Precipitation prediction method based on climate partition and LGBM model

Abstract

The invention discloses a rainfall prediction method based on climate zoning and an LGBM model, and relates to the technical field of meteorological processing, and the method comprises the steps: obtaining rainfall associated multi-source data covering a target area, and constructing a structured original data set; dividing the structured original data set according to a climate zoning rule of a target area and carrying out precipitation scene exclusive data preprocessing to generate a standardized precipitation modeling data set of each climate area; based on the standardized rainfall modeling data set of each climate region, training an LGBM rainfall prediction model, and generating an optimal rainfall prediction model matched with each climate region; and obtaining rainfall related data of a target area in a to-be-predicted time period, dividing the rainfall related data according to climate regions, respectively inputting the corresponding optimal rainfall prediction models, generating rainfall prediction results of the climate regions, and fusing the rainfall prediction results according to spatial positions to form a complete rainfall prediction distribution diagram of the target area. And through model adaptation optimization of precipitation multi-source associated data fusion, a precipitation prediction scheme adapted to complex climate and topographic features is formed.

Description

Technical Field

[0001] This invention relates to the field of meteorological processing technology, specifically to a precipitation prediction method based on climate zoning and the LGBM model. Background Technology

[0002] Precipitation, as a core component of the Earth's water cycle, directly regulates ecosystem evolution, agricultural production layout, and water resource security through its spatiotemporal distribution characteristics. It is also a key driving factor for meteorological disasters such as floods and droughts. Due to climate and topographic variations in the location of precipitation areas, regional precipitation exhibits significant spatiotemporal heterogeneity, leading to frequent extreme precipitation events. Therefore, accurately identifying the core influencing factors of precipitation and quantifying the intensity and interaction mechanisms of these factors are crucial prerequisites for improving the accuracy of precipitation forecasting and formulating regionally differentiated disaster prevention and mitigation strategies. These are also current research hotspots and challenges in the fields of atmospheric science and meteorology.

[0003] Currently, research techniques for studying the impact of precipitation factors fall into two main categories: traditional statistical methods and conventional machine learning methods. Both types of techniques have significant limitations. Traditional statistical methods, based on the core assumption of linear correlation, have mature technical systems, low computational costs, and strong interpretability, but they also have inherent defects: the linear assumption limits their ability to capture nonlinear relationships and cannot identify threshold effects; they lack the ability to handle high-dimensional variables, making it difficult to quantify the synergistic regulation mechanisms and interaction effects of multiple factors; and they are insufficient in capturing small-scale local precipitation characteristics.

[0004] Conventional machine learning methods are data-driven and break through the linear assumption, showing advantages in scenarios such as extreme precipitation event identification and short-term precipitation prediction. However, they have inherent defects of black box models: they cannot accurately quantify the contribution ratio and positive and negative effects of each influencing factor; they lack local interpretability and cannot clearly define the threshold range of factors under different climate zones and different precipitation intensities; they are difficult to intuitively present the interaction effect form and contribution intensity between influencing factors; and the model's generalization ability is significantly affected by the data distribution.

[0005] Furthermore, existing research suffers from systemic framework deficiencies: the research scale is limited to single climate zones or a few watersheds, lacking systematic studies at the national scale; variable selection is one-sided, with insufficient comprehensive consideration of underlying surface factors and cloud microphysical factors; the spatial variability of influencing factors has not been clarified, ignoring regional heterogeneity; and there is a lack of targeted nonlinear framework quantitative coordination mechanisms, resulting in an understanding of precipitation formation mechanisms that remains at the level of phenomenological correlations. These deficiencies make it difficult for existing research results to effectively support regionally differentiated precipitation regulation strategies and meteorological disaster responses, necessitating the development of new technological solutions to overcome existing limitations. Summary of the Invention

[0006] The purpose of this invention is to provide a precipitation prediction method based on climate zoning and LGBM model. By constructing a multi-source precipitation correlation data fusion + precipitation scene-specific preprocessing + LGBM model precipitation adaptation optimization, a precipitation prediction scheme adapted to complex climate and terrain features is formed.

[0007] To achieve the above objectives, this application proposes the following solution: On the one hand, this application provides a precipitation prediction method based on climate zoning and the LGBM model, specifically including the following steps: S1. Obtain precipitation-related multi-source data covering the target area and construct a structured raw dataset; S2. Based on the climate zoning rules of the target area, the structured raw dataset is divided into climate zones and precipitation scenario-specific data preprocessing is performed to generate standardized precipitation modeling datasets for each climate zone. S3. Based on the standardized precipitation modeling datasets of each climate zone, LGBM precipitation prediction models are trained separately. Bayesian optimization is used to independently tune the core hyperparameters of the LGBM precipitation prediction models of each climate zone. The models are then validated by time series cross-validation to generate the optimal precipitation prediction model that is adapted to each climate zone. S4. Obtain precipitation-related data for the target area within the forecast period and divide it according to climate zones. Input the corresponding optimal precipitation prediction model to generate precipitation prediction results for each climate zone and merge them according to spatial location to form a complete precipitation prediction distribution map of the target area.

[0008] In some specific implementation plans, precipitation-related multi-source data include meteorological station observation data, satellite remote sensing data, atmospheric reanalysis data, and geographic auxiliary data; The precipitation-related multi-source data were filtered to obtain valid data. The valid data were then classified into several subsets according to data type. For each subset, a three-dimensional structured attribute containing timestamps, spatial latitude and longitude, and factor values ​​was constructed to obtain the structured raw dataset. The structured raw dataset includes a subset of core target data, a subset of thermal factor data, a subset of dynamic factor data, a subset of cloud microphysical factor data, and a subset of underlying surface and auxiliary factor data.

[0009] In some specific implementation plans, the specific process for preprocessing precipitation scenario-specific data is as follows: S21. Perform spatiotemporal alignment and resolution unification on the structured original dataset. Unify the temporal window and spatial resolution of all subsets in the structured original dataset to the precipitation observation standard, while preserving topographic details and clarifying the correspondence between the interpolation method of each subset and the original resolution. S22. After spatiotemporal alignment and resolution unification, the core target data subset is cleaned by threshold filtering and hierarchical sampling strategy for precipitation data cleaning. S23. Derive collaborative features of dynamic factors from the subset of dynamic factor data after spatiotemporal alignment and resolution unification; S24. Impute missing values ​​for each subset after processing in steps S21-S23; S25. Factor standardization and global spatiotemporal fusion are performed on each subset after processing in step S24 to obtain single-table structured data. S26. Divide the single-table structured data to obtain standardized precipitation modeling datasets for each climate zone.

[0010] In some specific implementation schemes, the specific process of step S21 is as follows: The time windows of all subset data are unified to a daily scale that matches the precipitation observation period, and the spatial resolution is unified to a preset resolution: For subsets with resolutions lower than the preset resolution, bilinear interpolation is used to increase the resolution of the subsets. For subsets with resolutions higher than or equal to the preset resolution, nearest neighbor interpolation is used to decrease the resolution of the subsets, while preserving terrain and land use details. Precipitation scenario-specific factors were added to the underlying surface and auxiliary factor data subsets after resolution unification.

[0011] In some specific implementation schemes, the resolution-unified underlying surface and auxiliary factor data subset includes DEM data at a preset resolution, and the topographic relief F is calculated based on the DEM data at the preset resolution. h Topographic relief is used as a specific factor for precipitation scenarios. Topographic relief F h The calculation process is as follows: F h It equals the maximum elevation within the rectangular window minus the minimum elevation within the rectangular window, where the rectangular window is the adjacent raster range of the DEM data at the preset resolution.

[0012] In some specific implementation schemes, the dynamic factor data subset includes wind speed (WS) and wind direction (WD) from the ERA5 dataset. The specific process for deriving the dynamic factor co-features is as follows: Based on wind speed WS and wind direction WD, calculate the wind speed-wind direction synergy factor F used to quantify water vapor transport intensity. ws_wd F ws_wd =cos[WS×(WD-180°)].

