Composite heat wave dynamic risk prediction method, system, equipment and medium

Abstract

The invention relates to a composite heat wave dynamic risk prediction method, system and device and a medium. The method comprises the steps of obtaining time sequence observation data and prediction time lag of a target area, and performing preprocessing to obtain standardized grid point time sequence meteorological data; performing change feature extraction according to the standardized grid point time sequence meteorological data, and determining a dry and wet partition label of the target area according to the dryness index; performing feature splicing on the obtained dry and wet partition labels, daily potential evapotranspiration, daily standardized rainfall evapotranspiration indexes, threshold temperature, a multivariable change feature matrix and grid point static features to obtain feature tensors; inputting the feature tensor into a pre-trained composite heat wave risk prediction model to obtain a prediction result; and on the basis of the physical constraint base line, probability calibration, physical consistency filtering and risk grading are sequentially carried out on the prediction result to obtain a risk probability thermodynamic diagram. By adopting the method, a more accurate dynamic risk prediction result can be generated according to the composite heat wave evolution difference of different dry and wet zones.

Description

Technical Field

[0001] This invention belongs to the field of climate disaster prediction technology, and in particular relates to a method, system, equipment and medium for predicting dynamic risks of complex heat waves. Background Technology

[0002] With the development of climate change and high-resolution observation technology, spatiotemporal statistical and machine learning methods have emerged for the analysis and prediction of regional heat waves and drought complex events. These methods extract multivariate features at the raster scale through statistical regression or machine learning and make probabilistic or point value predictions of future short-term / medium-term climate risks.

[0003] In traditional techniques, modeling and predicting the combined risks of heat waves and droughts typically involves first identifying heat wave and drought events from historical observation and reanalysis data, and then establishing regression / classification models of occurrence probability or intensity based on statistical methods or global machine learning models.

[0004] However, most existing models fail to fully reflect the significant differences in the occurrence mechanism, frequency, intensity, and duration of combined dry and wet heat waves (CDHW) and combined wet heat waves (CWHW) in terms of parameter sharing or structural design. Using a single model parameter or a uniform training process to process all grid points can easily average the specific dynamics of different regions, thereby reducing the recall ability and intensity estimation accuracy of rare but high-impact events. Summary of the Invention

[0005] Therefore, it is necessary to provide a method, system, device, and medium for predicting the dynamic risk of composite heat waves by utilizing time-series characterization specific to dry and wet zones to provide more accurate risk prediction for different dry and wet zones, addressing the aforementioned technical problems.

[0006] Firstly, this application provides a method for predicting the dynamic risk of complex heat waves, including:

[0007] The time-series observation data and prediction time lag of the target area are acquired, and the time-series observation data are preprocessed to obtain standardized gridded time-series meteorological data; the time-series observation data includes daily maximum temperature, daily average temperature, daily minimum temperature, daily precipitation and timestamp;

[0008] Based on historical window cache meteorological data, the variation characteristics of each grid point are extracted according to standardized grid time-series meteorological data to obtain daily potential evapotranspiration, daily standardized precipitation evapotranspiration index, threshold temperature and multivariate variation characteristic matrix; historical window cache meteorological data includes historical grid time-series meteorological data and climate baseline data.

[0009] The dry and wet zone labels of the target area are determined based on the aridity index. The dry and wet zone labels, daily potential evapotranspiration, daily standardized precipitation evapotranspiration index, threshold temperature, multivariate variation feature matrix and grid static features are then concatenated according to a preset sliding window and prediction time lag to obtain the feature tensor. The grid static features include topographic elevation and land cover. The dry and wet zone labels include arid zone, semi-arid zone, semi-humid zone and humid zone.

[0010] The feature tensor is input into a pre-trained composite heat wave risk prediction model to obtain the prediction results. The composite heat wave risk prediction model includes a sequentially connected partition-aware temporal encoder, a spatial aggregation layer, a cross-regional attention layer, and a multi-task output head. The prediction results include the probability of composite heat wave events occurring at each grid point on the prediction time delay, the expected intensity of the prediction time delay, the number of days the prediction time delay lasts, and an uncertainty estimate. The probability of composite heat wave events occurring is either the probability of composite drought heat wave or the probability of composite wet heat wave.

[0011] Based on the physical constraint baseline, the prediction results are sequentially subjected to probability calibration, physical consistency filtering, and risk classification to obtain a risk probability heatmap.

[0012] In one embodiment, preprocessing of time-series observation data yields standardized gridded time-series meteorological data, including:

[0013] Outlier detection, time alignment, and missing value imputation are performed on time-series observation data to obtain gridded time-series raw data; missing value imputation is based on long-term climate statistical characteristics.

[0014] Based on a pre-defined unified grid, a bilinear interpolation algorithm is used to spatially interpolate the gridded time series raw data to obtain the gridded time series raw data; the unified grid is a unified prediction network grid used for the prediction network;

[0015] Standardized gridded time-series meteorological data are obtained by standardizing the original gridded time-series data.

[0016] In one embodiment, based on historical window cached meteorological data, the variation characteristics of each grid point are extracted according to standardized gridded time-series meteorological data to obtain daily potential evapotranspiration, daily standardized precipitation evapotranspiration index, threshold temperature, and multivariate variation feature matrix, including:

[0017] Based on the improved Hargreaves method, the daily potential evapotranspiration of each grid point is calculated daily according to the daily maximum temperature, daily average temperature, daily minimum temperature and daily precipitation.

[0018] The daily water balance is calculated based on the daily potential evapotranspiration and daily precipitation, and the cumulative water surplus or deficit is obtained by sliding accumulation of the daily water balance.

[0019] The daily standardized precipitation evapotranspiration index of each grid point is obtained by fitting the probability distribution of the cumulative water balance and standardizing it.

[0020] Based on the historical daily maximum temperature data of the climate baseline period, the same daily location for obtaining time-series observation data is used as a preset step size sliding window. The 90th percentile value is statistically calculated to obtain the threshold temperature of each grid point.

[0021] The variation characteristics of historical gridded time-series meteorological data and standardized gridded time-series meteorological data are extracted to obtain a multivariate variation characteristic matrix.

[0022] In one embodiment, the composite heatwave risk prediction model obtains the prediction results through the following method:

[0023] According to the dry and wet partition labels, the feature tensor is input into the corresponding partition-aware temporal encoder to obtain the low-dimensional temporal representation vector of each grid point;

[0024] Based on adjacency weights derived from geographical proximity and climate similarity, a spatial aggregation representation is obtained by fusing neighborhood information into the low-dimensional temporal representation vectors of each grid point through convolution operations.

[0025] Based on the cross-regional attention mechanism, we capture the long-range teleconnections and cross-regional patterns of the spatial aggregation representation to obtain the local grid representation;

[0026] Based on the prediction time lag, the probability of occurrence, intensity regression value and duration distribution parameters of the composite heat wave event at each grid point are predicted in parallel according to the local grid point representation, and the uncertainty is estimated by using the Bayesian approximation method to obtain the prediction results.

[0027] In one embodiment, the composite heatwave risk prediction model is obtained through the following method:

[0028] Acquire meteorological datasets and extract historical data from them through sample windows to obtain data samples and prediction candidate windows; the meteorological datasets include daily maximum temperature, daily average temperature, daily minimum temperature, daily precipitation, and annual average precipitation distributed according to a preset uniform grid.

[0029] The data samples are divided into training and validation sets according to daily and monthly scales, and corresponding data labels are generated for the training and validation sets based on the data change characteristics within the prediction candidate window.

[0030] The objective function is constructed by weighting the classification loss, regression loss, partition weighted loss, physical consistency penalty term and probability calibration loss. Based on the objective function, the composite heat wave risk prediction model is trained end-to-end according to the training set to obtain the model parameters. The partition-aware temporal encoders include arid zone temporal encoders based on short-term TCN and GRU, semi-arid zone temporal encoders based on short-term TCN, medium-term transformer and fully connected fusion layer, semi-humid zone temporal encoders based on short-term TCN, long-term transformer and attention fusion, and humid zone encoders based on long-term transformer and global flat pooling layer.

[0031] The model parameters are validated using the validation set, and the final composite heat wave risk prediction model is determined based on the classification indicators, regression indicators, and calibration indicators obtained from the validation.

[0032] In one embodiment, generating corresponding data labels for the training set and validation set based on data change characteristics within the prediction candidate window includes:

[0033] Based on the preset dry and wet zone division criteria, dry and wet zone labels are obtained according to the dryness index of the grid points.

[0034] If the daily maximum temperature is greater than the threshold temperature for the corresponding day for at least 3 consecutive days in the prediction candidate window, the prediction candidate window is identified as a heat wave event label.

[0035] In the prediction candidate window, if the daily standardized precipitation evapotranspiration index of the grid point is less than the preset drought threshold, the prediction candidate window is determined to be a drought event label; if the daily standardized precipitation evapotranspiration index of the grid point is greater than the preset wet threshold, the prediction candidate window is determined to be a wet event label.

[0036] In the prediction candidate window, if both heat wave event labels and drought event labels are present, it is determined to be a combined drought heat wave; if both heat wave event labels and wet event labels are present, it is determined to be a combined wet heat wave.

[0037] Record the occurrence date, duration, cumulative heat, and cumulative SPEI index of the combined drought heat wave or combined wet heat wave to obtain an event sequence label set.

[0038] In one embodiment, based on a physical constraint baseline, the prediction results are sequentially subjected to probability calibration, physical consistency filtering, and risk classification to obtain a risk probability heatmap, including:

[0039] Based on statistical calibration, the probability of occurrence of compound heat wave events is probabilistically calibrated, and the probability of occurrence of compound heat wave events after probability calibration and the expected intensity of prediction time lag are filtered for physical consistency using daily potential evapotranspiration and daily standardized precipitation evapotranspiration index, so as to obtain prediction results with reasonable probability and intensity.

[0040] According to the preset probability risk rules, the prediction results are classified into high / medium / low risk levels, and the risk level, the probability of occurrence of the composite heat wave event after probability calibration, the expected intensity of the prediction time delay, the duration of the prediction time delay in days, and the uncertainty estimate are used to generate a rasterized heat map, thus obtaining a risk probability heat map.

[0041] Secondly, this application also provides a composite heat wave dynamic risk prediction system, including:

[0042] The observation module is used to acquire time-series observation data and prediction time lags for the target area, and to preprocess the time-series observation data to obtain standardized gridded time-series meteorological data; the time-series observation data includes daily maximum temperature, daily average temperature, daily minimum temperature, daily precipitation and timestamp;

[0043] The index module is used to extract the variation characteristics of each grid point based on historical window cache meteorological data and standardized grid time-series meteorological data to obtain daily potential evapotranspiration, daily standardized precipitation evapotranspiration index, threshold temperature and multivariate variation characteristic matrix; historical window cache meteorological data includes historical grid time-series meteorological data and climate baseline data.

[0044] The feature module is used to determine the dry and wet zone labels of the target area based on the aridity index, and to concatenate the dry and wet zone labels, daily potential evapotranspiration, daily standardized precipitation evapotranspiration index, threshold temperature, multivariate variation feature matrix, and grid static features according to a preset sliding window and prediction time lag to obtain a feature tensor; the grid static features include topographic elevation and land cover; the dry and wet zone labels include arid zone, semi-arid zone, semi-humid zone, and humid zone.

[0045] The prediction module is used to input feature tensors into a pre-trained composite heat wave risk prediction model to obtain prediction results. The composite heat wave risk prediction model includes a sequentially connected partition-aware temporal encoder, a spatial aggregation layer, a cross-regional attention layer, and a multi-task output head. The prediction results include the probability of composite heat wave events occurring at each grid point on the prediction time lag, the expected intensity of the prediction time lag, the duration of the prediction time lag in days, and an uncertainty estimate. The probability of composite heat wave events occurring is either the probability of composite drought heat wave or the probability of composite wet heat wave.

