heatwave event occurs in the target area R within the time interval [t, t+24h], then record a heatwave forecast label. Otherwise, remember the heatwave forecast label. .

[0100] (v) Automated factor selection and determination of optimal combination

[0101] To achieve reproducibility and statistical robustness in the factor selection process, this invention uses a historical near-sample dataset and employs the following deterministic rules to determine the number of model factors, factor categories, and final factor combinations. The model is a data-driven prediction model that uses candidate factor combinations as independent variables and the predicted labels of corresponding samples as dependent variables. It learns the statistical or nonlinear mapping relationship between independent and dependent variables by training on historical sample data, and outputs the corresponding dependent variable prediction result or risk probability when new independent variable data is input.

[0102] (1) Define the overall score

[0103] After training the model for any candidate factor combination S (containing k factors), cross-validation is used to obtain the cross-validation mean indices: cv_AUC_mean, cv_Accuracy_mean, cv_Brier_mean, and cv_R2_mean, and the training set indices: train_AUC, train_Accuracy, and train_R2 are obtained.

[0104] Define Brier skill items:

[0105] ;

[0106] Define an overfitting penalty term (penalizing only the portion of training that is significantly better than cross-validation):

[0107] ;

[0108] ;

[0109] ;

[0110] ;

[0111] Based on this, candidate factor combinations are defined. Overall score:

[0112] ;

[0113] The cross-validation method used in this invention adopts a cross-validation approach based on tropical cyclone numbers. All approaching samples are divided into several non-overlapping sample groups according to their corresponding tropical cyclone numbers. During the cross-validation process, one group of samples is selected as the validation set each time, and the remaining samples are used as the training set for model training and evaluation. This ensures that multiple samples corresponding to the same tropical cyclone process will not appear in both the training set and the validation set at the same time, thereby avoiding overfitting problems caused by time correlation and process correlation.

[0114] (2) Determine the number of candidate factors k

[0115] Let the number of candidate factors range from k ∈ {3, 4, ..., 10}. For each k, randomly select B groups of factor combinations of size k from the candidate factor pool. Training and evaluation are performed. The random selection process here can be implemented using a fixed random seed or a pre-defined combination sequence to ensure the reproducibility of the factor number determination process. The average comprehensive score corresponding to the number of candidate factors k is defined as:

[0116] ;

[0117] Select the candidate factors with the highest average composite score. As the number of model factors.

[0118] (3) Exhaustive enumeration of factors based on categories to select the optimal factor combination

[0119] After determining the number of model factors as Then, enumerate all factors in the candidate factor pool F with the number of factors being... Factor combinations That is, satisfying All possible combinations; train the model for each candidate combination separately and calculate its comprehensive score (Score). Finally, the factor combination with the highest comprehensive score is selected as the optimal factor combination of the model and used for business prediction. In order to improve computational efficiency and avoid the repeated representation of the same physical process, the factors of the same physical category can be limited to at most one in any candidate combination.

[0120] This invention constructs a candidate factor pool categorized by physical processes and systematically addresses the issues of significant differences in dominant warming processes across different regions and the reliance on expert experience for factor selection through automated factor number determination, factor category constraints, and exhaustive screening mechanisms. This method can adaptively screen key influencing factors based on regional characteristics, improving the model's generalization performance in cross-regional applications.

[0121] (vi) Model training and early warning issuance

[0122] Based on a determined optimal combination of factors, a classification model is trained that can output the probability of a heat wave occurring in a target area within the next 24 hours, based on current meteorological conditions.

[0123] ;

[0124] The model training is based on historical approach samples; during operational phases, when a new tropical cyclone approaches the target area, corresponding meteorological factors are acquired in real time as model input, and the probability of heatwave risk for the next 24 hours is output. ,

[0125] Set early warning threshold When the predicted risk probability at a certain close sample time satisfies:

[0126] ;

[0127] If the target area is deemed to have a significant risk of heatwave in the next 24 hours, an early warning trigger signal will be issued; otherwise, no early warning will be triggered.