[0013] In some specific implementation schemes, the missing value imputation process is as follows: The method employs climate zone mean combined with inverse distance weighted interpolation (IDW). First, the mean factor of the original structured dataset is calculated as the initial value for missing values ​​based on the climate zone to which it belongs. Then, the IDW method is used to correct this by combining the effective factor values ​​of the three surrounding raster cells at preset resolutions. The final missing value V is equal to the sum of the products of the effective factor values ​​of the three surrounding raster cells and the inverse square of their corresponding distances, divided by the sum of the inverse squares of the distances of the three surrounding raster cells. This is expressed as: V= ; in, v i Indicates the first i The effective factor value of each surrounding raster. di Indicates the first i The distance from each surrounding grid cell to the location of the missing value.

[0014] In some specific implementation schemes, the process of obtaining structured data in a single table is as follows: Dimensional normalization: For all factors in each subset except the core target data subset after processing in step S24, Z-score normalization is applied to adapt to the normal distribution characteristics of precipitation factors and avoid extreme values ​​distorting the data distribution. Global spatiotemporal fusion: Based on daily-scale timestamps and preset resolution latitude and longitude raster, the processed subsets of thermal factor data, dynamic factor data, cloud microphysical factor data, underlying surface and auxiliary factor data are used as factor columns and fused with the target column of the core target data subset to form a single-table structured data.

[0015] In some specific implementation schemes, step S3, generating the optimal precipitation prediction model adapted to each climate zone, specifically includes: S31. Construct the LGBM base model, retain the original GOSS gradient one-sided sampling and EFB exclusive feature binding core mechanism of the LGBM base model, and initialize the initial training parameters of the LGBM base model. S32. Based on the precipitation characteristics of the standardized precipitation modeling dataset, adjust the retention ratio of high-gradient samples and the sampling ratio of low-gradient samples in the GOSS gradient one-sided sampling; and according to the precipitation physical mechanism classification, bind the high-dimensional factors in the standardized precipitation modeling dataset to obtain the LGBM precipitation prediction model. S33. Set the core hyperparameter search range for the LGBM precipitation prediction model for precipitation scenarios, and use Bayesian optimization method to iteratively optimize the model with the goal of minimizing the root mean square error and maximizing the coefficient of determination. S34. Use the time series ten-fold cross-validation method to verify the effectiveness of the optimal hyperparameter combination, ensuring that the training set time is earlier than the validation set time: S35. Based on the validated optimal hyperparameter combination, the LGBM precipitation prediction model is trained using the full training set to obtain the optimal precipitation prediction model.

[0016] In some specific implementation schemes, the following steps are also included: S5. Based on the preset climate zoning rules, perform zoning verification on the standardized precipitation modeling dataset and the optimal precipitation prediction model for each climate zone, and output the targeted explanation results of the zoning precipitation mechanism. S6. Embed the optimal precipitation prediction model into the SHAP interpretation framework, perform multi-dimensional SHAP analysis on the standardized precipitation modeling datasets of each climate zone, and output SHAP targeted interpretation results. S7. The results of the targeted interpretation of regional precipitation mechanisms and the results of the SHAP targeted interpretation are structurally integrated to construct a comparative system of global-regional precipitation driving mechanisms, and to clarify the common driving laws and regional unique regulation mechanisms of precipitation in different climate zones.

[0017] The advantages of this invention over the prior art are as follows: 1. This invention integrates multi-source data related to precipitation influencing factors, including ground observations, satellite remote sensing, atmospheric reanalysis, and geographic auxiliary data (such as DEM and NDVI), to construct a precipitation-related dataset covering multiple dimensions such as thermodynamics, dynamics, cloud microphysics, and underlying surface, providing richer physical information for the model. Simultaneously, addressing the high spatiotemporal heterogeneity and imbalanced sample characteristics of precipitation data, a dedicated preprocessing workflow was designed, including spatiotemporal alignment to a 1km daily scale, outlier cleaning, stratified sampling to balance the sample, factor derivation based on physical mechanisms (such as topographic relief and wind speed-direction synergy factors), and missing value imputation using climate zone means combined with inverse distance weights. These processes effectively improve data quality and completeness, avoid the interference of data bias on model accuracy, and lay a solid foundation for subsequent high-precision modeling.

[0018] 2. This invention performs deep adaptation and optimization of the LGBM model based on the characteristics of precipitation data: The proportion of gradient one-sided sampling (GOSS) is adjusted to ensure that high-gradient information from extreme precipitation samples is fully learned; Exclusive Feature Bundling (EFB) grouping is redesigned according to precipitation physical mechanisms to avoid information confusion between factors of different physical processes; Bayesian optimization and time-series 10-fold cross-validation are used to accurately search for optimal hyperparameters (such as the number of leaf nodes, tree depth, learning rate, etc.) in millions of samples, while ensuring that the training set time is always earlier than the validation set to prevent time information leakage. The resulting optimized model can accurately characterize the nonlinear and non-stationary characteristics of precipitation, significantly improving the predictive ability for extreme precipitation events.

[0019] 3. The optimized LGBM model is embedded into the SHAP interpretation framework to deeply analyze the precipitation mechanism from four dimensions: identifying core driving factors at the national scale (such as relative humidity and convective available potential energy) through feature importance ranking; quantifying the positive and negative contributions and magnitudes of each factor using waterfall plots; revealing the critical triggering thresholds of key factors using bee swarm plots (such as a significant increase in precipitation probability when relative humidity > 80%); and mining synergistic effects among multiple factors through dependency graphs (such as saturated humidity and high cloud top height jointly promoting precipitation). This interpretability not only verifies the physical rationality of the model's predictions but also provides a clear basis for model iterative optimization, thereby effectively improving prediction accuracy.

[0020] Based on a complete technical process of multi-source data fusion, model optimization, interpretable analysis, and regional validation, and utilizing SHAP feature importance analysis technology, relative humidity (RH), convective available potential energy (CAPE), K index (KI), and ice-water path (IWP) were accurately identified as the four core driving factors of precipitation at the national scale. Their average SHAP values ​​were significantly higher than other variables, highly consistent with climatological theories of precipitation formation. Based on visualization analysis methods using SHAP waterfall plots and bee swarm plots, the direction and threshold of influence of each meteorological variable on precipitation at the national scale were quantified. Surface temperature (LST), evaporation (E), and wind direction (WD) were identified as positive contributing factors, while ice-water path (IWP), K index (KI), and cloud top height (CTH) were identified as negative contributing factors. Through 7 The regional validation scheme of the macroclimate zone clearly reveals the spatial heterogeneity of precipitation driving mechanisms and identifies the dominant influencing factors in different climate zones (CAPE has a prominent influence in the northern subtropical humid zone, RH is dominant in the marginal tropical humid zone, and CTH plays a key role in the plateau temperate semi-arid zone). Utilizing SHAP four-dimensional visualization technology, it achieves a full-chain explanation from the importance of global variables to the contribution of local single samples, elucidating the synergistic effects of multivariate variables (such as RH saturation and high CTH synergistically promoting precipitation), breaking through the black-box limitations of traditional machine learning models and providing quantitative support for precipitation mechanism research. At the same time, the technical process constructed in this invention has broad reusability and can be directly transferred to the study of influencing factors of other meteorological elements such as temperature and wind speed, providing a standardized technical paradigm for the accurate simulation and mechanism analysis of meteorological elements. Attached Figure Description

[0021] Figure 1 A flowchart of a precipitation prediction method based on climate zoning and LGBM model provided in an embodiment of the present invention; Figure 2 This is a performance verification diagram of the LGBM model for precipitation simulation provided in an embodiment of the present invention; Figure 3 A global interpretation waterfall diagram of SHAP provided for embodiments of the present invention; Figure 4The SHAP global interpretation beehive graph provided in this embodiment of the invention; Figure 5 This is a SHAP and variable interaction dependency graph provided in an embodiment of the present invention; in, Figure 5 (a) in the graph represents the interaction dependency graph between SHAP and RH. Figure 5 (b) in the graph represents the interaction dependency graph between SHAP and variable KI; Figure 5 (c) in the graph represents the interaction dependency graph between SHAP and CAPE; Figure 5 (d) in the graph represents the interaction dependency graph between SHAP and IWP; Figure 5 (e) in the graph represents the interaction dependency graph between SHAP and CTH; Figure 5 In the graph, (f) represents the interaction dependency graph between SHAP and LST. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the present invention or its application or use. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of the invention.