[0046] The post-processing module is used to perform probability calibration, physical consistency filtering, and risk classification on the prediction results based on the physical constraint baseline, so as to obtain a risk probability heatmap.

[0047] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the above-described composite heat wave dynamic risk prediction methods.

[0048] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of any of the above-described composite heat wave dynamic risk prediction methods.

[0049] The aforementioned methods, systems, equipment, and media for predicting the dynamic risks of complex heat waves explicitly include key elements such as daily maximum temperature, daily average temperature, daily minimum temperature, and daily precipitation in their time-series observation data. This comprehensively covers the core driving factors of both the heat and moisture sides in the formation of complex heat waves, avoiding incomplete characterization of complex events due to missing elements. By using daily potential evapotranspiration and the daily standardized precipitation evapotranspiration index, the water balance relationship between precipitation and evapotranspiration is transformed into a quantifiable index, accurately distinguishing between drought and wet states. This avoids underestimating the drought risk exacerbated by high-temperature evapotranspiration using a single precipitation index. Threshold temperatures ensure that the heat wave judgment thresholds for different dry and wet zones conform to their long-term climate background, avoiding misjudgments or omissions caused by a globally uniform threshold. The multivariate change feature matrix integrates meteorological parameter trends and composite event correlation indicators, comprehensively characterizing the long-term evolution and short-term fluctuation characteristics of complex heat waves, providing a multi-dimensional scientific basis for the type determination and risk quantification of complex heat waves. The four categories of labels—arid, semi-arid, semi-humid, and humid—describe the fundamental differences in the formation mechanisms of composite heat waves across different arid and humid regions. This provides a zoning basis for the zoning-aware temporal encoder, ensuring that the encoder can optimize feature extraction strategies for different regions. The zoning-aware encoder, spatial aggregation layer, and cross-regional attention layer comprehensively learn the evolution patterns of composite heat waves from three dimensions: temporal, spatial, and cross-regional, preserving zoning characteristics while not neglecting inter-regional correlation effects. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 This is a flowchart illustrating the dynamic risk prediction method for composite heat waves of the present invention.

[0052] Figure 2 This is a flowchart illustrating the steps of step S104.

[0053] Figure 3 This is a schematic diagram illustrating the construction process of the composite heat wave risk prediction model of the present invention;

[0054] Figure 4 This is a structural diagram of the composite heat wave dynamic risk prediction system of the present invention. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0056] In one embodiment, such as Figure 1 As shown, a method for predicting the dynamic risk of complex heat waves is provided. This embodiment illustrates the application of this method to a terminal. It is understood that this method can also be applied to a server, and further to a system including both a terminal and a server, and implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0057] S101. Obtain time-series observation data and prediction time lag for the target area, and preprocess the time-series observation data to obtain standardized gridded time-series meteorological data; the time-series observation data includes daily maximum temperature, daily average temperature, daily minimum temperature, daily precipitation and timestamp.

[0058] Indicatively, the target area refers to the specific geographical area within which the user needs to conduct dynamic risk prediction for complex heat waves. This can be a river basin, a provincial-level administrative unit, or a climate-sensitive area spanning multiple administrative regions. Time-series observation data is the fundamental data reflecting the continuous changes of meteorological elements in the target area over time. It can be ground-based meteorological station observation data, satellite remote sensing inversion data, and meteorological reanalysis datasets. Optionally, for short-term risk prediction, time-series observation data needs to be accurate to a daily scale; for long-term risk prediction such as quarterly prediction, time-series observation data can be at a monthly scale. The prediction lag refers to the time interval from the current point in time when the prediction is executed to the possible occurrence of a future complex heat wave event. For example, the prediction lag can be 3 days (the risk of a complex heat wave in the next 3 days), or it can be 7 days, 14 days, etc. The determination of the prediction lag takes into account both the typical evolution cycle of complex heat waves (e.g., most complex heat waves take 3-7 days from incubation to occurrence) and the meteorological response characteristics of the target area. For example, heat waves in arid areas develop rapidly, so the prediction lag can be shorter; heat waves in humid areas couple slowly with the humid background, so the prediction lag can be longer.

[0059] Furthermore, the time-series observation data undergoes preprocessing, including quality control, spatial unification, and standardization. Quality control removes anomalies from the raw data, including eliminating outliers exceeding physically reasonable ranges, correcting time inconsistencies caused by timestamp misalignment, and appropriately imputing missing values. Spatial unification, based on a pre-defined unified prediction grid (e.g., 0.5° × 0.5°), converts observation data from different sources and spatial scales into a unified raster format, ensuring that meteorological parameters at each grid point directly correspond to geographic coordinates. Standardization involves climate normalization of each meteorological parameter, such as standardizing based on the mean and standard deviation of the climate baseline data. This eliminates differences in parameter magnitude caused by variations in climate background at different grid points; for example, daily precipitation in arid regions is generally lower than in humid regions, and standardization prevents these magnitude differences from masking the changing trend.

[0060] S102. Based on historical window cache meteorological data, extract the change characteristics of each grid point according to the standardized grid time series meteorological data to obtain the daily potential evapotranspiration, daily standardized precipitation evapotranspiration index, threshold temperature and multivariate change characteristic matrix; historical window cache meteorological data includes historical grid time series meteorological data and climate baseline data.

[0061] Historical window cached meteorological data serves as a historical reference dataset supporting current feature extraction. Historical gridded time-series meteorological data refers to standardized gridded meteorological data of the target area over a relatively long period, used to extract long-term trends and event correlation features. The climate baseline period refers to a relatively stable 30-year period, such as 1960-1990, conforming to the climate baseline period standards recommended by the World Meteorological Organization (WMO). It is used to calculate various meteorological thresholds, such as heat wave thresholds and wet / dry state baselines, ensuring the climatic representativeness of the thresholds and avoiding threshold deviations caused by short-term climate fluctuations. Daily potential evapotranspiration (daily-PET) is a key indicator reflecting the evaporation and transpiration capacity of land surface and plants. Its magnitude directly affects the regional water supply and demand balance; a high PET means rapid water consumption, which can exacerbate drought and amplify heat wave intensity. For example, daily-PET can be calculated grid-by-grid and day-by-day based on parameters such as daily maximum temperature, daily average temperature, daily minimum temperature, and daily precipitation from standardized gridded data, combined with calculation methods suitable for data availability in the target area, such as the improved Hargreaves method and the Penman-Monteith method. The Daily Standardized Precipitation Evapotranspiration Index (daily-SPEI) is a quantitative indicator of dryness and wetness that integrates precipitation (water input) and potential evapotranspiration (water output). It can accurately characterize the degree of water supply and demand imbalance on a daily scale. A negative daily-SPEI indicates a water deficit, with smaller values ​​indicating more severe drought. A positive daily-SPEI indicates a water surplus, with larger values ​​indicating greater humidity. For example, the daily-SPEI can be calculated based on daily precipitation and daily potential evapotranspiration to determine the daily water balance. This balance is then accumulated over a sliding cumulative period, such as 30 days or 90 days, to obtain a cumulative water surplus / deficit sequence. Further, through probability distribution fitting (e.g., Logistic distribution) and standardization, the cumulative water surplus / deficit is converted into an index value conforming to a normal distribution, ultimately yielding the grid-specific daily-SPEI. Threshold temperature is a key benchmark for distinguishing between normal high temperatures and heat waves. When the daily maximum temperature at a certain grid point continuously exceeds this threshold, it can be preliminarily identified as a heat wave event. For example, the threshold temperature can be based on the historical daily maximum temperature data of the climate baseline period in the historical window cache data. Taking the same calendar day of the current time series observation data as the center, a sliding window of ±7 days is set, and the 90th percentile value of all daily maximum temperature data within the sliding window is counted to obtain the 90th percentile threshold of the daily maximum temperature of each grid point, i.e., T90,d, and thus the threshold temperature.

[0062] The multivariate variation feature matrix is ​​a high-dimensional feature set that integrates the dynamic trends of meteorological parameters with the correlation indicators of complex events. It aims to capture the dynamic evolution patterns influencing the occurrence of complex heat waves from multiple dimensions. Specifically, it includes the changing trend characteristics of meteorological parameters, such as the interannual rate of increase in daily maximum temperature, the magnitude of decrease in daily precipitation, and the increasing trend of the number of days without precipitation. These changing trend characteristics directly reflect the changes in the probability of complex heat waves occurring; for example, the faster the temperature rises, the higher the frequency of heat waves. It also includes the correlation indicators of complex events, such as the past occurrence frequency, the rate of change in duration, and the cumulative growth rate of intensity of complex heat waves in different dry and wet zones. These correlation indicators can quantify the evolution intensity of complex heat waves. For example, based on historical window cache data and current standardized grid data, the above trends and correlation indicators are statistically analyzed according to a preset time window. The multidimensional indicators of each grid point are integrated into a vector, and then arranged in grid order to form a matrix, i.e., the multivariate variation feature matrix.

[0063] For example, taking the Yangtze River Basin, which is mainly a humid region, as the target area, based on historical window cache data from 1960 to 2024 and current standardized grid data, the features of each grid point are extracted. Specifically, the improved Hargreaves method is used to calculate the daily-PET, and the daily-SPEI is calculated based on the 15-day sliding cumulative period. The daily maximum temperature of each grid point within the climate baseline data is statistically analyzed to obtain the maximum temperature dataset. The 90th percentile threshold of the maximum temperature dataset is used as the threshold temperature of 32℃. Based on the data of the past 30 years, the daily maximum temperature rise rate is 0.25℃ / 10 years, the CWHW occurrence frequency is 1.3 times / year, and the increase in the number of consecutive days is 0.26 days / year. These are integrated into a multivariate change feature matrix.

[0064] S103. Determine the dry and wet zone labels of the target area based on the aridity index, and then perform feature splicing on the dry and wet zone labels, daily potential evapotranspiration, daily standardized precipitation evapotranspiration index, threshold temperature, multivariate variation feature matrix, and grid static features according to a preset sliding window and prediction time lag to obtain the feature tensor; grid static features include topographic elevation and land cover; the dry and wet zone labels include arid zone, semi-arid zone, semi-humid zone, and humid zone.

[0065] Indicatively, the dry / wet zoning labels are category identifiers based on the aridity index (AI) of each grid point in the target area. The aridity index is the ratio of annual average precipitation to annual average potential evapotranspiration. Grid points with an aridity index ≤ 0.2 are classified as arid regions, characterized by perennial water deficit and the highest probability of heat waves coupled with drought. Grid points with an aridity index between 0.2 and 0.5 are classified as semi-arid regions, where water supply and demand fluctuate greatly, and both CDHW and CWHW are likely to occur with high intensity. Grid points with an aridity index between 0.5 and 0.65 are classified as semi-humid regions, with moderate water conditions and a lower frequency and intensity of combined heat waves compared to semi-arid regions. Grid points with an aridity index > 0.65 are classified as humid regions, with a significant water surplus and a higher probability and faster growth of CWHW. For example, the classification process requires statistical analysis and matching of the above rules for each grid point based on the aridity index in the historical window cache data, assigning a unique dry / wet zone label to each grid point, and ultimately forming a dry / wet zone label map covering the target area.