[0128] Warning threshold This can be determined using validation set methods, such as selecting a threshold that optimizes the prediction effect or business cost. Below is a screening process for one type of warning threshold, including the following steps:

[0129] S61, Extract the predicted heatwave risk probability for all samples in the validation set. and real labels;

[0130] S62, generate a series of candidate warning thresholds in the probability interval [0,1] according to a preset step size;

[0131] S63, for each candidate warning threshold, calculate the sensitivity (to represent the proportion correctly predicted when a heat wave occurs), the specificity (to represent the proportion correctly predicted when a heat wave does not occur), the accuracy (to represent the proportion correctly predicted for all samples), and the Brier score (to represent the probability prediction error value), and calculate the Youden index for each candidate warning threshold based on the sensitivity and specificity.

[0132] S64. Based on the minimum value constraints corresponding to sensitivity, specificity and accuracy, candidate warning thresholds that do not meet the requirements are eliminated; for two candidate warning thresholds whose Youden index differs from the preset Youden index, the candidate warning threshold with the larger Brier score is eliminated.

[0133] S65, calculate the cost of missed alarms and the cost of false alarms corresponding to different candidate warning thresholds, and calculate the total cost; with the optimization objectives of maximizing the Youden index and accuracy and minimizing the total cost, construct a weighted model of Youden index, accuracy and total cost, and select the optimal candidate warning threshold from the candidate warning thresholds;

[0134] S66, put the optimal candidate warning threshold back into the historical sample data to verify whether the warning effect meets the warning requirements. If it does, directly output the optimal candidate warning threshold as the final warning threshold. If it does not, adjust the step size of the candidate threshold and return to step S62 to reselect the candidate warning threshold until the optimal candidate warning threshold that meets the warning requirements is selected as the final warning threshold.

[0135] Example

[0136] This example uses a region in South China as an example. Those skilled in the art can extend and apply it to other regions affected by tropical cyclones without departing from the technical concept of this invention.

[0137] (1) Selecting the research area:

[0138] First, a region in South China was selected as the target area R, with its spatial range defined as 21°–25°N and 108°–118°E. This region covers the main high-temperature-prone areas of South my country, characterized by frequent tropical cyclone activity, complex topography, and significant heat wave impact. Considering the scale and influence range of the outer wind circle of a tropical cyclone, outward expansion parameters Δlat=5° and Δlon=10° were set to obtain the cyclone approach area R′, with its spatial range of 16°–30°N and 98°–128°E. Through these settings, when the center of a tropical cyclone enters the R′ region, it can be considered that the cyclone has the potential to influence the target area R.

[0139] (2) Data acquisition:

[0140] This example requires data spanning the warm season (May-September) from 1981 to 2024. Data from 1981 to 2015 is used for model training, and data from 2016 to 2024 is used for independent prediction testing to simulate a formal operational environment. The following two types of data are obtained: 1. Tropical cyclone track information: The CMA optimal track dataset provided by the China Meteorological Administration Tropical Cyclone Data Center is used. This dataset contains information on the location, intensity, and timing of tropical cyclone centers, accurately depicting the evolution of tropical cyclones. 2. Meteorological background data were obtained from the European Centre for Medium-Range Weather Forecasts (ECMWF) Generation 5 Reanalysis data. Hourly data were selected for the target area R (20°N–27°N, 106°E–120°E) during the warm season (May–September) from 1981 to 2024. These data included variables such as 2m air temperature, total precipitation, total cloud cover, net solar radiation, net thermal radiation, latent heat flux, sensible heat flux, soil temperature, soil moisture content, planetary boundary layer height, u and v components of the 10m wind field, u and v components of the 850hPa wind field, and 850hPa vertical motion.

[0141] (3) Define a heat wave event:

[0142] Within the target area R, regional heat wave events are identified according to the following rules:

[0143] High temperature threshold setting: A combination of absolute and relative thresholds is used. Grid points with a daily maximum temperature (Tmax) ≥ 33℃ and a daily minimum temperature (Tmin) ≥ 26℃, and which simultaneously exceed the 75th percentile threshold for the same period in history (1981-2010), are defined as high temperature grid points.

[0144] Regional heat wave determination: Calculate the proportion of high-temperature grid points C(d) in the target area. When C(d) exceeds 15% for 3 consecutive days or more, the continuous process is defined as a regional heat wave event. Figure 2The temporal distribution characteristics of heat waves and tropical cyclones in the target area from 1981 to 2024 are shown, indicating a clear correlation between the two.

[0145] (4) Construction of close-up samples and generation of forecast labels:

[0146] By traversing all 6-hour best track records of tropical cyclones in the Northwest Pacific officially released by the China Meteorological Administration from 1981 to 2024, the 6-hour time when the latitude and longitude of the tropical cyclone center falls within the approach area R' is recorded as the approach sample time t. In this embodiment, a total of 9241 approach sample times were extracted.