[0024] At the same time, it should be understood that, for ease of description, the dimensions of the various parts shown in the accompanying drawings are not drawn according to actual scale.

[0025] Furthermore, for clarity and brevity, descriptions of well-known structures, functions, and configurations may have been omitted. Those skilled in the art will recognize that various changes and modifications can be made to the examples described herein without departing from the spirit and scope of this disclosure.

[0026] Techniques, methods, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and equipment should be considered part of the specification.

[0027] In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values.

[0028] Example 1 like Figure 1 As shown, this embodiment provides a precipitation prediction method based on climate zoning and the LGBM model, specifically including the following steps: S1. Obtain precipitation-related multi-source data covering the target area and construct a structured raw dataset; Precipitation-related multi-source data includes meteorological station observation data, satellite remote sensing data, atmospheric reanalysis data, and geographic auxiliary data; The precipitation-related multi-source data were filtered to obtain valid data. The valid data were then classified into several subsets according to data type. For each subset, a three-dimensional structured attribute containing timestamps, spatial latitude and longitude, and factor values ​​was constructed to obtain the structured raw dataset. The structured raw dataset includes a subset of core target data, a subset of thermal factor data, a subset of dynamic factor data, a subset of cloud microphysical factor data, and a subset of underlying surface and auxiliary factor data.

[0029] S2. Based on the climate zoning rules of the target area, the structured raw dataset is divided into climate zones and precipitation scenario-specific data preprocessing is performed to generate standardized precipitation modeling datasets for each climate zone. Specifically, the preprocessing process for precipitation scenario-specific data is as follows: S21. Perform spatiotemporal alignment and resolution unification on the structured original dataset. Unify the temporal window and spatial resolution of all subsets in the structured original dataset to the precipitation observation standard, while preserving topographic details and clarifying the correspondence between the interpolation method of each subset and the original resolution. The specific process of step S21 is as follows: The time windows of all subset data are unified to a daily scale that matches the precipitation observation period, and the spatial resolution is unified to a preset resolution: For subsets with resolutions lower than the preset resolution, bilinear interpolation is used to increase the resolution of the subsets. For subsets with resolutions higher than or equal to the preset resolution, nearest neighbor interpolation is used to decrease the resolution of the subsets, while preserving terrain and land use details. Precipitation scenario-specific factors were added to the underlying surface and auxiliary factor data subsets after resolution unification.

[0030] The resolution-unified subset of underlying surface and auxiliary factor data includes DEM data at a preset resolution. The topographic relief F is calculated based on the DEM data at the preset resolution. h Topographic relief is used as a specific factor for precipitation scenarios. Topographic relief F h The calculation process is as follows: F hIt equals the maximum elevation within the rectangular window minus the minimum elevation within the rectangular window, where the rectangular window is the adjacent raster range of the DEM data at the preset resolution.

[0031] S22. After spatiotemporal alignment and resolution unification, the core target data subset is cleaned by threshold filtering and hierarchical sampling strategy for precipitation data cleaning. S23. Derive collaborative features of dynamic factors from the subset of dynamic factor data after spatiotemporal alignment and resolution unification; The dynamic factor data subset includes wind speed (WS) and wind direction (WD) from the ERA5 dataset. The specific process for deriving the dynamic factor co-features is as follows: Based on wind speed WS and wind direction WD, calculate the wind speed-wind direction synergy factor F used to quantify water vapor transport intensity. ws_wd F ws_wd =cos[WS×(WD-180°)].

[0032] S24. Impute missing values ​​for each subset after processing in steps S21-S23; The specific process of missing value imputation is as follows: The method employs climate zone mean combined with inverse distance weighted interpolation (IDW). First, the mean factor of the original structured dataset is calculated as the initial value for missing values ​​based on the climate zone to which it belongs. Then, the IDW method is used to correct this by combining the effective factor values ​​of the three surrounding raster cells at preset resolutions. The final missing value V is equal to the sum of the products of the effective factor values ​​of the three surrounding raster cells and the inverse square of their corresponding distances, divided by the sum of the inverse squares of the distances of the three surrounding raster cells. This is expressed as: V= ; in, v i Indicates the first i The effective factor value of each surrounding raster. di Indicates the first i The distance from each surrounding grid cell to the location of the missing value.

[0033] S25. Factor standardization and global spatiotemporal fusion are performed on each subset after processing in step S24 to obtain single-table structured data. The specific process of obtaining structured data from a single table is as follows: Dimensional normalization: For all factors in each subset except the core target data subset after processing in step S24, Z-score normalization is applied to adapt to the normal distribution characteristics of precipitation factors and avoid extreme values ​​distorting the data distribution. Global spatiotemporal fusion: Based on daily-scale timestamps and preset resolution latitude and longitude raster, the processed subsets of thermal factor data, dynamic factor data, cloud microphysical factor data, underlying surface and auxiliary factor data are used as factor columns and fused with the target column of the core target data subset to form a single-table structured data.

[0034] S26. Divide the single-table structured data to obtain standardized precipitation modeling datasets for each climate zone.

[0035] S3. Based on the standardized precipitation modeling datasets of each climate zone, LGBM precipitation prediction models are trained separately. Bayesian optimization is used to independently tune the core hyperparameters of the LGBM precipitation prediction models of each climate zone. The models are then validated by time series cross-validation to generate the optimal precipitation prediction model that is adapted to each climate zone. Specifically, step S3, which generates the optimal precipitation prediction model adapted to each climate zone, includes: S31. Construct the LGBM base model, retain the original GOSS gradient one-sided sampling and EFB exclusive feature binding core mechanism of the LGBM base model, and initialize the initial training parameters of the LGBM base model. S32. Based on the precipitation characteristics of the standardized precipitation modeling dataset, adjust the retention ratio of high-gradient samples and the sampling ratio of low-gradient samples in the GOSS gradient one-sided sampling; and according to the precipitation physical mechanism classification, bind the high-dimensional factors in the standardized precipitation modeling dataset to obtain the LGBM precipitation prediction model. S33. Set the core hyperparameter search range for the LGBM precipitation prediction model for precipitation scenarios, and use Bayesian optimization method to iteratively optimize the model with the goal of minimizing the root mean square error and maximizing the coefficient of determination. S34. Use the time series ten-fold cross-validation method to verify the effectiveness of the optimal hyperparameter combination, ensuring that the training set time is earlier than the validation set time: S35. Based on the validated optimal hyperparameter combination, the LGBM precipitation prediction model is trained using the full training set to obtain the optimal precipitation prediction model.

[0036] S4. Obtain precipitation-related data for the target area within the forecast period and divide it according to climate zones. Input the corresponding optimal precipitation prediction model to generate precipitation prediction results for each climate zone and merge them according to spatial location to form a complete precipitation prediction distribution map of the target area.

[0037] S5. Based on the preset climate zoning rules, perform zoning verification on the standardized precipitation modeling dataset and the optimal precipitation prediction model for each climate zone, and output the targeted explanation results of the zoning precipitation mechanism. S6. Embed the optimal precipitation prediction model into the SHAP interpretation framework, perform multi-dimensional SHAP analysis on the standardized precipitation modeling datasets of each climate zone, and output SHAP targeted interpretation results. S7. The results of the targeted interpretation of regional precipitation mechanisms and the results of the SHAP targeted interpretation are structurally integrated to construct a comparative system of global-regional precipitation driving mechanisms, and to clarify the common driving laws and regional unique regulation mechanisms of precipitation in different climate zones.