[0066] Furthermore, grid static features are geographical and surface features that do not change over time but have a significant impact on the formation and intensity of complex heat waves. These mainly include topographic elevation (DEM) and land cover. Topographic elevation reflects the altitude of the grid points and is usually obtained through digital elevation models. The impact of topographic elevation on complex heat waves is mainly reflected in the direct influence of altitude on temperature and the influence of topographic relief on precipitation distribution. For example, for every 100 meters increase in altitude, the temperature decreases by approximately 0.6℃. Higher altitude grid points have lower heat wave thresholds and weaker intensity; for instance, the frequency of heat waves in the Qinghai-Tibet Plateau region is significantly lower than in the plains region. Windward slopes of mountains receive more precipitation, easily forming a humid background and reducing the probability of CDHW (Current Distance, Temperature, and Windward Movement); leeward slopes receive less precipitation, easily exacerbating drought. Land cover reflects the type of land cover at the grid points, such as vegetation cover, urban buildings, water bodies, and bare land, and is usually obtained through the interpretation of satellite remote sensing data. Different land cover types exhibit significant differences in surface thermal properties and evapotranspiration capacity. Urban built-up areas have low surface albedo and small heat capacity, making them prone to forming heat island effects and exacerbating heat waves. Vegetated areas have strong evapotranspiration capacity and can alleviate heat waves through evaporative cooling. Water bodies have large heat capacity, which can regulate the surrounding air temperature and reduce the probability of heat waves.

[0067] Specifically, the construction of the feature tensor integrates dynamic meteorological features, static geographical features, and time-scale information into a unified dimensional input format. For example, dynamic features are extracted using a preset sliding window. The length of the sliding window needs to be set according to the feature response requirements of different arid and humid zones. For instance, a 14-30 day window is used in arid zones, focusing on short-term temperature abrupt changes, while a 60-90 day window is used in humid zones, focusing on long-term arid and humid backgrounds. Dynamic feature sequences within the corresponding window are extracted from daily potential evapotranspiration, daily standardized precipitation evapotranspiration index, threshold temperature, and multivariate variation feature matrix. Furthermore, for each preset prediction lag, the extracted dynamic feature sequences are associated with a time scale. For example, a 1-day prediction lag matches the dynamic features of the past 14 days, and a 7-day prediction lag matches the dynamic features of the past 30 days, ensuring that the time span of the dynamic features matches the time scale of the prediction lag. That is, short-term predictions require short-term features to capture abrupt changes, while medium-term predictions require long-term features to capture trends. The dynamic feature sequences matching the prediction lag, static features, and arid / humid zone labels are concatenated according to a grid-time-feature dimensional structure to obtain the final feature tensor.

[0068] S104. Input the feature tensor into the pre-trained composite heat wave risk prediction model to obtain the prediction results. The composite heat wave risk prediction model includes a sequentially connected partition-aware temporal encoder, a spatial aggregation layer, a cross-regional attention layer, and a multi-task output head. The prediction results include the probability of composite heat wave events occurring at each grid point on the prediction time delay, the expected intensity of the prediction time delay, the number of days the prediction time delay lasts, and the uncertainty estimate. The probability of composite heat wave events occurring is either the probability of composite drought heat wave or the probability of composite wet heat wave.

[0069] The overall architecture of the composite heat wave risk prediction model follows the logic of temporal feature extraction, spatial information aggregation, cross-regional correlation capture, and multi-task prediction. It consists of a sequentially connected partition-aware temporal encoder, a spatial aggregation layer, a cross-regional attention layer, and a multi-task output head. The first three layers are responsible for extracting comprehensive features from the input feature tensor, including partition-specific temporal patterns, regional spatial correlations, and cross-regional remote interactions. The last layer outputs multi-dimensional prediction results in parallel based on these comprehensive features. All components work together to ensure the accuracy and comprehensiveness of the prediction results. Illustratively, the partition-aware temporal encoder designs dedicated temporal coding sub-modules for arid, semi-arid, semi-humid, and humid regions to accommodate the different temporal evolution patterns of composite heat waves in different arid and humid zones. Specifically, the arid region submodule adopts a structure combining short-term TCN (Temporal Convolutional Network) and GRU (Gated Recurrent Unit). The TCN captures local temporal features of short-term temperature surges and heat wave onset over 14-30 days through multi-scale convolutional kernels, while the GRU performs temporal dependency modeling on the local features output by the TCN, strengthening the logic for judging heat waves with continuous high temperatures and avoiding misjudgment of isolated high temperatures. The humid region submodule adopts a structure combining long-term Transformer and global average pooling layer. The Transformer captures the temporal correlation between long-term daily-SPEI, precipitation, and heat waves over 60-90 days through a self-attention mechanism, while the global average pooling layer weakens isolated short-term fluctuations. The semi-arid and semi-humid region submodules adopt a dual-branch fusion structure, namely a combination of short-term TCN and medium / long-term Transformer, to capture short-term signals triggered by heat waves and medium / long-term trends in dry and wet backgrounds, respectively. The fusion layer then integrates the two types of features to adapt to the complex heat wave patterns driven by their hybrid nature.

[0070] The spatial aggregation layer reflects the spatial correlation between the occurrence and evolution of complex heat waves. The characteristics of a single grid point cannot reflect the coupling effect at the regional scale. For example, heat waves from adjacent grid points may spread to each other, and precipitation anomalies in upstream areas may affect the dryness and wetness of downstream areas. The spatial aggregation is realized based on the adjacency weights of geographical proximity and climate similarity. Specifically, an adjacency matrix between grid points is constructed. If the geographical distance between two grid points is less than a preset threshold, they are determined to be adjacent. Further, the climate similarity of adjacent grid points, such as the similarity of annual average precipitation and annual average temperature, is calculated. Geographical distance and climate similarity are weighted and fused to obtain the adjacency weight. That is, the closer the distance and the more similar the climate, the greater the weight. The low-dimensional temporal representation vector of each grid point is fused with neighborhood information through convolution operations or graph neural networks (GNN) to obtain a spatial aggregation representation that includes neighborhood correlation.

[0071] The combined heat waves in different arid and humid zones are not completely independent; there are long-distance teleconnections or cross-regional influences. For example, heat waves in arid zones may affect the temperature distribution in humid zones through atmospheric circulation, and precipitation anomalies in semi-humid zones may change the arid and humid background of semi-arid zones. Therefore, the cross-regional attention layer captures the long-distance correlations between different zones to avoid the loss of global information caused by the model focusing only on local zones. The implementation of cross-regional attention is based on an attention mechanism. Specifically, the spatial aggregate representation of each zone is used as a key and a value, and the spatial aggregate representation of the current grid point is used as a query. The attention weights of the query and each key are calculated. The weights reflect the degree of influence of the long-distance zone on the current grid point. For example, the influence weight of the arid zone on the humid zone is dynamically learned based on historical correlation data. The values ​​of each zone are weighted and summed based on the attention weights to obtain the local grid point representation containing long-distance cross-regional correlations.

[0072] The multi-task output head contains four parallel sub-modules: a compound heat wave event occurrence probability sub-module, which outputs the probability of compound events occurring at each grid point within the prediction time lag using a Sigmoid activation function, with values ​​ranging from 0 to 1. Specifically, the compound drought heat wave occurrence probability corresponds to the probability of heat waves and droughts occurring simultaneously, and the compound wet heat wave occurrence probability corresponds to the probability of heat waves and wetness occurring simultaneously; a prediction time lag expected intensity sub-module, which outputs the quantitative value of the expected intensity of the compound event using a linear activation function, such as the cumulative heat obtained by regression based on the intensity statistics of historical compound events and the current features; a prediction time lag duration days sub-module, which outputs the expected duration of the compound event, and can use a distribution prediction to output the probability distribution parameter of the duration days, such as the λ value of a Poisson distribution; and an uncertainty estimation sub-module, which quantifies the reliability of the prediction results using MC-Dropout (Monte Carlo dropout) or deep ensemble methods.

[0073] Furthermore, in the prediction results, the probability of a compound heat wave event is the likelihood of a compound event occurring at a certain grid point within the prediction time lag, ranging from 0 to 1, with higher values ​​indicating higher risk. The expected intensity of the prediction time lag is the average intensity expected when the compound event occurs, using accumulated heat as the core quantitative indicator. The calculation logic involves summing the daily temperature anomalies exceeding the threshold within the prediction time lag. This indicator reflects the destructive potential of the compound event; higher intensity indicates more severe impact. The duration of the prediction time lag is the number of consecutive days from the occurrence to the end of the predicted compound event, taking a positive integer value to reflect the duration of the compound event's impact. Uncertainty estimation is a quantitative assessment of the reliability of the prediction results, usually presented as a confidence interval or standard deviation. Narrower confidence intervals and smaller standard deviations indicate more reliable prediction results; conversely, caution should be exercised when considering such results.

[0074] S105. Based on the physical constraint baseline, the prediction results are sequentially subjected to probability calibration, physical consistency filtering, and risk classification to obtain a risk probability heatmap.

[0075] The physical constraint baseline is a reasonable forecasting range standard constructed based on fundamental meteorological principles, historical observation patterns, and the formation mechanism of complex heat waves. Its core function is to filter out abnormal results in model forecasts that do not conform to physical laws, ensuring the scientific validity and reliability of the forecast results. Illustratively, the physical constraint baseline includes variable correlation constraints, i.e., reasonable correlations between different meteorological variables and forecast results, such as the negative correlation constraint between daily-SPEI and CDHW probability, and the positive correlation constraint between PET and complex heat wave intensity. It also includes numerical range constraints, i.e., the physically reasonable range of values ​​for the forecast results. For example, the duration of a complex event is usually 3-15 days; if the predicted duration is >20 days, this constraint is violated. The cumulative heat of a complex heat wave intensity is usually no more than 50℃·day; if the predicted intensity is >80℃·day, it requires further verification.

[0076] Specifically, the core objective of probability calibration is to address the mismatch between model-predicted probabilities and actual occurrence frequencies. This can be achieved using temperature scaling and Platt scaling. Temperature scaling introduces a learnable temperature parameter T, divides the original logarithmic probability of the model output by T, and then applies sigmoid activation to adjust the probability distribution. For example, T > 1 can reduce probability concentration and avoid overconfidence. Platt scaling maps the original output to calibrated probabilities by fitting a logistic regression model to the original output. The calibration process needs to optimize calibration parameters based on the correspondence between predicted probabilities and actual occurrence frequencies in historical predictions to ensure that the calibrated probabilities meet the reliability curve close to the diagonal. Physical consistency filtering is a process of verifying the reasonableness and correcting anomalies of calibrated prediction results based on a physical constraint baseline. Specifically, it verifies whether the correlation between the probability of occurrence of composite heat waves, expected intensity and daily potential evapotranspiration, and daily standardized precipitation evapotranspiration index conforms to the constraint baseline, and whether the prediction lag duration and expected intensity are within a physically reasonable range. It also verifies whether there are significant abrupt changes in the prediction results of adjacent grid points. For example, if the CDHW probability of a certain grid point is 0.8, while the probability of all surrounding grid points is <0.2, it is necessary to judge whether it is reasonable in combination with topography and land cover differences. If it is unreasonable, it is adjusted to be close to the mean of the surrounding areas.

[0077] Optionally, prediction results that violate the physical constraint baseline can be corrected and marked. Minor violations are corrected to reasonable values ​​based on the constraint baseline, while serious violations are marked as low reliability and noted in subsequent risk diagrams to remind users to refer to them with caution.

[0078] Furthermore, risk grading is the process of transforming quantitative prediction results into intuitive high / medium / low risk categories. A risk probability heatmap visualizes the risk level, calibrated probability, expected intensity, duration (in days), and uncertainty in a grid map format. A risk probability heatmap typically contains multiple parallel submaps, each corresponding to different dimensions of risk information. Examples include a heatmap showing the probability of a combined heatwave, an expected intensity heatmap, a duration heatmap, and a comprehensive risk level heatmap.