[0147] For each approaching sample time t, determine whether the target area R will experience a regional heat wave event as defined in step (3) within the next 24-hour time window [t, t+24h]. If it does, assign a positive sample label y(t)=1; otherwise, assign a negative sample label y(t)=0. In this embodiment, the final number of positive samples is 1043 and the number of negative samples is 8198, with a positive-to-negative sample ratio of approximately 7.9:1.

[0148] (5) Construction of candidate factor pool

[0149] Based on the acquired ERA5 reanalysis data, this embodiment constructs 41 candidate factors covering 10 types of physical processes for each approach sample time t, forming a candidate factor pool F. All factors are averaged within the target region R, with some terrain-sensitive factors calculated in specific sub-regions. The factor naming convention is: [Variable]_[Time Window]_[Statistic]_[Processing Method]_[Region], and the specific configuration is shown in Table 1.

[0150] Table 1:

[0151] The abbreviations in the factor names in Table 1 have the following meanings: month indicates the month sequence factor of the month to which the sample time belongs; T2m indicates the air temperature at 2 meters above the ground, used to characterize the regional background thermal state; STL1 indicates the temperature of the first soil layer, used to characterize the thermal condition of shallow soil; SWdown indicates the downward shortwave radiation flux at the surface; Rnet indicates the net surface radiation, used to reflect the surface energy balance characteristics; OMEGA850 indicates the vertical velocity of the 850 hPa pressure layer, used to characterize the large-scale subsidence or ascent in the middle and lower layers; SM_L1 indicates the soil moisture of the first layer, used to reflect the land surface moisture conditions; PR indicates precipitation, used to characterize the cumulative impact of previous precipitation; V10_nanling indicates the meridional component of the wind speed at 10 meters above the ground in the Nanling Mountains, used to characterize the topographically related downslope wind (foehn wind) effect; V850 indicates the meridional component of the atmospheric wind field at the 850 hPa pressure layer; tc_azimuth_from_SC_deg The azimuth of the tropical cyclone relative to the target area is represented by: move_speed_ms; vmax_ms; vmax_ms; Bowen ratio (the ratio of sensible heat flux to latent heat flux); and EF (evaporation fraction), which represents the proportion of latent heat flux in the distribution of surface energy. The time windows in the factor names (e.g., 12h, 24h, 1d, 5d, 10d) represent historical time windows looking backward from the sample time t. Statistical calculations for the time windows use left-closed, right-open intervals. Statistics include the mean or accumulator. The "raw" in the factor name indicates that the data used are raw physical quantity data, without anomaly reduction, standardization, or normalization. R represents the target area affected by the heat wave. SC in the factor name represents South China, corresponding to the study area R; nanling represents the Nanling region, corresponding to the Nanling sub-region that characterizes the topographic effect. Its spatial range can be limited to longitude 110°E–115°E and latitude 23.5°N–25°N.

[0152] (6) Automated factor selection and determination of optimal combination

[0153] (6.1) Model Training and Evaluation Framework: This embodiment uses the XGBoost classifier as the base learning model, but the invention is not limited to this model. Its core hyperparameters are set as follows: number of trees n_estimators=400, maximum depth max_depth=3, learning rate=0.02, and subsample rate subsample=0.7. A grouped cross-validation strategy based on tropical cyclone numbers is introduced during model training and evaluation: all samples are divided into several groups according to the cyclone number (10 groups in this example). During cross-validation, one group is selected sequentially as the validation set, and the remaining groups are used as the training set. This ensures that multiple samples corresponding to the same tropical cyclone do not appear simultaneously in the training and validation sets, thereby reducing the risk of overfitting due to spatiotemporal correlation. For each validation set, four evaluation metrics are calculated: AUC, Accuracy, Brier score, and R².

[0154] (6.2) Determining the number of candidate factors k: Let the range of the number of candidate factors be k∈{3,4,…,10}. For each k value, randomly select B=500 factor combinations of size k from the candidate factor pool F with replacement. Train the XGBoost model independently for each combination and calculate the comprehensive score Score(S) according to the cross-validation strategy described above. For each k value, calculate the average score of its 500 samplings. .like Figure 3 As shown, the average composite score reaches a peak of 0.763 when k=6, significantly higher than other factor numbers (0.750 when k=5, 0.755 when k=7), indicating that the 6-factor model achieves the optimal balance between predictive performance and model simplicity. Therefore, the optimal number of factors is determined to be k*=6.