[0038] Understandably, current research on regional precipitation influencing factors mainly relies on traditional statistical methods and conventional machine learning methods. Both of these techniques and research frameworks have significant limitations, making it difficult to meet the needs of refined, mechanistic research and application transformation. Traditional statistical methods, based on the core assumption of linear correlation, cannot capture the nonlinear relationship and threshold effect between precipitation and influencing factors. They also lack the ability to handle high-dimensional variables and are difficult to characterize small-scale local precipitation features. Conventional machine learning methods suffer from "black box" defects, failing to accurately quantify factor contributions and interaction effects, lacking local interpretability, and their generalization ability is easily affected by data distribution.

[0039] Meanwhile, existing research suffers from systemic defects such as scale limitations, one-sided variable selection, neglect of spatial variation patterns, and lack of a nonlinear collaborative quantification framework. These defects result in an understanding of precipitation formation mechanisms that remains at the level of phenomenological correlation, making it difficult to support the formulation of regionally differentiated disaster prevention and mitigation strategies.

[0040] This invention aims to address the aforementioned technical bottlenecks and framework deficiencies, focusing on overcoming three main categories of problems: First, breaking through the linear constraints of traditional statistical methods to accurately capture the synergistic effects of multiple factors and threshold features; second, addressing the black-box problem of conventional machine learning methods by constructing a locally interpretable system that quantifies the contribution ratio of factors, positive and negative effects, and regionally differentiated thresholds, clarifying the multi-factor interaction mechanism and enhancing the generalization stability of the model; and third, improving the overall research framework by integrating four core factors—thermal, dynamic, cloud microphysics, and underlying surface—to reveal the spatial variation patterns of factor effects, construct a multi-factor nonlinear interaction quantification system, clarify the inherent logical chain of factors-process-precipitation, and ultimately form a precipitation mechanism analysis technique adapted to complex climate and topographic features, providing support for improving precipitation prediction accuracy and formulating disaster prevention and mitigation strategies.

[0041] To better illustrate the accuracy of the model constructed in this application, a quantitative interpretation method for precipitation influencing factors is presented. The core of this interpretation method is an integrated technical solution that combines multi-source precipitation correlation data fusion, precipitation scenario-specific preprocessing, LGBM model precipitation adaptation optimization, SHAP precipitation mechanism targeted interpretation, and climate zone regional validation. This solution addresses the core characteristics of precipitation prediction, such as nonlinearity, high spatiotemporal heterogeneity, and multi-factor synergistic driving forces, solving the entire process from data input to mechanism output. The specific steps are as follows: Step 1: Constructing the Structured Original Dataset To meet the requirements of capturing the spatiotemporal characteristics of precipitation and the representativeness of climate, multi-source precipitation data covering seven major climate regions were collected. All data sources underwent availability screening, eliminating invalid data sources with less than 90% temporal or spatial coverage, retaining only valid continuous time series data covering the seven major climate regions from 2000 to 2022. The filtered data was then organized into five subsets according to data type. Each subset was constructed with a three-dimensional structured attribute consisting of timestamps, spatial latitude and longitude, and factor values ​​to ensure subsequent spatiotemporal correlation and fusion, ultimately integrating into a structured raw dataset D0.

[0042] Each subset is specifically composed of data from 2000 to 2022, with the original spatial resolution and carrying a unique spatiotemporal tag: 1. Core target data subset D 0T The data consists of daily precipitation data from 2,481 meteorological stations of the Meteorological Bureau, in mm, including samples with full intensity gradients of no precipitation, light rain, moderate rain, heavy rain, and rainstorms, covering 7 major climate zones, and used as target variables for model training and validation.

[0043] 2. Subset D of thermal factor data 0H The TRIMS LST surface temperature dataset, in K; and the ERA5 dataset, containing sensible and latent heat fluxes, in W / m², used to capture the driving effect of surface-atmosphere heat exchange on precipitation.

[0044] 3. Dynamic factor data subset D 0D The ERA5 dataset contains surface pressure (SP, hPa), wind speed (WS, m / s), wind direction (WD, °), K-index (KI, K), and convective available potential energy (CAPE, J / kg), used to quantify dynamic conditions such as vertical atmospheric motion and stability.

[0045] 4. Cloud microphysical factor data subset D 0C The CLARA-A3 cloud attribute dataset includes cloud top temperature (CTT) in Kelvin, cloud top height (CTH) in meters, ice water path (IWP) and liquid water path (LWP) in kg / m², directly depicting the intra-cloud processes of precipitation formation.

[0046] 5. Subset D of underlying surface and auxiliary factor data 0U SRTM 90m resolution DEM data, which can extract elevation, slope and aspect; 0.05°×0.05° NDVI dataset, reflecting vegetation cover; 30m resolution CLCD land use classification data, distinguishing between cultivated land, forest land, urban land and other types; 1km resolution AOD dataset, representing aerosol concentration, which affects cloud condensation nucleus formation.

[0047] This step outputs a structured original dataset D0, containing D 0T D 0H D 0D D 0C D 0U Five structured subsets with spatiotemporal labels serve as the sole input for step 2.

[0048] Step 2: Precipitation scenario-specific data preprocessing (Input: D0 → Processing: Spatiotemporal alignment / cleaning / derivation / standardization / partitioning → Output: D1) To address the spatiotemporal heterogeneity, multi-factor dimensional differences, and precipitation intensity threshold effects of precipitation data, unified spatiotemporal alignment, targeted sub-processing, global standardization, and dataset partitioning were performed on the five subsets of D0. All processing was based on timestamps and spatial latitude and longitude to achieve spatiotemporal correlation and fusion of the five subsets, ultimately outputting a single-table structured standardized modeling dataset. Each row corresponds to a spatiotemporal sample, each column corresponds to a precipitation correlation factor, and the last column is the target variable, i.e., the daily precipitation value.

[0049] Step 2.1: Spatiotemporal alignment and resolution unification (execute on all subsets of D0) The temporal window and spatial resolution of each subset of D0 are unified with the precipitation observation standard, while preserving topographic details and clarifying the correspondence between the interpolation method of each subset and the original resolution: 1. Time Alignment: Unify the time window of all subsets to 20:00 Beijing time to 20:00 the next day, matching the daily precipitation observation period of the meteorological station. For hourly granular data, such as D... 0H D 0D The hourly ERA5 data was used to calculate a 24-hour average and convert it to a daily scale; for data at the daily scale and above, the timestamps were directly calibrated. After processing, all subsets had a daily time granularity and completely consistent timestamps.

[0050] 2. Spatial Alignment: Unify the spatial resolution of all subsets to 1km × 1km. For subsets with a resolution lower than 1km, including D... 0H D 0D D 0C Bilinear interpolation is used to improve resolution and ensure spatial continuity of meteorological factors; for subsets D with a resolution of 1 km or higher... 0U The 90m DEM and 30m CLCD use nearest neighbor interpolation to reduce resolution while preserving topographic and land use details, while the 1km AOD and 0.05°NDVI directly retain 1km resolution.

[0051] 3. Derivation of precipitation-specific topographic features: Regarding D 0U Calculate the topographic relief F from the interpolated 1km DEM data. hThis is a factor specific to precipitation scenarios. Plain text calculation method: F h It equals the maximum elevation within a 3×3 window minus the minimum elevation within a 3×3 window. The 3×3 window represents the adjacent raster range of a 1km resolution DEM data point. After calculation, F... h Added as a new factor to D 0U .

[0052] Step 2.2: Special cleaning of precipitation data (only for the subset D of the target variable) 0T implement) Regarding D 0T To address the issues of an excessively high percentage of samples from days with no precipitation and the presence of outliers due to instrument malfunctions, a threshold screening + stratified sampling strategy was employed. 1. Removal of outlier zero values: If a weather station has 30 consecutive days of zero values... 0T The value was 0, but three other stations in the same climate zone around it all had precipitation during the same period. This was determined to be an abnormal 0 value caused by instrument malfunction, and the sample was directly removed.