[0079] In the aforementioned dynamic risk prediction method for complex heat waves, time-series observation data can fully characterize the coupling process between heat waves and extreme moisture conditions, avoiding missed detections of complex events due to the lack of certain elements, and improving the completeness of the prediction. Preprocessing converts the raw data into a standardized grid format, eliminating spatial resolution and dimensional differences between different data sources, thus improving the spatial completeness and comparability of the prediction. Threshold temperatures are calculated based on climate baseline data, avoiding overestimation / underestimation of thresholds caused by short-term climate fluctuations, ensuring the long-term stability of heat wave determination. Trend indicators in the multivariate change feature matrix are calculated based on long-term historical data, accurately reflecting the long-term evolution of complex heat waves, avoiding trend misjudgments caused by short-term data, and improving the long-term robustness of the prediction. The target area is divided into four dry and wet zones according to the aridity index, avoiding misjudgments caused by a globally uniform threshold. Simultaneously, zone labels provide the model with regional prior information, enabling the zone-aware time-series encoder to optimize feature extraction for different zones, thereby improving the prediction accuracy of each zone and avoiding prediction biases in areas with significant regional differences, as is common with traditional globally uniform models. The composite heat wave risk prediction model employs a hierarchical architecture consisting of a zonal-aware temporal encoder, a spatial aggregation layer, a cross-regional attention layer, and a multi-task output head. This architecture accurately captures the unique temporal patterns of each zonal region, enhancing the representational ability of temporal features. The spatial aggregation layer integrates neighborhood information to characterize the spatial diffusion effect of composite heat waves, avoiding the local limitations of single-grid-point predictions. The multi-task output head simultaneously outputs the probability of occurrence, expected intensity, duration, and uncertainty, meeting both the early warning needs of emergency departments and the intensity / duration response planning needs of agricultural departments, thus improving the decision-making adaptability of the prediction results. Physical constraint baselines are used to perform probability calibration and physical consistency filtering on the prediction results, ensuring that probabilistic predictions more closely reflect real risks, eliminating contradictory results that violate meteorological laws, ensuring that the output results conform to objective scientific principles, avoiding absurd predictions caused by data-driven approaches, and improving the scientific rigor and reliability of the prediction results, providing credible and usable risk information for decision-making.

[0080] In one embodiment, preprocessing of time-series observation data yields standardized gridded time-series meteorological data, including:

[0081] S11. Perform outlier detection, time alignment, and missing value imputation on the time series observation data to obtain gridded time series raw data; missing value imputation is based on long-term climate statistical characteristics.

[0082] As an example, based on fundamental meteorological principles, a physically reasonable threshold is set. Data that does not meet this threshold is labeled as anomalies, and potential outliers are screened using statistical methods. For instance, a 3... The principle is to calculate the mean of a meteorological element in a long-term series. with standard deviation If a data point exceeds If the value falls outside the 3σ range, it is considered a statistical outlier. Minor outliers, such as those slightly exceeding the 3σ range, can be marked and retained to avoid accidental deletion of extreme weather event records; severe outliers, such as those exceeding the physical threshold, are directly removed and supplemented in subsequent missing value imputation stages.

[0083] Furthermore, to determine the target time standard, the local time of the target region can be used. Data from non-target time zones can be converted to different time zones to verify the consistency of the day boundary. Specifically, hourly data that crosses the day boundary can be merged into daily-scale data according to the target day boundary to ensure that the time interval corresponding to each day is completely consistent.

[0084] Missing value imputation can be based on long-term climate statistical characteristics. For example, the contemporaneous mean imputation method, suitable for short-term missing values, can be used. Specifically, all historical observation data for the target grid point within the climate baseline period and contemporaneous with the missing date are extracted, and the mean is calculated as the imputation value. If contemporaneous data is insufficient, the contemporaneous range is expanded. Alternatively, the nearest similar day imputation method, suitable for days with complete data before and after the missing date, can be used. Specifically, similar days 3-5 days before and after the missing date are selected, and the mean values ​​of meteorological elements on these similar days are calculated as the imputation values.

[0085] S12. Based on a preset unified grid, a bilinear interpolation algorithm is used to perform spatial interpolation on the grid point time series raw data to obtain the grid point time series raw data; the unified grid is a unified prediction network grid used for the prediction network.

[0086] In a schematic manner, the spatial resolution is set to 0.5° × 0.5°. The geographic coordinate range of the unified grid must completely cover the target area and extend outward by 1-2 grid points to avoid spatial aggregation errors caused by the lack of neighboring data for grid points at the edge of the target area. The coordinate system adopts the WGS84 geographic coordinate system to ensure coordinate matching of data from different sources. A unified grid based on latitude and longitude projection is used. After the preset unified grid is determined, a grid list needs to be generated, including the unique ID, latitude and longitude coordinates, and whether it belongs to the target area for each grid point, to ensure that subsequent interpolation data can accurately match each grid point. Optionally, bilinear interpolation is used for spatial interpolation. Specifically, based on the four nearest known grid points around the unknown grid point, the unknown value is calculated through two linear interpolations. In the horizontal direction of the rectangle, i.e., the longitude direction, the known grid point values ​​on the upper and lower sides are linearly interpolated to obtain the corresponding interpolation values ​​of the unknown grid point on the upper and lower sides. Furthermore, in the vertical direction, i.e., the latitude direction, the interpolation results of the upper and lower sides are linearly interpolated again to finally obtain the value of the unknown grid point.

[0087] S13. Standardize the original gridded time series data to obtain standardized gridded time series meteorological data.

[0088] The standardization of meteorological data needs to retain information about relative climate anomalies, that is, the degree of deviation of the element value of a certain grid point on a certain day from its long-term climate average. The Z-score standardization method is chosen, which calculates the standardized value based on the statistical characteristics of the climate baseline data. ,in, Let i be the standardized value of the i-th grid point on day t. The original meteorological element values ​​for the i-th grid point on day t are... Let be the mean value of this element at the i-th grid point during the climatic baseline period. Let be the standard deviation of this element at the i-th grid point during the climatic baseline period, and be the standardized value. The positive or negative sign indicates whether the value is higher or lower than the long-term mean, and the absolute value indicates the degree of anomaly, which meets the prediction requirements of complex heat wave extreme anomaly events; each grid point and Calculations are based on long-term climate characteristics to avoid misjudging the same degree of anomaly between arid and humid regions. The climate baseline period is used for calculation. and The time base is selected from a period of relatively stable climate without significant trend changes to avoid statistical distortion caused by climate abrupt changes. For example, for each grid point i in the target area, all diurnal data of a certain meteorological element, such as daily maximum temperature, within the climate base period of 1960-1990 are extracted, and the mean is calculated. with standard deviation For spatially unified gridded time-series data, the data for day t of each grid point i is... Substitute into the Z-score formula to calculate the standardized value. .

[0089] In one embodiment, based on historical window cached meteorological data, the variation characteristics of each grid point are extracted according to standardized gridded time-series meteorological data to obtain daily potential evapotranspiration, daily standardized precipitation evapotranspiration index, threshold temperature, and multivariate variation feature matrix, including:

[0090] S21. Based on the improved Hargreaves method, the daily potential evapotranspiration of each grid point is calculated daily according to the daily maximum temperature, daily average temperature, daily minimum temperature and daily precipitation.

[0091] Indicative of daily potential evapotranspiration ,in, Daily average temperature ; The difference between the daily maximum temperature and the daily minimum temperature ; extraterrestrial radiation ; Daily precipitation Furthermore, , ,in, The solar constant can be taken as 0.0820. , This is the Earth-Sun distance correction factor. Solar declination angle , Local geographical latitude , Sunrise and sunset corners Furthermore, , , , where J represents the accumulated days of the year.

[0092] S22. Calculate the daily water balance based on the daily potential evapotranspiration and daily precipitation, and then perform a sliding accumulation of the daily water balance to obtain the cumulative water surplus or deficit.

[0093] Cumulative Water Deficit (WSD) refers to the cumulative difference between daily precipitation and daily potential evapotranspiration at a target grid point within a sliding time window. It highlights the trend of water surplus or deficit on seasonal or longer time scales. For example, based on the evolution cycle of composite heat waves and the wet and dry characteristics of the target area, a 90-day sliding window is chosen as the default length. This covers the seasonal wet and dry variations in most areas and captures the typical cycle of drought and heat wave coupling. Optionally, if the target area has a special wet and dry rhythm, such as a short rainy season in tropical regions, a 60-day window can be selected. ,in, Daily cumulative precipitation ; This represents the daily potential evapotranspiration.

[0094] S23. The daily standardized precipitation evapotranspiration index of each grid point is obtained by fitting the probability distribution of the cumulative water balance and standardizing it.

[0095] The daily standardized precipitation evapotranspiration index (daily-SPEI) is a quantitative indicator of dryness and wetness obtained by fitting and standardizing the cumulative water surplus / deficit (WSD) series to a probability distribution. This eliminates spatiotemporal heterogeneity and enables cross-regional and cross-temporal comparisons of dryness and wetness. Specifically, standardization transforms the WSD of all grid points into an index value conforming to a standard distribution, ensuring that a daily-SPEI of -1.0 indicates mild drought in both arid and humid regions, thus guaranteeing consistency in the determination of dryness and wetness. Furthermore, for the WSD time series of the target grid points, the location parameters of the Logistic distribution are solved using the maximum likelihood estimation method. and scale parameters The position parameters The central trend of the WSD series is close to the long-term average water surplus / deficit, and the scale parameter is... Reflecting the degree of dispersion of WSD, The larger the value, the more drastic the fluctuations in water balance. Furthermore, each WSD value is substituted into the fitted Logistic distribution function to calculate its corresponding cumulative probability F(x), which is the probability that the WSD is less than or equal to that value. This probability ranges from [0,1], reflecting the relative position of the current water balance in the historical sequence. For example, F(x) = 0.1 indicates that the WSD value is below 90% of the historical values, belonging to a period of significant drought. The cumulative probability F(x) is converted into a SPEI value conforming to a standard normal distribution through an inverse function. For example, when F(x) ≤ 0.5, the negative conversion reflects the degree of drought, that is, the smaller the value, the more severe the drought; when F(x) > 0.5, the positive conversion reflects the degree of humidity, that is, the larger the value, the more significant the humidity. The resulting daily-SPEI value is usually divided into extreme drought, severe drought, moderate drought, and mild drought with -2.0, -1.5, -1.0, and -0.5 as boundaries, and the humidity state is divided with 0.5 as the boundary. This classification standard provides a clear basis for determining the type of complex heat waves.

[0096] S24. Based on the historical daily maximum temperature data of the climate baseline period, the same day location for obtaining time series observation data is used as the preset step size sliding window. The 90th percentile value is statistically calculated to obtain the threshold temperature of each grid point.

[0097] Indicatively, the threshold temperature refers to the 90th percentile threshold (T90,d) of daily maximum temperature statistically analyzed by wet and dry zones. The calculation of grid point T90,d is based on the historical daily maximum temperatures during the 1960-1990 climate baseline period, employing a same-day sliding window strategy to eliminate the influence of seasonal fluctuations. Specifically, a 15-day sliding window (±7 days) is constructed centered on the target grid point calendar day. All daily maximum temperature data within this sliding window during the climate baseline period are extracted, and their 90th percentile value is calculated. This value is the T90,d for the target grid point calendar day. All grid points within the target area are categorized according to their wet and dry zone labels, ensuring that each zone contains a sufficient number of grid points. The arithmetic mean of the same-day T90,d for all grid points within each zone is calculated to obtain the threshold temperature for the target grid point calendar day. This process is repeated to obtain the threshold temperature sequence for each zone over the entire year (365 days).

[0098] S25. Extract the variation characteristics of historical gridded time-series meteorological data and standardized gridded time-series meteorological data to obtain a multivariate variation characteristic matrix.