[0155] (6.3) Obtaining the optimal factor combination: Under the constraint of k*=6, enumerate all possible 6-factor combinations from the 41 candidate factors (approximately 4.5×10). 6To improve computational efficiency and ensure the physical independence of factors, this embodiment imposes the following constraints: at most one factor from the same time window and the same physical category can be selected. For example, T2m_24h_mean_raw_SC and T2m_48h_mean_raw_SC cannot be selected simultaneously; at most one factor can be selected from the preheating condition category. This constraint reduces the number of combinations to approximately 50,000, which can be completed within 48 hours on a conventional high-performance computing platform. Finally, the factor combination with the highest comprehensive score is selected as the optimal factor combination for the model. The optimal factor combination determined in this embodiment is: month order, "T2m_24h_mean_raw_SC", "OMEGA850_12h_mean_raw_SC", "SM_L1_7d_mean_raw_SC", "V10_nanling_72h_mean_raw_nanling", and "EF_24h_raw_SC". This combination has a cross-validation AUC of 0.927, an accuracy of 91.0%, and a Brier score of 0.062, demonstrating excellent predictive performance. Its physical significance is clear: the monthly sequence represents the seasonal background, the 24-hour average temperature reflects recent thermal accumulation, the 12-hour average 850 hPa subsidence represents the cyclone's peripheral dynamic forcing, the 7-day soil moisture represents the land surface moisture memory effect, the 72-hour meridional wind in the Lingnan mountainous area represents the topographic foehn effect, and the 24-hour evapotranspiration fraction represents the land surface energy distribution state. This combination fully reflects the physical mechanism of multi-scale, multi-process coupling under the scenario of a tropical cyclone approaching.

[0156] (7) Model training and early warning release:

[0157] During the operational phase, the classification model is fully trained based on the aforementioned optimal factor combination. When a new tropical cyclone enters the approaching area R′, the meteorological factors at the corresponding time are acquired in real time as model input, and the model outputs the probability of a heat wave occurring in the target area within the next 24 hours, with a value ranging from [0,1]. In this embodiment, the warning threshold... A threshold of 0.35 is preferred, as this threshold can significantly reduce the false alarm rate under low false alarm rate constraints in historical sample validation; in other implementations, It can be adjusted according to the business side's need to balance false alarms and missed alarms.

[0158] When the predicted risk probability is greater than or equal to 0.35, the target area is deemed to have a high risk of heat wave in the next 24 hours, and an early warning signal is triggered; otherwise, no early warning is triggered.

[0159] The overall prediction results of the model are shown in the attached figure. Figure 4As shown. Under the warning threshold conditions, the model was validated by simulating real-world business scenarios during the independent forecast period (2016-2024). The model demonstrated stable and effective predictive capabilities under the approaching tropical cyclone scenario. Statistical results show that when the warning threshold is reached... When the coefficient of performance (COP) is 0.35, the model achieves a hit rate of 88.2% for actual cyclone processes that trigger heat waves, with a false alarm rate of 15.6%. This demonstrates that the method of this invention can provide reliable early warning of regional heat wave risks during operational phases and has significant practical application value.

[0160] This invention combines the heatwave risk probability output by the classification model with the early warning threshold to form clear early warning triggering rules, enabling the prediction results to be directly used for operational early warning issuance. This method can be extended to various regions affected by tropical cyclones, providing a stable and implementable technical solution for the early prevention of regional heatwave disaster chains under the influence of tropical cyclones.