[0053] 2. Stratified sampling to balance the sample: D 0T The precipitation intensity was divided into 5 levels: no precipitation (0 mm), light rain (0.1-10 mm), moderate rain (10.1-25 mm), heavy rain (25.1-50 mm), and torrential rain (greater than 50 mm). Proportional sampling was performed for each level to ensure that the proportion of samples of each intensity in the dataset was balanced and to avoid the model being biased towards the prediction of no precipitation samples.

[0054] Step 2.3: Derivation of precipitation-specific features and imputation of missing values ​​(performed on all subsets of D0) 1. Derivation of synergistic characteristics of dynamic factors: for D 0D Calculate the wind speed-wind direction synergy factor F from the given wind speed (WS) and wind direction (WD). ws_wd , used to quantify water vapor transport intensity. F ws_wd It equals WS multiplied by the cosine of (WD minus 180°). After calculation, F... ws_wd Added as a new factor to D 0D .

[0055] 2. Missing Value Imputation: A climate zone mean combined with inverse distance weighted interpolation (IDW) method was used. First, the mean factor values ​​for the seven major climate zones to which the samples belonged were calculated as initial missing values. Then, the IDW method was used to correct these missing values ​​by incorporating the effective factor values ​​from the three surrounding 1km grids. The final imputed value was equal to the sum of the products of the effective factor values ​​of the three surrounding grids and the inverse square of their corresponding distances, divided by the sum of the inverse squares of the distances from the three surrounding grids. After processing, the percentage of missing values ​​in all subsets of D0 was less than 0.1%, and the data integrity met the modeling requirements.

[0056] Step 2.4: Factor standardization and global spatiotemporal fusion (performed on all subsets of D0) 1. Dimensional normalization: For D0, divide D by... 0T All factors other than the target variable are standardized using Z-scores to fit the normal distribution characteristics of precipitation factors and avoid extreme values ​​distorting the data distribution. x_std equals (x_ori minus u) divided by sigma. Where x_std is the standardized factor value, x_ori is the original factor value, u is the mean of the factor across the entire dataset, and sigma is the standard deviation of the factor across the entire dataset. After standardization, all factors have uniform dimensions, a mean of 0, and a variance of 1.

[0057] 2. Global Spatiotemporal Fusion: Based on daily-scale timestamps + 1km latitude and longitude grids, the processed D... 0H D 0D D 0C D 0U As a factor column, with D 0T The target columns are merged to form a single table of structured data.

[0058] Step 2.5: Dataset Partitioning (Executed on the merged single-table data) The merged single-table structured data is divided into training set D in an 8:2 ratio. 1train and verification set D 1val The training and validation sets were stored in a standardized precipitation modeling dataset for subsequent model training. The splitting process satisfied a double consistency constraint: the distribution of samples from the seven major climate zones and the five precipitation intensity levels in the training and validation sets were completely consistent, avoiding generalization bias caused by distribution imbalance. The final training set consisted of approximately 7.6 million samples, and the validation set consisted of approximately 1.9 million samples.

[0059] This step outputs the standardized precipitation modeling dataset D1, which includes the training set D. 1train Validation set D 1val This is the sole input for model training in step 3. The data is in a single-table structured format and can be directly imported into the LGBM model for training.

[0060] Step 3: LGBM Model Precipitation Adaptation Optimization (Input: D1 → Processing: Initialization + Structure Adaptation + Parameter Optimization + Validation → Output: M) opt ) Using D1 as the sole training and validation data, and based on the LGBM framework, model structure adaptation, Bayesian hyperparameter tuning, and time-series cross-validation were performed tailored to the characteristics of precipitation data. The entire process aimed to minimize RMSE and maximize R². 2 For a dual objective function, the data flow direction is clearly defined: D 1train Used for model training and parameter iteration, D 1val Used for model validation and performance evaluation, avoiding leakage of time information throughout the process.

[0061] Step 3.1: Initialization of the LGBM basic model The original GOSS gradient one-sided sampling and EFB exclusive feature binding core mechanism of LGBM are retained. The basic model Minit is initialized and the core framework parameters are set to an optimizable state. The initial parameters adopt industry-standard values, including 31 leaf nodes, 10 tree depth, and 0.1 learning rate.

[0062] Step 3.2: Adaptation of precipitation data to a specific model structure (for M) init (Adjustment, parameterless iteration) Based on the characteristics of precipitation data, such as a low proportion of extreme samples and a clear physical mechanism of the factors, M was adjusted. init The core logic of GOSS and EFB: 1. GOSS Sampling Optimization: To adapt to the low proportion of extreme precipitation samples, the high gradient sample retention ratio a_ratio was adjusted from 20% to 30% in the conventional groundwater scenario, and the low gradient sample sampling ratio b_ratio was adjusted from 10% to 15%, ensuring that the model fully captures the high gradient features of extreme precipitation.

[0063] 2. EFB Feature Bundling Adjustment: Based on the precipitation physical mechanism classification, high-dimensional factors in D1 are bundled to replace the conventional classification method, avoiding information loss caused by bundling factors of different physical mechanisms. Specifically, they are divided into four groups: the convection potential group bundles dynamic factors such as CAPE and KI; the cloud structure group bundles cloud microphysical factors such as CTT, CTH, and IWP; the heat exchange group bundles thermal factors such as LST and sensible heat flux; and the underlying surface group bundles underlying surface factors such as altitude, Fh, and NDVI.

[0064] Step 3.3: Bayesian hyperparameter tuning Considering the characteristics of precipitation data, such as multi-factor nonlinear interaction, millions of samples, and weak multicollinearity, a search range for core hyperparameters was defined, and iterative tuning was carried out using Bayesian optimization methods. 1. Core hyperparameter search range, customized for precipitation scenarios: number of leaf nodes ranges from 30 to 80, capturing nonlinear relationships in precipitation through more leaf nodes; tree depth ranges from 50 to 70, characterizing multi-factor collaborative mechanisms through deeper tree depth; learning rate ranges from 0.05 to 0.15, balancing training efficiency and accuracy with a moderate learning rate; number of decision trees ranges from 200 to 300, fitting multi-factor interactions through more decision trees to avoid underfitting the model; L1 regularization strength ranges from 0.01 to 0.05, reducing the regularization strength to preserve factor interaction information.

[0065] 2. Iterative optimization process: The first step is to select a set of parameters from the parameter search range and substitute them into M. init The model to be validated is obtained; the second step is to use D... 1trainTrain the model to be validated using D 1val Computational models RMSE and R 2 The third step is to update the parameter search direction based on the Bayesian posterior probability distribution and select the next set of better parameters; repeat the above steps until RMSE and R0 are achieved. 2 Numerical convergence yields the optimal combination of hyperparameters.

[0066] Step 3.4: 10-fold cross-validation of time series data To address the time-series correlation characteristics of precipitation data, a 10-fold time-series cross-validation method is used instead of random cross-validation to prevent leakage of time information. (D) 1train The dataset is divided into 10 consecutive subsets in chronological order. Nine subsets are selected as the training set and one subset as the validation set at each iteration, ensuring that the training set's execution precedes the validation set's execution throughout the process. If the average RMSE and R² of 10-fold validation are... 2 If the result is consistent with the convergence result, the optimal parameter combination is determined to be valid; if the result is inconsistent, return to the Bayesian hyperparameter tuning step to re-optimize.

[0067] Step 3.5: Determining the Optimal Model Based on the validated optimal hyperparameter combination, using the full D... 1train Conduct model training using D 1val Model validation was conducted, and the precipitation-adapted optimized LGBM model M was finally obtained. opt The optimal hyperparameters determined through optimization are: 56 leaf nodes, 60 tree depth, 0.09 learning rate, 233 decision trees, and 0.02 L1 regularization strength.

[0068] This step outputs the optimal precipitation prediction model (precipitation-adaptive optimized LGBM model) M opt The trained end-to-end prediction model is the core input for steps 4 and 5.