[0099] Indicatively, the multivariate variation feature matrix contains two core features, both extracted from historical and current time windows. Among them, the meteorological parameter variation trend features reflect the long-term evolution of meteorological elements. Examples include temperature trends, such as calculating the annual average change of daily maximum temperature over the past 30 years through linear regression to obtain the interannual rate of increase of daily maximum temperature, and the increase in the heat wave threshold (T90,d). It also includes precipitation trends, such as the interannual decrease in daily precipitation and the interannual increase in the number of days without precipitation, as well as moisture trends, such as the interannual decrease in daily SPEI and the interannual increase in daily potential evapotranspiration. Optionally, the meteorological parameter variation trend features are extracted using statistical methods such as linear regression and Mann-Kendall trend test to ensure the significance and reliability of the trends. Composite event correlation indicators reflect the dynamic changes in the frequency, intensity, and duration of composite heat waves. Examples include frequency indicators, such as the annual frequency of composite drought heat waves (CDHW) and composite wet heat waves (CWHW) in different regions; duration indicators, such as the interannual growth rate of the average duration days of CDHW / CWHW; and intensity indicators, such as the interannual growth rate of the cumulative heat of CDHW / CWHW. Optionally, the frequency indicators are extracted from historical composite heat wave event records to ensure a direct correlation with composite events.

[0100] Furthermore, for each grid point, the meteorological parameter change trend characteristics and composite event correlation indicators are integrated into a feature vector. Each element in the vector corresponds to a quantitative value of a feature. If the target area contains N grid points and the feature vector dimension of each grid point is D, then the feature vectors of the N grid points are arranged in grid point order to form an N×D matrix, which is the multivariate change feature matrix.

[0101] In one embodiment, such as Figure 2 As shown, the composite heat wave risk prediction model obtains its prediction results through the following methods:

[0102] S201. According to the dry and wet partition labels, input the feature tensor into the corresponding partition-aware temporal encoder to obtain the low-dimensional temporal representation vector of each grid point.

[0103] Indicatively, the feature tensor already contains the dry / wet zone label for each grid point, such as one-hot encoding, where arid regions are [1,0,0,0] and wet regions are [0,0,0,1]. The label recognition module parses the zone affiliation of each grid point and directs the temporal feature subsequence of that grid point to the dedicated temporal encoder of the corresponding zone. The arid zone encoder targets the characteristics of the arid zone CDHW dominated by short-term heat waves. The TCN captures the local temporal patterns of temperature surges and heat wave onset within 14-30 days through 5 convolutional kernels. The GRU performs temporal dependency modeling on the local features output by the TCN, filtering out interference from isolated high temperatures, such as single-day high temperatures that do not form continuous heat waves. Finally, it outputs a 128-dimensional low-dimensional temporal vector, focusing on preserving the correlation information between short-term temperature abrupt changes and the drought background. The semi-arid region encoder addresses the hybrid driving features of heat wave triggering and the coupling of dry and wet backgrounds in semi-arid regions. A short-term TCN with two convolutional layers extracts heat wave signals over 30-45 days, while a medium-term Transformer with two self-attention layers captures the dry and wet trends of the 90-day daily-SPEI. The MLP fusion layer integrates the two types of features through weighted summation, such as a heat wave signal weight of 0.4 and a dry and wet trend weight of 0.6, outputting a 256-dimensional vector while preserving the temporal correlation between the short-term and medium-term. The semi-humid region encoder addresses the more significant influence of dry and wet backgrounds in semi-humid regions. It simplifies the TCN to a single layer to weaken short-term heat wave interference, strengthens the Transformer to three layers and extends it to a 45-60 day window to capture slow drought accumulation, and assigns a 0.6 weight to the dry and wet features output by the Transformer. The final output is a 192-dimensional vector, highlighting the coupling between long-term dry and wet trends and weak heat wave signals. The encoder for humid regions addresses the characteristic that the CWHW of humid regions is dominated by a long-term humid background. The Transformer uses a 4-layer self-attention mechanism to capture the long-term correlation between daily-SPEI, precipitation, and heat waves in a 60-90 day window. For example, in a humid background, it takes 7-10 days for heat to accumulate and form a heat wave. The global average pooling layer smooths and isolates short-term fluctuations. For example, if a single day's high temperature is alleviated by precipitation, the output is a 256-dimensional vector, which focuses on preserving the temporal information of long-term humidity and slow heat accumulation.

[0104] S202. Based on the adjacency weights of geographical proximity and climate similarity, the low-dimensional temporal representation vectors of each grid point are fused with neighborhood information through convolution operation to obtain a spatial aggregated representation.

[0105] In a schematic representation, adjacency weights are used to quantify the association strength between the current grid point and its neighboring grid points, avoiding the neglect of climate differences by simply using weights based on geographical distance. Specifically, based on the latitude and longitude coordinates of a unified grid, the Euclidean distance between the current grid point and other grid points is calculated. A distance threshold is set; grid points with a distance less than the threshold are considered geographically neighboring grid points, while those with a distance greater than the threshold have a weight of 0. The weight of geographical proximity is calculated using a distance decay function, meaning the closer the distance, the greater the weight. For example, a distance d=10km has a weight of 0.8, and d=50km has a weight of 0.1, consistent with the physical law that the influence of nearby neighbors is stronger than that of distant neighbors. Furthermore, for geographically neighboring grid points, their climate similarity with the current grid point is calculated, with a similarity range of [0,1]. The higher the similarity, such as both being arid regions with a dryness index difference <0.1, the larger the correction coefficient, such as 1.0; the lower the similarity, such as the current grid point being in an arid region and the neighboring grid points being in a humid region, the smaller the correction coefficient, such as 0.3. Optionally, the geographical proximity weight is multiplied by the climate similarity correction coefficient to obtain the final adjacency weight between the current grid point and each neighboring grid point, which ranges from [0,1]. The larger the weight, the stronger the spatial influence of the neighboring grid points on the current grid point.

[0106] Spatial aggregation integrates neighborhood features through information transfer and weighted fusion. This can be achieved using graph neural networks (GNNs) or convolutional operations. For example, an undirected graph can be constructed using all grid points in the target region as nodes and adjacency weights as edge weights. For instance, the edge weight between the current grid point i and its neighboring grid point j is... Grid point j has the same weight as i. Furthermore, for each node, the low-dimensional temporal representation vectors of all its neighboring nodes are collected and summed using adjacency weights, such as the neighborhood vector. The contribution is × The neighborhood feature aggregation value is obtained, and this aggregation value is concatenated with the low-dimensional temporal vector of the current node. The feature is then fused through a 1-layer MLP with ReLU activation function to eliminate dimensional redundancy. The vector output by the MLP is the spatial aggregation representation of the current grid point, which has the same dimension as the low-dimensional temporal vector, but contains the spatial association information of the neighborhood grid points.

[0107] Optionally, after spatial aggregation, the effect needs to be verified by neighborhood consistency test. That is, for the spatial aggregation representation of a certain grid point, the cosine similarity between its representation and the aggregation representation of neighboring grid points is calculated. If the similarity is significantly higher than that of non-neighboring grid points, it indicates that the spatial association information has been effectively integrated. If there is no difference in similarity, the distance threshold or climate similarity calculation method needs to be adjusted to ensure the aggregation effect.

[0108] S203. Based on the cross-regional attention mechanism, the remote teleconnections and cross-regional patterns of the spatial aggregation representation are captured to obtain the local grid representation.

[0109] Indicatively, cross-regional attention is an improvement on the self-attention mechanism of Transformer. It uses the spatial aggregated representations of different regions as Key and Value, and the spatial aggregated representation of the current grid point as Query. Attention weights quantify the influence of distant regions on the current grid point. Specifically, the spatial aggregated representations of all grid points are divided into four groups based on dryness / wetness: arid region group, semi-arid region group, semi-humid region group, and humid region group. Features in each group are transformed into dimensions through a linear mapping layer. Key and Dimension The value, at the same time, transforms the spatial aggregation representation of the current grid point into a dimension through another linear mapping layer. The query is then further processed. The similarity between the query and all partition keys is calculated, such as dot product similarity, to obtain the original similarity between the current grid point and each remote partition grid point. The original similarity is normalized using the Softmax function to obtain attention weights, ranging from [0,1]. The sum of the weights is 1; a larger weight indicates a stronger influence of the remote partition grid point's features on the current grid point. For example, the similarity between the query for an arid region and another arid region key is high, with a weight close to 0.3, while the similarity with a humid region key is low, with a weight close to 0.05. The values ​​of each remote partition grid point are weighted and summed according to the attention weights to obtain a cross-region feature aggregation value. This aggregation value is then residually connected to the spatial aggregation representation of the current grid point and stabilized using a single LayerNorm (layer normalization) layer, ultimately outputting a local grid point representation.

[0110] Optionally, for partition pairs without significant long-range associations, such as most grid points between arid and humid regions, their attention weights can be set to 0 using a mask to reduce unnecessary computation, such as retaining only cross-regional associations between arid and semi-arid regions, and between semi-humid and humid regions. Based on the strength of cross-regional associations statistically analyzed from historical data, soft constraints can be applied to the attention weights, such as setting the lower limit of the weights for partition pairs with strong historical associations to 0.1, ensuring that the model prioritizes learning verified long-range association patterns.

[0111] S204. Based on the prediction time delay, the probability of occurrence, intensity regression value and duration distribution parameters of the composite heat wave event at each grid point are predicted in parallel according to the local grid point representation, and the uncertainty is estimated by using the Bayesian approximation method to obtain the prediction results.

[0112] To illustrate, the probability prediction of compound heat wave events targets two categories: compound drought heat wave (CDHW) and compound wet heat wave (CWHW). It employs two independent MLPs and sigmoid activation structures. Each MLP contains two fully connected layers with a hidden layer dimension of 64. The input is a local grid representation, and the output is the probability of occurrence of the corresponding category. For example, the CDHW probability submodule outputs 0.7, indicating that there is a 70% probability that CDHW will occur at that grid point within the future prediction time lag; the CWHW probability submodule outputs 0.2, indicating that there is a 20% probability that CWHW will occur.

[0113] The intensity prediction with a prediction time lag uses cumulative heat, i.e., the sum of daily temperature anomalies exceeding a threshold within the prediction time lag, to quantify the intensity. It employs an MLP and linear activation structure: the MLP contains two fully connected layers with a hidden layer dimension of 64, and uses the ReLU activation function. The input is a local grid representation, and the output is the expected intensity value. Optionally, to avoid predicted values ​​exceeding physically reasonable ranges, a truncation operation is added after the output layer to ensure the physical reasonableness of the intensity prediction.

[0114] The prediction of the duration of the time lag in the forecast is based on positive integers and follows a Poisson distribution. It employs an MLP (Multi-Level Processing) and a Poisson distribution parameterization structure, where the MLP outputs the rate parameter of the Poisson distribution. and This reflects the expected duration in days. The probability of a duration of 3 or 5 days can be calculated using the probability mass function of a Poisson distribution. The day with the highest probability is taken as the expected duration, or it can be output directly. As expected.

[0115] Uncertainty stems from model parameter uncertainty and data noise. MC-Dropout is used for Bayesian approximation estimation. Specifically, a Dropout layer is retained in the MLP with a multi-task output head, with a dropout probability of 0.2. This layer is enabled during both training and inference phases. By randomly dropping some neurons multiple times, such as 50 times, 50 independent prediction results are obtained. Statistics are calculated for these 50 prediction results. The mean is used as the deterministic prediction value, and the standard deviation is used as the uncertainty. Alternatively, a 95% confidence interval is calculated. The larger the standard deviation / the wider the confidence interval, the higher the uncertainty.