[0161] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0162] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A 24-hour forecasting method for heat waves in areas approaching tropical cyclones, characterized in that, The method includes the following steps: S1. Based on the spatial range of the impact of the tropical cyclone on the target area, a proximity window is set, and the moment when the latitude and longitude of the tropical cyclone center fall within the proximity window is defined as the proximity sample time. S2, obtain the historical best path data of tropical cyclones and the corresponding meteorological data of the target area corresponding to each approaching sample time, and extract all variable factors that have an impact on the heat wave by combining the topographic data of the target area, and use them as candidate factors to generate a candidate factor pool; divide the variable factors in the candidate factor pool into several physical categories according to their physical meaning, and each physical category includes one or more variable factors; S3. Based on meteorological data of the target area, identify regional heat wave events within the target area. For each approaching sample time, construct a heat wave forecast label for the next 24 hours based on the identified regional heat wave event. Combine the tropical cyclone number to obtain a set of sample data. Summarize all sample data to generate a historical sample dataset. S4. Divide the historical sample dataset into W sample subsets according to the tropical cyclone number; set the range of the number of candidate factors; for each candidate factor number k, randomly select several groups of candidate factor combinations of size k with replacement from the candidate factor pool, and select only one variable factor for each physical category. Use the candidate factor combination as the independent variable input and the heat wave forecast label as the dependent variable. Use W-1 sample subsets as the training set to train the classification model independently, and use the remaining set as the validation set to validate the prediction model. Calculate the comprehensive score of each candidate factor combination. S5. Select the optimal number of candidate factors based on the average comprehensive score of candidate factor combinations with different numbers of candidate factors. Enumerate all possible factor combinations from the candidate factor set according to the optimal number of candidate factors, train the classification model, select the factor combination with the highest comprehensive score as the optimal factor combination for the target area, and output the corresponding classification model as the prediction model to predict the probability of a heat wave in the target area in the next 24 hours.

2. The 24-hour forecasting method for heat waves in areas triggered by approaching tropical cyclones according to claim 1, characterized in that, Step S1 further includes: Target areas where heat waves may occur Define it as a rectangular window with latitude and longitude coordinates: ； In the formula, and These are the minimum and maximum latitudes of the target region R, respectively. and These are the minimum and maximum longitudes of the target area R, respectively; Expand the target area outwards by preset latitude and longitude ranges in the south, north, west, and east directions to obtain the proximity area window. : ； Among them, Δlat and Δlon are the latitudinal expansion parameter and the longitude expansion parameter, respectively, which are related to the spatial range of the impact of tropical cyclones on the target area; The moment when the latitude and longitude of the tropical cyclone center fall within the approach area window R′ is defined as the approach sample moment.

3. The 24-hour forecasting method for heat waves in areas triggered by approaching tropical cyclones according to claim 1, characterized in that, In step S2, the historical best path data of tropical cyclones and the corresponding meteorological data of the target area are obtained at each approach time. Combined with the topographic data of the target area, all variable factors that have an impact on the heat wave are extracted and used as candidate factors to generate a candidate factor pool. The variable factors in the candidate factor pool are divided into several physical categories according to their physical meaning. Each physical category includes one or more variable factors. The optimal track data for tropical cyclones shall include at least the latitude and longitude information of the tropical cyclone center and the corresponding time; the meteorological data shall include at least 2m air temperature, total precipitation, total cloud cover, net solar radiation at the surface, net thermal radiation at the surface, latent heat flux at the surface, sensible heat flux at the surface, soil temperature, soil moisture content, planetary boundary layer height, 10m wind field data, 850hPa wind field data, and 850hPa vertical motion.

4. The 24-hour forecasting method for heat waves in areas triggered by approaching tropical cyclones according to claim 1, characterized in that, The physical categories include at least time-related categories, preheating conditions, radiation conditions, large-scale subsidence, soil moisture, pre-precipitation, foehn effect, atmospheric wind field, tropical cyclone attributes, and land surface processes.

5. The 24-hour forecasting method for heat waves in areas triggered by approaching tropical cyclones according to claim 1, characterized in that, Step S3, the process of identifying regional heat wave events within the target area based on meteorological data of the target area, includes the following steps: Set a high-temperature threshold Tthr, and determine the daily maximum temperature Tmax,i(d) of each grid point or station in the target area on any date d, defining an indicator variable. : ； According to the indicator variable Calculate the proportion of high-temperature grid points within the target area and use it as the high-temperature coverage rate within the target area. : ； Where M is the number of grid points within the target area; When the high-temperature coverage rate C(d) in a region is not lower than the set percentage threshold P% for N consecutive days, the continuous process is defined as a regional heat wave event.

6. The 24-hour forecasting method for heat waves in areas triggered by approaching tropical cyclones according to claim 1, characterized in that, In step S3, the process of constructing heatwave forecast labels for the next 24 hours based on the identified regional heatwave events includes the following steps: For each approach time t, if a regional heat wave event occurs in the target area R within the time interval [t, t+24h], then record the heat wave forecast label corresponding to that approach time t. The value is 1; otherwise, record the heat wave forecast label corresponding to the time t when the sample approaches. The value of is 0.