[0069] Step 4: Targeted explanation of SHAP precipitation mechanism (Input: M) opt +D1→Processing: Multi-dimensional SHAP analysis + Physical verification→Output: S all ).

[0070] M opt Embedding the SHAP interpretation framework and using D1 as the analysis data, this study designs four precipitation-specific interpretation dimensions based on the three-element theory of water vapor-dynamics-cloud microphysics in precipitation formation. By combining data-driven analysis results with precipitation climatology theory, it avoids the interpretation bias of purely data-driven approaches and ultimately outputs interpretation results of precipitation mechanisms at the national scale.

[0071] Step 4.1: Identification of core driving factors at the national scale Input D1 into M optThe absolute SHAP values ​​of all precipitation-related factors were calculated, and the factors were sorted from high to low according to their absolute SHAP values. The core driving factors of precipitation at the national scale, such as relative humidity (RH), convective available potential energy (CAPE), K index (KI), and ice-water pathway (IWP), were identified, and the contribution ratio of each factor was quantified.

[0072] Step 4.2: Precipitation-Specific 4D SHAP Analysis A four-dimensional specific analysis was conducted to address the threshold triggering, multi-factor synergy, and positive-negative effect differentiation characteristics of precipitation, and the rationality of the results was verified by combining physical mechanisms. 1. Quantification of Positive and Negative Effects: The SHAP values ​​of each core factor are calculated using SHAP waterfall plots to clarify the direction of the factor's effect on precipitation and distinguish between positive and negative contributions. For example, surface temperature (LST) has a positive contribution, as high LST can enhance surface evaporation and provide sufficient water vapor; cloud top height (CTH) has a negative contribution, as a high CTH means that the cloud top is too high, making it difficult for water vapor to settle and form precipitation.

[0073] 2. Threshold Effect Identification: Using the SHAP swarm diagram, the critical thresholds for each core factor's influence on precipitation are determined. For example, when the relative humidity (RH) is greater than 80%, the SHAP value is significantly positive, and the probability of precipitation increases dramatically; when the convective available potential energy (CAPE) is greater than 1000 J / kg, the promoting effect on extreme precipitation reaches saturation.

[0074] 3. Multi-factor interaction analysis: By using the SHAP dependency graph, we can uncover the factor interaction mechanisms unique to precipitation scenarios. For example, the synergistic effect of relative humidity (RH) and cloud top height (CTH) is observed. When RH is saturated and CTH is high, the SHAP value is significantly positive, and the two work together to promote precipitation. When RH is saturated but CTH is low, the SHAP value is close to 0, and there is no significant synergistic effect.

[0075] 4. Physical mechanism correlation verification: The above analysis results are compared with the three-element theory of water vapor-dynamic-cloud microphysical precipitation to verify the physical rationality of core factors, effect direction, threshold range, and interaction mechanism, and to eliminate false correlations generated by pure data-driven approaches.

[0076] This step of the model outputs the SHAP-targeted interpretation results of the national-scale precipitation mechanism. all It includes the ranking of core driving factors, the positive and negative effects of factors, the critical threshold, the multi-factor interaction mechanism, and the physical mechanism verification conclusions, which serve as the benchmark reference for the partition verification in step 5.

[0077] Step 5: Climate Zoning Verification (Input: M) opt +D1+R cli →Processing: Partition slicing + precision verification + partition SHAP analysis →Output: S cli ).

[0078] To address the strong regional heterogeneity of precipitation, this study uses climate zones as independent validation units and employs automated batch execution logic to complete dataset partitioning, model generalization accuracy validation, and partition SHAP mechanism interpretation. It also integrates regional difference results to construct a national-regional integrated precipitation mechanism interpretation system.

[0079] Step 5.1: Automated Slicing of Dataset Partitions According to partitioning rule R cli The spatial boundary of D1 is divided into 7 independent and non-overlapping climate zone subsets D1 by latitude and longitude spatial matching and overlay analysis. 1clin n takes values ​​from 1 to 7, corresponding to the seven major climate zones. During the slicing process, the precipitation intensity stratification structure and factor integrity of each subset are preserved to ensure that the distribution of the zonal samples is consistent with the full sample.

[0080] Step 5.2: Verification of the generalization accuracy of the partitioning model M opt Input the 7 climate zone subsets D respectively 1clin Using the validation set data within each partition, the partition evaluation metric RMSE is calculated. clin and R 2 clin Execution accuracy determination and local optimization: 1. Accuracy determination: If R 2 clin ≥0.7 and RMSE clin If the accuracy is below the preset threshold, determine M. opt It exhibits good generalization ability in this region and requires no secondary optimization.

[0081] 2. Local optimization: If the accuracy of a certain climate zone is not up to standard, the Bayesian optimization and time series cross-validation process in step 3 is reused based on the samples of that zone to fine-tune the hyperparameters of the model in the region, so as to obtain a sub-model specifically adapted to that climate zone and ensure the uniformity of prediction accuracy across the entire region.

[0082] Step 5.3: Explanation of the Partition SHAP Targeting Mechanism For each climate zone subset D 1clin The SHAP interpretation process in step 4 is fully reproduced to achieve a refined analysis of regional precipitation mechanisms. 1. Identify region-specific core driving factors and conduct horizontal comparisons with core factors at the national scale to extract regional differences. For example, the CAPE contribution is prominent in the northern subtropical humid region, relative humidity (RH) dominates in the marginal tropical humid region, and cloud top height (CTH) plays a key regulatory role in the plateau temperate semi-arid region.

[0083] 2. Quantify the positive and negative contributions of factors within the region and the critical trigger thresholds, and explain the differences in conjunction with the regional climate background. For example, the effective precipitation trigger threshold of relative humidity in arid areas is significantly higher than that in humid areas; the CAPE saturation threshold for promoting extreme precipitation in plateau areas is lower than that in plain areas.

[0084] 3. Explore the unique multi-factor synergistic patterns in each climate zone, and verify the physical rationality of the interaction mechanism by combining regional characteristics such as monsoon, topography, and continental climate.

[0085] 4. Compare the regional interpretation results with regional precipitation climatology theories, eliminate spurious correlations, and ensure that the conclusions conform to the laws of regional atmospheric motion, water vapor transport, and cloud microphysical processes.

[0086] Step 5.4: Integration and Comparative Analysis of Partition Results By structurally integrating the precision indicators, SHAP interpretation results, and factor difference patterns of seven climate zones, a comparative system of national-regional precipitation driving mechanisms is constructed to clarify the common driving patterns and regional unique regulation mechanisms of precipitation in different climate zones.

[0087] This step of the model outputs the SHAP-targeted interpretation results of precipitation mechanisms in 7 major climate zones. cli It includes model accuracy indicators for each climate zone, ranking of regional core driving factors, positive and negative effects of factors and critical thresholds, multi-factor interaction patterns, conclusions on national-regional differences, and physical mechanism verification reports, which are the final standardized output results of this invention.