[0116] In one embodiment, such as Figure 3 As shown, the composite heat wave risk prediction model is obtained through the following method:

[0117] S301. Obtain the meteorological dataset and extract historical data from the meteorological dataset through the sample window to obtain data samples and prediction candidate windows; the meteorological dataset includes daily maximum temperature, daily average temperature, daily minimum temperature, daily precipitation and annual average precipitation distributed according to a preset uniform grid.

[0118] Indicatively, meteorological datasets must meet three major standards: spatial uniformity, temporal continuity, and parameter completeness. Core meteorological parameters include daily maximum temperature, daily average temperature, daily minimum temperature, daily precipitation, and annual average precipitation data distributed according to a pre-defined uniform grid, such as 0.5°×0.5°. By removing grid data with mismatched spatial resolution or poor temporal continuity, sample bias caused by data quality issues is avoided. For multi-source data, such as ERA5 and ground station data, a master data and supplementary verification mode is adopted, that is, ERA5 is the master, and ground station data is used for key grid verification to ensure the reliability of the dataset.

[0119] Furthermore, the sample window needs to adapt to the evolution pattern of historical accumulation and future outbreak of complex heat waves. A sliding window is used to construct sample pairs of historical inputs and future labels. Specifically, the lengths of the historical input window (sample window) and the future prediction window (prediction candidate window) are set. The length of the historical input window needs to match the characteristic response period of different regions, such as 14-30 days in arid areas and 60-90 days in humid areas, to ensure the capture of early incubation signals of complex heat waves. The length of the prediction candidate window needs to be consistent with the preset prediction lag, such as 1 day, 3 days, 7 days, or 14 days, to ensure that the labels are aligned with subsequent prediction targets. For example, selecting a 30-day historical input window and a 7-day prediction candidate window means using meteorological data from the past 30 days at a certain grid point to predict the risk of a complex heat wave in the next 7 days.

[0120] Optionally, the historical input data for each sample pair can be converted into a tensor format of grid points, time, and parameters, and the predicted candidate window can be marked with the time interval for generating labels to ensure that the sample format is consistent with the requirements of subsequent model input.

[0121] S302. Divide the data samples into training and validation sets according to daily and monthly scales, and generate corresponding data labels for the training and validation sets based on the data change characteristics within the prediction candidate window.

[0122] For illustration, daily-scale datasets such as ERA5 are used as training sets, and monthly-scale datasets such as ground station data are used as validation sets. Data labels are quantitative annotations of whether a compound heat wave occurs within the prediction candidate window, what type of compound heat wave occurs, and its intensity and duration. For example, within the prediction candidate window, a heat wave is determined based on a threshold temperature, and the dry / wet state is determined based on the daily SPEI. If both heat wave and drought conditions exist within the candidate window, it is labeled as a compound drought heat wave (CDHW); if both heat wave and wet conditions exist, it is labeled as a compound wet heat wave (CWHW). For samples labeled CDHW or CWHW, further quantitative indicators within the candidate window are extracted, including the date of occurrence, duration, cumulative heat, and cumulative SPEI, as labels for model intensity regression and duration prediction.

[0123] S303. A target function is constructed based on a weighted combination of classification loss, regression loss, partition weighted loss, physical consistency penalty term and probability calibration loss. Based on the target function, the composite heat wave risk prediction model is trained end-to-end according to the training set to obtain the model parameters. The partition-aware temporal encoders include arid zone temporal encoders based on short-term TCN and GRU, semi-arid zone temporal encoders based on short-term TCN, medium-term transformer and fully connected fusion layer, semi-humid zone temporal encoders based on short-term TCN, long-term transformer and attention fusion, and humid zone encoders based on long-term transformer and global flat pooling layer.

[0124] As an illustration, the objective function uses a weighted summation form, integrating five types of loss terms. Each type of loss term corresponds to an optimization objective of the model, and the weights are dynamically adjusted according to partitioning requirements and task importance. Among them, the classification loss ( To optimize the prediction accuracy of the probability of compound heat waves, a weighted binary cross-entropy loss is used. This addresses the sample imbalance problem of CDHW (Conditional High-Degree Heat Wave) with few samples and many non-compound event samples by assigning higher weights to minority class samples, thus preventing the model from biased towards the majority class. Regression loss ( To optimize the prediction accuracy of intensity and duration days, Huber loss is used. This loss is more robust to outliers, using MSE for small errors and MAE for large errors, avoiding model parameter oscillations caused by extreme samples. Partition-weighted loss ( This is used to adapt to the differences in sample characteristics across different zones, assigning differentiated weights to the losses in different arid and humid zones. For example, the CDHW samples in arid zones have high information value and are assigned a loss weight of 1.8; the CWHW samples in humid zones are the focus and are assigned a weight of 1.5; and the semi-arid / semi-humid zones are assigned a weight of 1.2. The weights are determined based on the impact of the combined heat wave in each zone and the scarcity of samples, ensuring that the model prioritizes learning the patterns of high-impact, low-sample zones. A physical consistency penalty term (…) This is used to constrain the model output to conform to meteorological principles and avoid physically contradictory results. For example, if the predicted CDHW intensity is positively correlated with the input daily-SPEIS, violating the negative correlation between stronger drought and higher intensity, a penalty is applied. Probabilistic calibration loss ( This is used to optimize the matching degree between the predicted probability and the actual occurrence frequency. The Brier score is used, which quantifies the mean square error between the predicted probability and the true label. The lower the score, the better the probability calibration.

[0125] Furthermore, differentiated training is performed on the partition-aware temporal encoder. Different initial parameters are used for each partition sub-encoder; for example, the TCN convolutional kernels in arid regions are randomly initialized, while the Transformer attention weights in humid regions are independently initialized. This avoids bias in initial parameters towards any particular partition and ensures that each sub-module can autonomously learn the partition-specific temporal patterns. Furthermore, when calculating the partition-weighted loss, only the sample loss for the corresponding partition is fed back to that partition sub-encoder. For example, the loss for arid region samples is only used to update the parameters of the TCN and GRU in arid regions, and the loss for humid region samples is only used to update the parameters of the Transformer in humid regions, ensuring that the optimization direction of the sub-module is consistent with the partition feature requirements. The parameters of the spatial aggregation layer and the cross-regional attention layer are shared by all partitions and are updated jointly during training through the losses of samples from all partitions, ensuring that the information interaction parameters between different partitions can learn the global cross-regional association patterns. Optionally, the model uses end-to-end training with the AdamW optimizer and an initial learning rate of 1e-3. A cosine annealing learning rate strategy is adopted, in which the learning rate decreases slowly according to a cosine curve in the early stage of training and stabilizes at a low learning rate in the later stage to avoid parameter oscillation in the later stage of training and ensure that the model converges to the optimal solution. The comprehensive performance index of the validation set is used as the monitoring object. If the validation set performance does not improve for 10 consecutive training epochs, training is stopped to avoid overfitting.

[0126] S304. Validate the model parameters based on the validation set, and determine the final composite heat wave risk prediction model based on the classification index, regression index and calibration index obtained from the validation.

[0127] The classification metrics are used to assess the accuracy of predicting the probability of a compound heat wave. The Area Under the ROC Curve (AUC) quantifies the model's ability to distinguish between occurrences and non-occurrences; an AUC closer to 1 indicates better classification performance. The Point of Detection (POD) calculates the proportion of samples that actually occurred and were predicted by the model to have occurred, reflecting the model's recall ability for compound events. The False Alarm Rate (FAR) calculates the proportion of samples that did not actually occur but were predicted by the model to have occurred, reflecting the model's false alarm rate. Regression metrics are used to assess the accuracy of predicting intensity (cumulative heat) and duration. The Root Mean Square Error (RMSE) quantifies the average deviation between predicted and actual values; a smaller RMSE indicates higher regression accuracy. The Mean Absolute Error (MAE) is a more robust deviation metric for outliers, supplementing RMSE in assessing the prediction accuracy for extreme events. The Continuous Ranking Probability Score (CRPS) quantifies the matching degree between the predicted and actual distributions of duration; a smaller CRPS indicates more accurate distribution predictions. The calibration metrics are used to evaluate the matching degree between the predicted probability and the actual occurrence frequency. Among them, the Brier score quantifies the mean square error between the predicted probability and the true label, with a score range of [0, 0.25]. The closer to 0, the better the calibration. The reliability curve groups the predicted probabilities and calculates the actual occurrence frequency of each group. If the frequency is close to the probability, it indicates that the model is well calibrated.

[0128] The validation set was divided into four groups according to dry and wet partitions. The classification, regression, and calibration indices were calculated for each group. The weighted average of the partitioned and grouped indices was used to obtain the global comprehensive index. The global comprehensive index corresponding to the model parameters in different training rounds was compared, and the parameter combination with the highest classification AUC, lowest regression RMSE, and lowest calibration Brier score was selected as the final model.

[0129] In one embodiment, generating corresponding data labels for the training set and validation set based on data change characteristics within the prediction candidate window includes:

[0130] S31. Based on the preset dry and wet zone division criteria, obtain the dry and wet zone labels according to the dryness index of the grid points.

[0131] Indicatively, the arid region is defined as having an aridity index ≤ 0.2, indicating an extreme water deficit year-round, with PET (particulate matter) far exceeding precipitation, prolonged drought duration, and the highest probability of heat waves coupled with drought. This threshold accurately defines the extreme drought background. The semi-arid region is defined as having an aridity index < 0.5, indicating large fluctuations in water supply and demand, frequent alternations between drought and wetness, and the possibility of both CDHW (cooling-drying wind) and CWHW (cooling-drying wind) with high intensity. This threshold range matches the climatic characteristics of seasonal drought in the semi-arid region. The semi-humid region is defined as having an aridity index < 0.65, indicating moderate water conditions, with precipitation concentrated in the rainy season and drought mostly seasonal. This threshold takes into account the monsoon region characteristics of wet summers and dry winters. The humid region is defined as having an aridity index > 0.65, indicating a significant water surplus, relatively even distribution of precipitation throughout the year, prolonged wetness, and a high probability of CWHW. This threshold matches the core characteristic of abundant annual precipitation in the humid region.

[0132] S32. If the daily maximum temperature is greater than the threshold temperature of the corresponding day for at least 3 consecutive days in the prediction candidate window, the prediction candidate window is identified as a heat wave event label.

[0133] A heatwave is defined as a condition for determining whether a heatwave has occurred, requiring at least three consecutive days with daily maximum temperatures exceeding a threshold temperature. Short-term high temperatures lasting less than three consecutive days are not considered heatwaves because they do not create a significant cumulative effect, thus avoiding interference from short-term fluctuations in composite event labels. If two heatwave events occur only 1-2 days apart and the threshold temperature is not reached, they are considered as a single heatwave, and their duration is calculated by merging the two events. This avoids splitting a single continuous heatwave into multiple short-term events and ensures that the labels reflect the actual duration of the heatwave. For example, the threshold temperature for each day within the predicted candidate window is extracted, and the actual daily maximum temperature within the candidate window is compared with the corresponding threshold temperature. Dates with a daily maximum temperature ≥ the threshold are marked as high-temperature days. The number of consecutive high-temperature days within the candidate window is counted. If the number is ≥ 3 days, a heatwave is determined to have occurred within the candidate window, and a heatwave event label = 1 is generated. If the number of consecutive high-temperature days is < 3 days, a heatwave is determined not to have occurred, and a heatwave event label = 0 is generated. Furthermore, for windows determined to be heatwaves, the start date, end date, and duration of the heatwave are recorded simultaneously.

[0134] S33. In the prediction candidate window, if the daily standardized precipitation evapotranspiration index of the grid point is less than the preset drought threshold, the prediction candidate window is determined to be a drought event label; if the daily standardized precipitation evapotranspiration index of the grid point is greater than the preset wet threshold, the prediction candidate window is determined to be a wet event label.