7. The 24-hour forecasting method for heat waves in areas triggered by approaching tropical cyclones according to claim 1, characterized in that, Step S4, the process of calculating the comprehensive score for each candidate factor combination, includes: After training the classification model for any candidate factor combination S, cross-validation by grouping by tropical cyclone number is used to obtain the cross-validation mean indices: cv_AUC_mean, cv_Accuracy_mean, cv_Brier_mean, cv_R2_mean, and the training set indices: train_AUC, train_Accuracy, train_R2. Define Brier's skill items as: ； The overfitting penalty term is defined as follows: ； ； ； ； The comprehensive score of candidate factor combination S is calculated using the following formula: ； In the formula, cv_AUC_mean represents the mean AUC under cross-validation, which reflects the classification model's ability to distinguish heat wave events; cv_Accuracy_mean represents the mean accuracy under cross-validation; cv_Brier_mean represents the mean Brier score under cross-validation, which measures the error of probability prediction; and cv_R2_mean represents the R-squared score under cross-validation. 2 The mean value reflects how well the classification model fits the data; train_AUC represents the AUC value on the training set; train_Accuracy represents the accuracy on the training set; and train_R² represents the R-squared value on the training set. 2 value; gap_auc represents the overfitting gap of the AUC value; gap_r2 represents R value. 2 The overfitting gap is represented by the accuracy gap; gap_acc represents the overfitting gap in accuracy; and overfit_penalty represents the overall overfitting penalty term.

8. The 24-hour forecasting method for heat waves in the triggering area of ​​a tropical cyclone according to claim 1, characterized in that, Step S5 further includes: Based on different candidate factor numbers The average comprehensive score of candidate factor combinations Select the optimal number of candidate factors k*: ； In the formula, B is the number of randomly selected factor combinations. This represents the b-th group of factor combinations of size k randomly selected from the candidate factor pool. Representing factor combinations The overall score; Based on the optimal number of candidate factors, enumerate all factor combinations with a factor count of k* in the candidate factor pool F. For each factor combination, a classification model is trained separately and its comprehensive score is calculated. The factor combination with the highest comprehensive score is selected as the optimal factor combination for the classification model.

9. The 24-hour forecasting method for heat waves in areas triggered by approaching tropical cyclones according to claim 1, characterized in that, The method further includes: S6, using a predictive model to assess the probability of a heatwave occurring in the target area within the next 24 hours. Make predictions and set early warning thresholds. When the predicted risk probability at a certain close sample time satisfies: ； If the target area is deemed to have a significant risk of heatwave in the next 24 hours, an early warning trigger signal will be issued; otherwise, no early warning will be triggered.

10. The 24-hour forecasting method for heat waves in areas triggered by approaching tropical cyclones according to claim 1, characterized in that, The warning threshold The determination process includes the following steps: S61, Extract the predicted heatwave risk probability for all samples in the validation set. and real labels; S62, generate a series of candidate warning thresholds in the probability interval [0,1] according to a preset step size; S63, for each candidate warning threshold, calculate the sensitivity (to represent the proportion correctly predicted when a heat wave occurs), the specificity (to represent the proportion correctly predicted when a heat wave does not occur), the accuracy (to represent the proportion correctly predicted for all samples), and the Brier score (to represent the probability prediction error value), and calculate the Youden index for each candidate warning threshold based on the sensitivity and specificity. S64. Based on the minimum value constraints corresponding to sensitivity, specificity and accuracy, candidate warning thresholds that do not meet the requirements are eliminated; for two candidate warning thresholds whose Youden index differs from the preset Youden index, the candidate warning threshold with the larger Brier score is eliminated. S65, calculate the cost of missed alarms and the cost of false alarms corresponding to different candidate warning thresholds, and calculate the total cost; with the optimization objectives of maximizing the Youden index and accuracy and minimizing the total cost, construct a weighted model of Youden index, accuracy and total cost, and select the optimal candidate warning threshold from the candidate warning thresholds; S66, put the optimal candidate warning threshold back into the historical sample data to verify whether the warning effect meets the warning requirements. If it does, directly output the optimal candidate warning threshold as the final warning threshold. If it does not, adjust the step size of the candidate threshold and return to step S62 to reselect the candidate warning threshold until the optimal candidate warning threshold that meets the warning requirements is selected as the final warning threshold.

Images