[0088] like Figure 2 As shown, this plot displays LGBM scatter plots of observed and simulated precipitation, along with various evaluation indicators. The horizontal axis represents observed precipitation (mm), and the vertical axis represents simulated precipitation (mm). The precipitation data are daily precipitation data, with a sample size N of 9,540,660. △ indicates that the 99% confidence interval test has been passed. The color of the points represents density; darker colors indicate higher point density. R0 2 For fixed coefficients, RMSE is the root mean square error, MAE is the mean absolute error, Bias is the deviation, and KGE is the Kling-Gupta efficiency, where R... 2 The closer KGE is to 1, the better the model performance; the closer RMSE, MAE, and Bias are to 0, the better the model performance. Figure 3The waterfall plot clearly shows that the final predicted value (6.81) of the LGBM model nationwide was lower than the baseline value (7.027). The core value of this waterfall plot lies in the fact that it not only shows the final difference between the predicted value and the baseline value, but also quantifies the influence of each meteorological variable on the precipitation forecast through a visualization of the superimposed contributions of variables, making the strength of the interaction between variables and the overall effect immediately clear. From the quantitative results of the waterfall plot, the precipitation process is indeed the result of the combined effect of multiple meteorological variables, and the magnitude and direction of the influence of each variable are clearly identifiable: in terms of water vapor content, LST, E, WD, and RH all contributed positive values. The superposition of the positive contributions of these variables together provided a sufficient water vapor basis for precipitation, becoming an important force driving precipitation; while WS is shown as a weak negative contribution value in the waterfall plot, slightly inhibiting precipitation, and its influence is much smaller than that of water vapor supply variables. At the level of convective stability, LST and SP both show positive contributions, promoting precipitation by enhancing convective development. However, KI, CTH, and CAPE show significant negative contributions. The combined negative effects of these variables offset some of the convective-promoting effects, becoming key factors restricting the effective triggering and maintenance of convection. In intracloud condensation nuclei and microphysical processes, LWP, CFC, and AOD all contribute significant positive values, providing material and process support for precipitation particle collision, coalescence, and condensation growth, and are the core contributors to precipitation particle formation. IWP and CTT, on the other hand, show negative contributions, and their inhibitory effects contrast with the aforementioned positive contributions in the waterfall plot. The waterfall plot's specific quantification of the positive / negative contributing factors of each variable clearly shows that the variables, through the superposition of their positive and negative contributions of varying magnitudes, collectively influence the occurrence and development of precipitation. Furthermore, the quantification results clearly indicate that variables directly related to water vapor conditions and convective dynamics, due to their significant positive contributions, become the core driving factors of the model output. Other variables, either due to their weak negative contributions, the cancellation of their positive contributions, or the indirect nature of physical mechanisms, variable redundancy, or the characteristics of regional climate scenarios, exhibit smaller contributions in the waterfall plot, resulting in a relatively weak overall contribution to precipitation. In addition, Figure 4 The SHAP swarm diagram further corroborates the above description from the perspective of the correlation between the level of eigenvalue contribution and the direction of influence: variables such as IWP contribute significant positive SHAP values ​​when the eigenvalue contribution is high, strongly promoting precipitation forecasting; while variables such as KI and CTH show obvious negative SHAP values ​​when the eigenvalue contribution level is low, inhibiting precipitation. This indicates that variables directly related to water vapor and convective dynamics are the core drivers not only in terms of the superposition of contribution magnitudes, but also in terms of the continuous key role in the correlation between the level of eigenvalue and influence; while the contributions of other variables to precipitation are relatively weak overall.

[0089] Figure 5This is a SHAP variable interaction dependency plot, which can intuitively present the interaction relationships between various meteorological variables and their synergistic impact on precipitation forecast results. From Figure 5 As shown in (a), when RH is low, the SHAP value is mostly negative or a small positive value. In this case, low RH and low CTH have a suppressive or weakly promoting effect on precipitation. This is because low RH indicates insufficient atmospheric water vapor, while low CTH reflects weak vertical cloud development, lacking both water vapor and vertical dynamics, making effective precipitation difficult. When RH approaches saturation, high RH combined with high CTH makes a strong positive contribution to precipitation. This may be because high RH provides sufficient water vapor condensation conditions, while high CTH reflects sufficient vertical cloud development, allowing water vapor to efficiently transform into precipitation through condensation, collision, and coalescence in the vertical direction. The two work together to stimulate significant precipitation potential. However, when RH is high but CTH is low, the positive contribution to precipitation is significantly weakened, or even turns into a suppressive effect. This may be because although high RH brings sufficient water vapor, low CTH restricts vertical cloud development, making it difficult for water vapor to fully develop into effective precipitation cloud systems. Condensed water vapor cannot efficiently transform into precipitation particles, resulting in ineffective precipitation and the insufficient utilization of ample water vapor conditions. Figure 5 Images (b) and (c) together reveal the synergistic effect of KI and CAPE on precipitation: when both KI and CAPE are low, the atmosphere lacks both unstable structure and convective energy, and the corresponding regions in both images show mostly negative or small positive values, exhibiting an inhibitory or weakly promoting effect on precipitation; however, when both KI and CAPE are high, the strong unstable structure provided by high KI and the ample energy contained in high CAPE work synergistically, enabling water vapor to be efficiently converted into precipitation, and the overlapping high-value regions in both images show large positive values, making a strong positive contribution to precipitation; if only one factor is high, due to the lack of cooperation from the other factor, convection cannot develop efficiently, water vapor cannot be fully converted into precipitation, and the positive contribution to precipitation will be significantly weakened. Figure 5As shown in (d), when IWP is low, it mainly inhibits or weakly promotes precipitation. This is because low IWP means insufficient ice-phase water vapor content in the cloud, while low CTH reflects weak vertical cloud development, lacking both ice-phase water vapor foundation and vertical dynamics. Water vapor cannot efficiently transform into precipitation through processes such as condensation, collision, and freezing, making precipitation difficult to occur effectively. When both IWP and CTH are high, the combination of high IWP and high CTH has a strong positive contribution to precipitation. This may be because high IWP provides sufficient conditions for ice-phase water vapor condensation, while high CTH reflects sufficient vertical cloud development. Ice-phase particles can efficiently grow and transform into precipitation particles in the vertical direction through processes such as condensation, collision, and frost, and the two synergistically stimulate precipitation potential. When IWP is high but CTH is low, the positive contribution to precipitation is significantly weakened. This may be because although high IWP brings sufficient ice-phase water vapor, low CTH limits the vertical development of clouds. Ice-phase particles cannot grow sufficiently in the vertical direction, and water vapor cannot fully develop into an effective precipitation cloud system, resulting in ineffective precipitation. The sufficient ice-phase water vapor conditions are not fully utilized. Figure 5 In (e), when CTH is low and CTT is high, the effect on precipitation is mostly inhibitory or weakly promoting. This is because low CTH reflects weak vertical cloud development, while high CTT means that the atmospheric temperature at the cloud top is warmer, making it difficult for water vapor to cool sufficiently, which is not conducive to the efficient formation of precipitation particles, thus making it difficult for precipitation to occur effectively. When CTH is high and CTT is low, the combination of high CTH and low CTT has a strong positive contribution to precipitation. This may be because high CTH provides ample space for vertical cloud development, and low CTT means that the cloud top is at a lower temperature at high altitude, where water vapor can easily cool, condense, or even freeze. Both create favorable conditions for the formation and growth of precipitation particles, thereby stimulating precipitation potential. However, when CTH is high but CTT is also high, the positive contribution to precipitation is significantly weakened. This may be because although high CTH provides space for vertical cloud development, high CTT limits the cooling efficiency of water vapor at high altitudes, making it difficult for water vapor to grow efficiently into large precipitation particles through the ice phase process. The condensed water vapor cannot be fully converted into effective precipitation, making it difficult for precipitation to occur effectively. Figure 5In (f) of the study, when both LST and KI are low, they primarily inhibit or weakly promote precipitation. This is because low LST results in less heat transfer from the surface to the atmosphere, weakening surface evaporation and leading to a lack of water vapor. Low KI reflects a stable vertical structure of atmospheric temperature and humidity, indicating a lack of both sufficient water vapor and unstable conditions conducive to vertical water vapor movement and condensation, thus hindering effective precipitation. When both LST and KI are high, they make a strong positive contribution to precipitation. This is because high LST results in more heat transfer from the surface to the atmosphere, which further moistens the lower atmosphere through enhanced evaporation; while high KI indicates significant atmospheric instability. The combination of sufficient water vapor and a strong unstable structure makes air masses easily lifted and continue to rise, fully stimulating precipitation potential. When LST is high but KI is low, the positive contribution to precipitation is significantly weakened. This is because although the high LST brings low-level water vapor and energy, the low KI restricts the development of vertically unstable atmospheric structures, making it difficult for sufficient water vapor to fully participate in the convection process in the vertical direction. As a result, low-level water vapor and energy cannot be efficiently converted into precipitation particles, making it difficult for precipitation to occur effectively.