[0135] For illustration, a drought event is defined as a daily-SPEI < -0.5, corresponding to a mild drought level. At this point, a regional water supply-demand imbalance is evident, and when coupled with heat waves, it is prone to forming a CDHW (Cold-Dry, High-Hyper) pattern. A wet event is defined as a daily-SPEI > 0.5, corresponding to a mild wet level. At this point, the region has a water surplus and high air humidity, and when coupled with heat waves, it is prone to forming a CWHW (Cold-Wide, High-Hyper) pattern. For example, the daily grid point daily-SPEI values ​​within the prediction candidate window are extracted. The daily daily-SPEI is compared with the judgment threshold. Dates with daily-SPEI < -0.5 are marked as drought days, and dates with daily-SPEI > 0.5 are marked as wet days. Optionally, if there are ≥ 3 drought days within the candidate window, a drought event label = 1 is generated; if there are ≥ 3 wet days, a wet event label = 1 is generated; if both drought and wet days are < 3 days, it is judged as a normal state label = 0, and no drought / wet event label is generated.

[0136] For windows classified as drought / wet, the average daily-SPEI within the candidate window is calculated to quantify the severity of drought / wetness.

[0137] S34. In the prediction candidate window, if both heat wave event labels and drought event labels exist simultaneously, it is determined to be a combined drought heat wave; if both heat wave event labels and wet event labels exist simultaneously, it is determined to be a combined wet heat wave.

[0138] Indicatively, if the overlap between a heat wave event and a drought / wet event is ≥1 day, it indicates that the two have formed a synergistic effect, which is consistent with the synergistic occurrence of composite events. Within the same candidate window, CDHW and CWHW cannot coexist. That is, if the heat wave event label = 1 and the drought event label = 1, it is initially determined to be a CDHW candidate; if the heat wave event label = 1 and the wet event label = 1, it is initially determined to be a CWHW candidate.

[0139] S35. Record the occurrence date, duration, cumulative heat, and cumulative SPEI index of the combined drought heat wave or combined wet heat wave to obtain the event sequence label set.

[0140] Indicatively, the occurrence date is the start date of the composite event, the duration is the complete duration of the composite event, the cumulative heat is the sum of the differences between the daily maximum temperature and the threshold temperature within the duration, and the cumulative SPEI is the sum of the daily SPEI within the duration. For each sample labeled CDHW or CWHW, the above parameters are extracted from the meteorological dataset and the aforementioned labels to form a single event label vector.

[0141] In one embodiment, based on a physical constraint baseline, the prediction results are sequentially subjected to probability calibration, physical consistency filtering, and risk classification to obtain a risk probability heatmap, including:

[0142] S41. Based on statistical calibration, the probability of occurrence of composite heat wave events is probabilistically calibrated, and the probability of occurrence of composite heat wave events after probability calibration and the expected intensity of prediction time lag are filtered for physical consistency using daily potential evapotranspiration and daily standardized precipitation evapotranspiration index, so as to obtain prediction results with reasonable probability and intensity.

[0143] Illustratively, temperature scaling or Platt scaling is used. Temperature scaling introduces a single learnable parameter T (temperature), divides the original logarithmic probability logit(p) by T, and then activates it via a sigmoid function. When T > 1, the probability distribution is flatter; when T < 1, the probability distribution is steeper. Platt scaling fits a logistic regression model to the original output. , where a and b are fitting parameters. Specifically, by extracting the original predicted probabilities and corresponding true labels of the model on the validation set, it is ensured that the validation set and the training set do not overlap. The calibration parameters are optimized using the validation set data, with the optimization objective being to minimize the Brier score. The optimized parameters are then applied to the probability of the current prediction result to calculate the calibrated probability of the compound heat wave event.

[0144] Furthermore, if the CDHW probability is calibrated If the daily-SPEI value is greater than 0.6, but the input daily-SPEI value is greater than 0.5, then this is considered a contradiction, and the daily-SPEI value will be adjusted accordingly. Lower it to below 0.3, or adjust it dynamically based on the SPEI value, such as lowering it to 0.2 when SPEI=0.8, to avoid the absurd result of high CDHW risk in a wet background. If the CWHW probability is calibrated... If the value is greater than 0.6, but the input daily-SPEI is less than -0.5, then... The value was lowered to below 0.3, consistent with the principle that a humid background is a prerequisite for the formation of CWHW. If the predicted expected intensity I > 30℃·day, but the daily average PET within the prediction time lag is < 5mm, then I is corrected to I × (PET / 5) to ensure that the intensity matches the moisture consumption capacity. If the predicted duration D > 20 days, D is truncated to 20 days, referencing the maximum duration of composite heat waves in historical observations, to avoid extreme and unreasonable values.

[0145] S42. According to the preset probability risk rules, classify the prediction results into high / medium / low risk levels, and generate a gridded heat map by combining the risk level, the probability of occurrence of the composite heat wave event after probability calibration, the expected intensity of the prediction time delay, the number of days of the prediction time delay, and the uncertainty estimate, so as to obtain a risk probability heat map.

[0146] Indicatively, low risk means the probability of a calibrated composite heatwave event is p < 0.3, or 0.3 ≤ p < 0.5 but the expected intensity I < 15℃·day and the duration D < 3 days. Medium risk means 0.3 ≤ p < 0.7 and (I > 15℃·day or D ≥ 3 days); or p ≥ 0.7 but I > 30℃·day and D < 2 days. High risk means p ≥ 0.7 and I > 20℃·day and D ≥ 4 days; or p ≥ 0.6 and I > 30℃·day and D ≥ 5 days.

[0147] Furthermore, optionally, the heatmap for the probability of a compound heat wave uses color mapping: blue for low probability, yellow for medium probability, and red for high probability. The grid probability data is converted to GeoTIFF raster format and matched with the administrative boundary map and topographic base map of the target area to ensure that the risk information of each raster accurately corresponds to the actual geographical location. The color mapping for the expected intensity heatmap uses light orange for low intensity (I < 15℃·day), orange for medium intensity (15℃·day ≤ I < 30℃·day), and dark orange for high intensity (I ≥ 30℃·day). The heatmap for duration uses light green for short duration (D < 3 days), green for medium duration (3 days ≤ D < 5 days), and dark green for long duration (I ≥ 5 days).

[0148] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0149] Based on the same inventive concept, this application also provides a composite heat wave dynamic risk prediction system for implementing the aforementioned composite heat wave dynamic risk prediction method. The solution provided by this system is similar to the implementation scheme described in the above method; therefore, the specific limitations of one or more embodiments of the composite heat wave dynamic risk prediction system provided below can be found in the limitations of the composite heat wave dynamic risk prediction method described above, and will not be repeated here.

[0150] In one exemplary embodiment, such as Figure 4 As shown, a composite heat wave dynamic risk prediction system is provided, comprising:

[0151] The observation module 401 is used to acquire time-series observation data and prediction time lag of the target area, and to preprocess the time-series observation data to obtain standardized gridded time-series meteorological data; the time-series observation data includes daily maximum temperature, daily average temperature, daily minimum temperature, daily precipitation and timestamp;

[0152] The index module 402 is used to extract the change characteristics of each grid point based on historical window cache meteorological data and standardized grid time-series meteorological data to obtain daily potential evapotranspiration, daily standardized precipitation evapotranspiration index, threshold temperature and multivariate change characteristic matrix; the historical window cache meteorological data includes historical grid time-series meteorological data and climate baseline data.

[0153] Feature module 403 is used to determine the dry and wet zone labels of the target area based on the aridity index, and to concatenate the dry and wet zone labels, daily potential evapotranspiration, daily standardized precipitation evapotranspiration index, threshold temperature, multivariate variation feature matrix and grid static features according to a preset sliding window and prediction time lag to obtain a feature tensor; the grid static features are topographic elevation and land cover; the dry and wet zone labels include arid zone, semi-arid zone, semi-humid zone and humid zone;

[0154] The prediction module 404 is used to input the feature tensor into the pre-trained composite heat wave risk prediction model to obtain the prediction result. The composite heat wave risk prediction model includes a partition-aware temporal encoder, a spatial aggregation layer, a cross-regional attention layer, and a multi-task output head connected in sequence. The prediction result includes the probability of composite heat wave events occurring at each grid point on the prediction time delay, the expected intensity of the prediction time delay, the number of days the prediction time delay lasts, and an uncertainty estimate. The probability of composite heat wave events occurring is either the probability of composite drought heat wave or the probability of composite wet heat wave.

[0155] The post-processing module 405 is used to perform probability calibration, physical consistency filtering and risk classification on the prediction results based on the physical constraint baseline, so as to obtain a risk probability heatmap.

[0156] In one embodiment, a data processing module is also included, for:

[0157] Outlier detection, time alignment, and missing value imputation are performed on time-series observation data to obtain gridded time-series raw data; missing value imputation is based on long-term climate statistical characteristics.

[0158] Based on a pre-defined unified grid, a bilinear interpolation algorithm is used to spatially interpolate the gridded time series raw data to obtain the gridded time series raw data; the unified grid is a unified prediction network grid used for the prediction network;

[0159] Standardized gridded time-series meteorological data are obtained by standardizing the original gridded time-series data.

[0160] In one embodiment, the index module 402 is further configured to:

[0161] Based on the improved Hargreaves method, the daily potential evapotranspiration of each grid point is calculated daily according to the daily maximum temperature, daily average temperature, daily minimum temperature and daily precipitation.

[0162] The daily water balance is calculated based on the daily potential evapotranspiration and daily precipitation, and the cumulative water surplus or deficit is obtained by sliding accumulation of the daily water balance.

[0163] The daily standardized precipitation evapotranspiration index of each grid point is obtained by fitting the probability distribution of the cumulative water balance and standardizing it.

[0164] Based on the historical daily maximum temperature data of the climate baseline period, the same day location for obtaining time series observation data is used as a preset step size sliding window. The 90th percentile value is statistically calculated to obtain the threshold temperature of each grid point.

[0165] The variation characteristics of historical gridded time-series meteorological data and standardized gridded time-series meteorological data are extracted to obtain a multivariate variation characteristic matrix.

[0166] In one embodiment, the prediction module 404 is further configured to:

[0167] According to the dry and wet partition labels, the feature tensor is input into the corresponding partition-aware temporal encoder to obtain the low-dimensional temporal representation vector of each grid point;

[0168] Based on adjacency weights derived from geographical proximity and climate similarity, a spatial aggregation representation is obtained by fusing neighborhood information into the low-dimensional temporal representation vectors of each grid point through convolution operations.

[0169] Based on the cross-regional attention mechanism, we capture the long-range teleconnections and cross-regional patterns of the spatial aggregation representation to obtain the local grid representation;

[0170] Based on the prediction time lag, the probability of occurrence, intensity regression value and duration distribution parameters of the composite heat wave event at each grid point are predicted in parallel according to the local grid point representation, and the uncertainty is estimated by using the Bayesian approximation method to obtain the prediction results.

[0171] In one embodiment, a model building module is also included, for:

[0172] Acquire meteorological datasets and extract historical data from them through sample windows to obtain data samples and prediction candidate windows; the meteorological datasets include daily maximum temperature, daily average temperature, daily minimum temperature, daily precipitation, and annual average precipitation distributed according to a preset uniform grid.

[0173] The data samples are divided into training and validation sets according to daily and monthly scales, and corresponding data labels are generated for the training and validation sets based on the data change characteristics within the prediction candidate window.

[0174] The objective function is constructed by weighting the classification loss, regression loss, partition weighted loss, physical consistency penalty term and probability calibration loss. Based on the objective function, the composite heat wave risk prediction model is trained end-to-end according to the training set to obtain the model parameters. The partition-aware temporal encoders include arid zone temporal encoders based on short-term TCN and GRU, semi-arid zone temporal encoders based on short-term TCN, medium-term transformer and fully connected fusion layer, semi-humid zone temporal encoders based on short-term TCN, long-term transformer and attention fusion, and humid zone encoders based on long-term transformer and global flat pooling layer.

[0175] The model parameters are validated using the validation set, and the final composite heat wave risk prediction model is determined based on the classification indicators, regression indicators, and calibration indicators obtained from the validation.

[0176] In one embodiment, a label generation module is also included, for:

[0177] Based on the preset dry and wet zone division criteria, dry and wet zone labels are obtained according to the dryness index of the grid points.

[0178] If the daily maximum temperature is greater than the threshold temperature for the corresponding day for at least 3 consecutive days in the prediction candidate window, the prediction candidate window is identified as a heat wave event label.

[0179] In the prediction candidate window, if the daily standardized precipitation evapotranspiration index of the grid point is less than the preset drought threshold, the prediction candidate window is determined to be a drought event label; if the daily standardized precipitation evapotranspiration index of the grid point is greater than the preset wet threshold, the prediction candidate window is determined to be a wet event label.

[0180] In the prediction candidate window, if both heat wave event labels and drought event labels are present, it is determined to be a combined drought heat wave; if both heat wave event labels and wet event labels are present, it is determined to be a combined wet heat wave.

[0181] Record the occurrence date, duration, cumulative heat, and cumulative SPEI index of the combined drought heat wave or combined wet heat wave to obtain an event sequence label set.

[0182] In one embodiment, a correction module is also included for:

[0183] Based on statistical calibration, the probability of occurrence of compound heat wave events is probabilistically calibrated, and the probability of occurrence of compound heat wave events after probability calibration and the expected intensity of prediction time lag are filtered for physical consistency using daily potential evapotranspiration and daily standardized precipitation evapotranspiration index, so as to obtain prediction results with reasonable probability and intensity.

[0184] According to the preset probability risk rules, the prediction results are classified into high / medium / low risk levels, and the risk level, the probability of occurrence of the composite heat wave event after probability calibration, the expected intensity of the prediction time delay, the duration of the prediction time delay in days, and the uncertainty estimate are used to generate a rasterized heat map, thus obtaining a risk probability heat map.

[0185] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps in the above method embodiments.

[0186] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0187] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0188] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method for predicting the dynamic risk of a composite heat wave, characterized in that, The method includes: The time-series observation data and prediction time lag of the target area are acquired, and the time-series observation data are preprocessed to obtain standardized gridded time-series meteorological data; the time-series observation data includes daily maximum temperature, daily average temperature, daily minimum temperature, daily precipitation and timestamp; Based on historical window cached meteorological data, the variation characteristics of each grid point are extracted according to the standardized grid time-series meteorological data to obtain daily potential evapotranspiration, daily standardized precipitation evapotranspiration index, threshold temperature and multivariate variation characteristic matrix; the historical window cached meteorological data includes historical grid time-series meteorological data and climate baseline period data. The dry and wet zone labels of the target area are determined based on the aridity index. The dry and wet zone labels, daily potential evapotranspiration, daily standardized precipitation evapotranspiration index, threshold temperature, multivariate variation feature matrix, and grid static features are then concatenated using a preset sliding window and the prediction time lag to obtain a feature tensor. The grid static features include topographic elevation and land cover. The dry and wet zone labels include arid, semi-arid, semi-humid, and humid regions. The feature tensor is input into a pre-trained composite heat wave risk prediction model to obtain the prediction result. The composite heat wave risk prediction model includes a sequentially connected partition-aware temporal encoder, a spatial aggregation layer, a cross-regional attention layer, and a multi-task output head. The prediction result includes the probability of composite heat wave events occurring at each grid point on the prediction time delay, the expected intensity of the prediction time delay, the number of days the prediction time delay lasts, and an uncertainty estimate. The probability of composite heat wave events occurring is either the probability of a composite drought heat wave or the probability of a composite wet heat wave. Based on the physical constraint baseline, the prediction results are sequentially subjected to probability calibration, physical consistency filtering, and risk classification to obtain a risk probability heatmap.

2. The method according to claim 1, characterized in that, The preprocessing of the time-series observation data to obtain standardized gridded time-series meteorological data includes: Outlier detection, time alignment, and missing value imputation are performed on the time-series observation data to obtain gridded time-series raw data; the missing value imputation is based on long-term climate statistical characteristics. Based on a preset unified grid, a bilinear interpolation algorithm is used to spatially interpolate the original time series data of the grid points to obtain the original time series data of the grid points; the unified grid is a unified prediction network grid used for the prediction network. The original gridded time-series data is standardized to obtain standardized gridded time-series meteorological data.

3. The method according to claim 1, characterized in that, The historical window-based cached meteorological data, based on the standardized gridded time-series meteorological data, extracts the variation characteristics of each grid point to obtain daily potential evapotranspiration, daily standardized precipitation evapotranspiration index, threshold temperature, and a multivariate variation feature matrix, including: Based on the improved Hargreaves method, the daily potential evapotranspiration of each grid point is calculated daily according to the daily maximum temperature, daily average temperature, daily minimum temperature and daily precipitation. The daily water balance is calculated based on the daily potential evapotranspiration and the daily precipitation, and the cumulative water surplus or deficit is obtained by sliding accumulation of the daily water balance. The daily standardized precipitation evapotranspiration index of each grid point is obtained by fitting the probability distribution of the accumulated moisture surplus and deficit and standardizing it. Based on the historical daily maximum temperature data of the climate baseline period data, the same daily location for obtaining time-series observation data is used as a preset step size sliding window. The 90th percentile value is statistically calculated to obtain the threshold temperature of each grid point. The variation characteristics of the historical gridded time-series meteorological data and the standardized gridded time-series meteorological data are extracted to obtain the multivariate variation feature matrix.

4. The method according to claim 3, characterized in that, The composite heatwave risk prediction model obtains the prediction results through the following method: According to the dry and wet partition labels, the feature tensor is input into the corresponding partition-aware temporal encoder to obtain the low-dimensional temporal representation vector of each grid point; Based on adjacency weights derived from geographical proximity and climate similarity, neighborhood information is fused into the low-dimensional temporal representation vectors of each grid point through convolution operations to obtain a spatial aggregated representation. Based on the cross-regional attention mechanism, the remote teleconnections and cross-regional patterns of the spatial aggregation representation are captured to obtain the local grid representation; Based on the prediction time lag, the probability of occurrence, intensity regression value and duration distribution parameters of the composite heat wave event at each grid point are predicted in parallel according to the local grid point representation, and the uncertainty is estimated using the Bayesian approximation method to obtain the prediction results.

5. The method according to claim 4, characterized in that, The composite heatwave risk prediction model was obtained through the following method: A meteorological dataset is acquired, and historical data is extracted from the meteorological dataset through a sample window to obtain data samples and prediction candidate windows; the meteorological dataset includes daily maximum temperature, daily average temperature, daily minimum temperature, daily precipitation, and annual average precipitation distributed according to a preset uniform grid. The data samples are divided into training and validation sets according to daily and monthly scales, and corresponding data labels are generated for the training and validation sets based on the data change characteristics within the prediction candidate window. An objective function is constructed based on a weighted combination of classification loss, regression loss, partition-weighted loss, physical consistency penalty term, and probability calibration loss. Based on the objective function, the composite heat wave risk prediction model is trained end-to-end according to the training set to obtain model parameters. The partition-aware temporal encoder includes an arid region temporal encoder based on short-term TCN and GRU, a semi-arid region temporal encoder based on short-term TCN, medium-term transformer, and fully connected fusion layer, a semi-humid region temporal encoder based on short-term TCN, long-term transformer, and attention fusion, and a humid region encoder based on long-term transformer and global flat pooling layer. The model parameters are validated based on the validation set, and the final composite heat wave risk prediction model is determined based on the classification index, regression index and calibration index obtained from the validation.

6. The method according to claim 5, characterized in that, The step of generating corresponding data labels for the training set and the validation set based on the data change features within the prediction candidate window includes: Based on the preset dry and wet zone division criteria, dry and wet zone labels are obtained according to the dryness index of the grid points; If the daily maximum temperature is greater than the corresponding daily threshold temperature for at least 3 consecutive days in the prediction candidate window, the prediction candidate window is identified as a heat wave event label. In the prediction candidate window, if the daily standardized precipitation evapotranspiration index of the grid point is less than the preset drought threshold, then the prediction candidate window is determined to be a drought event label; if the daily standardized precipitation evapotranspiration index of the grid point is greater than the preset wet threshold, then the prediction candidate window is determined to be a wet event label. In the prediction candidate window, if both the heat wave event label and the drought event label exist simultaneously, it is determined to be a combined drought heat wave; if both the heat wave event label and the wet event label exist simultaneously, it is determined to be a combined wet heat wave. Record the occurrence date, duration, cumulative heat, and cumulative SPEI index of the combined drought heat wave or the combined wet heat wave to obtain an event sequence label set.

7. The method according to claim 1, characterized in that, Based on the physical constraint baseline, the prediction results are sequentially subjected to probability calibration, physical consistency filtering, and risk classification to obtain a risk probability heatmap, including: The probability of the occurrence of the composite heat wave event is probabilistically calibrated based on statistical calibration, and the probability of the occurrence of the composite heat wave event after probability calibration and the expected intensity of the prediction time lag are filtered for physical consistency using the daily potential evapotranspiration and the daily standardized precipitation evapotranspiration index, so as to obtain the prediction result with reasonable probability and intensity. According to the preset probability risk rules, the prediction results are classified into high / medium / low risk levels, and the risk level, the probability of occurrence of the composite heat wave event after probability calibration, the expected intensity of the prediction time delay, the duration of the prediction time delay in days, and the uncertainty estimate are used to generate a rasterized heat map to obtain the risk probability heat map.

8. A composite heatwave dynamic risk prediction system, characterized in that, The system includes: The observation module is used to acquire time-series observation data and prediction time lag of the target area, and to preprocess the time-series observation data to obtain standardized gridded time-series meteorological data; the time-series observation data includes daily maximum temperature, daily average temperature, daily minimum temperature, daily precipitation and timestamp; The index module is used to extract the variation characteristics of each grid point based on the historical window cache meteorological data and the standardized grid time-series meteorological data to obtain the daily potential evapotranspiration, daily standardized precipitation evapotranspiration index, threshold temperature and multivariate variation characteristic matrix; the historical window cache meteorological data includes historical grid time-series meteorological data and climate baseline data. The feature module is used to determine the dry and wet zone labels of the target area based on the aridity index, and to concatenate the dry and wet zone labels with the daily potential evapotranspiration, daily standardized precipitation evapotranspiration index, threshold temperature, multivariate variation feature matrix and grid static features according to a preset sliding window and the prediction time lag to obtain a feature tensor; the grid static features are topographic elevation and land cover; the dry and wet zone labels include arid areas, semi-arid areas, semi-humid areas and humid areas; The prediction module is used to input the feature tensor into a pre-trained composite heat wave risk prediction model to obtain the prediction result. The composite heat wave risk prediction model includes a partition-aware temporal encoder, a spatial aggregation layer, a cross-regional attention layer, and a multi-task output head connected in sequence. The prediction result includes the probability of composite heat wave events occurring at each grid point on the prediction time delay, the expected intensity of the prediction time delay, the duration of the prediction time delay in days, and an uncertainty estimate. The probability of composite heat wave events occurring is either the probability of a composite drought heat wave or the probability of a composite wet heat wave. The post-processing module is used to perform probability calibration, physical consistency filtering, and risk classification on the prediction results based on the physical constraint baseline, so as to obtain a risk probability heatmap.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

Images