[0090] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Based on the technical essence of the present invention, any simple modifications, equivalent substitutions, and improvements made to the above embodiments within the spirit and principles of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A precipitation prediction method based on climate zoning and the LGBM model, characterized in that, Specifically, the following steps are included: S1. Obtain precipitation-related multi-source data covering the target area and construct a structured raw dataset; S2. Based on the climate zoning rules of the target area, the structured raw dataset is divided into climate zones and precipitation scenario-specific data preprocessing is performed to generate standardized precipitation modeling datasets for each climate zone. S3. Based on the standardized precipitation modeling datasets of each climate zone, LGBM precipitation prediction models are trained separately. Bayesian optimization is used to independently tune the core hyperparameters of the LGBM precipitation prediction models of each climate zone. The models are then validated by time series cross-validation to generate the optimal precipitation prediction model that is adapted to each climate zone. S4. Obtain precipitation-related data for the target area within the forecast period and divide it according to climate zones. Input the corresponding optimal precipitation prediction model to generate precipitation prediction results for each climate zone and merge them according to spatial location to form a complete precipitation prediction distribution map of the target area.

2. The precipitation prediction method based on climate zoning and the LGBM model according to claim 1, characterized in that, Precipitation-related multi-source data includes meteorological station observation data, satellite remote sensing data, atmospheric reanalysis data, and geographic auxiliary data; The precipitation-related multi-source data were filtered to obtain valid data. The valid data were then classified into several subsets according to data type. For each subset, a three-dimensional structured attribute containing timestamps, spatial latitude and longitude, and factor values ​​was constructed to obtain the structured raw dataset. The structured raw dataset includes a subset of core target data, a subset of thermal factor data, a subset of dynamic factor data, a subset of cloud microphysical factor data, and a subset of underlying surface and auxiliary factor data.

3. The precipitation prediction method based on climate zoning and the LGBM model according to claim 2, characterized in that, The specific process of preprocessing precipitation scenario-specific data is as follows: S21. Perform spatiotemporal alignment and resolution unification on the structured original dataset. Unify the temporal window and spatial resolution of all subsets in the structured original dataset to the precipitation observation standard, while preserving topographic details and clarifying the correspondence between the interpolation method of each subset and the original resolution. S22. After spatiotemporal alignment and resolution unification, the core target data subset is cleaned by threshold filtering and hierarchical sampling strategy for precipitation data cleaning. S23. Derive collaborative features of dynamic factors from the subset of dynamic factor data after spatiotemporal alignment and resolution unification; S24. Impute missing values ​​for each subset after processing in steps S21-S23; S25. Factor standardization and global spatiotemporal fusion are performed on each subset after processing in step S24 to obtain single-table structured data. S26. Divide the single-table structured data to obtain standardized precipitation modeling datasets for each climate zone.

4. The precipitation prediction method based on climate zoning and the LGBM model according to claim 3, characterized in that, The specific process of step S21 is as follows: The time windows of all subset data are unified to a daily scale that matches the precipitation observation period, and the spatial resolution is unified to a preset resolution: For subsets with resolutions lower than the preset resolution, bilinear interpolation is used to increase the resolution of the subsets. For subsets with resolutions higher than or equal to the preset resolution, nearest neighbor interpolation is used to decrease the resolution of the subsets, while preserving terrain and land use details. Precipitation scenario-specific factors were added to the underlying surface and auxiliary factor data subsets after resolution unification.

5. The precipitation prediction method based on climate zoning and the LGBM model according to claim 4, characterized in that, The resolution-unified subset of underlying surface and auxiliary factor data includes DEM data at a preset resolution. The topographic relief F is calculated based on the DEM data at the preset resolution. h Topographic relief is used as a specific factor for precipitation scenarios; Topographic relief F h The calculation process is as follows: F h It equals the maximum elevation within the rectangular window minus the minimum elevation within the rectangular window, where the rectangular window is the adjacent raster range of the DEM data at the preset resolution.

6. The precipitation prediction method based on climate zoning and the LGBM model according to claim 3, characterized in that, The dynamic factor data subset includes wind speed (WS) and wind direction (WD) from the ERA5 dataset. The specific process for deriving the dynamic factor co-features is as follows: Based on wind speed WS and wind direction WD, calculate the wind speed-wind direction synergy factor F used to quantify water vapor transport intensity. ws_wd F ws_wd =cos[WS×(WD-180°)].

7. The precipitation prediction method based on climate zoning and the LGBM model according to claim 3, characterized in that, The specific process of missing value imputation is as follows: The method employs climate zone mean combined with inverse distance weighted interpolation (IDW). First, the mean factor of the original structured dataset is calculated as the initial value for missing values ​​based on the climate zone to which it belongs. Then, the IDW method is used to correct this by combining the effective factor values ​​of the three surrounding raster cells at preset resolutions. The final missing value V is equal to the sum of the products of the effective factor values ​​of the three surrounding raster cells and the inverse square of their corresponding distances, divided by the sum of the inverse squares of the distances of the three surrounding raster cells. This is expressed as: V= ； in, v i Indicates the first i The effective factor value of each surrounding raster. di Indicates the first i The distance from each surrounding grid cell to the location of the missing value.

8. The precipitation prediction method based on climate zoning and the LGBM model according to claim 3, characterized in that, The specific process of obtaining structured data from a single table is as follows: Dimensional normalization: For all factors in each subset except the core target data subset after processing in step S24, Z-score normalization is applied to adapt to the normal distribution characteristics of precipitation factors and avoid extreme values ​​distorting the data distribution. Global spatiotemporal fusion: Based on daily-scale timestamps and preset resolution latitude and longitude raster, the processed subsets of thermal factor data, dynamic factor data, cloud microphysical factor data, underlying surface and auxiliary factor data are used as factor columns and fused with the target column of the core target data subset to form a single-table structured data.

9. A precipitation prediction method based on climate zoning and the LGBM model according to claim 8, characterized in that, Step S3, which generates the optimal precipitation prediction model adapted to each climate zone, specifically includes: S31. Construct the LGBM base model, retain the original GOSS gradient one-sided sampling and EFB exclusive feature binding core mechanism of the LGBM base model, and initialize the initial training parameters of the LGBM base model. S32. Based on the precipitation characteristics of the standardized precipitation modeling dataset, adjust the retention ratio of high-gradient samples and the sampling ratio of low-gradient samples in the GOSS gradient one-sided sampling; and according to the precipitation physical mechanism classification, bind the high-dimensional factors in the standardized precipitation modeling dataset to obtain the LGBM precipitation prediction model. S33. Set the core hyperparameter search range for the LGBM precipitation prediction model for precipitation scenarios, and use Bayesian optimization method to iteratively optimize the model with the goal of minimizing the root mean square error and maximizing the coefficient of determination. S34. Use the time series ten-fold cross-validation method to verify the effectiveness of the optimal hyperparameter combination, ensuring that the training set time is earlier than the validation set time: S35. Based on the validated optimal hyperparameter combination, the LGBM precipitation prediction model is trained using the full training set to obtain the optimal precipitation prediction model.

10. A precipitation prediction method based on climate zoning and the LGBM model according to claim 2, characterized in that, It also includes the following steps: S5. Based on the preset climate zoning rules, perform zoning verification on the standardized precipitation modeling dataset and the optimal precipitation prediction model for each climate zone, and output the targeted explanation results of the zoning precipitation mechanism. S6. Embed the optimal precipitation prediction model into the SHAP interpretation framework, perform multi-dimensional SHAP analysis on the standardized precipitation modeling datasets of each climate zone, and output SHAP targeted interpretation results. S7. The results of the targeted interpretation of regional precipitation mechanisms and the results of the SHAP targeted interpretation are structurally integrated to construct a comparative system of global-regional precipitation driving mechanisms, and to clarify the common driving laws and regional unique regulation mechanisms of precipitation in different climate zones.

Images