A dual-strategy enhanced local perception precipitation correction method and system

Abstract

The application belongs to the field of artificial intelligence, and specifically discloses a double-strategy enhanced local perception precipitation correction method and system. The method comprises the following steps: obtaining forecast precipitation data and auxiliary variable data of a research area to be corrected; determining the precipitation sub-area to which each grid in the research area belongs; classifying and mapping the forecast precipitation data and the auxiliary variable data, and identifying the precipitation occurrence state of each grid; outputting a zero-value result for the grid in the no-rain state; for the grid in the rain state, performing local noise suppression processing on the attention score matrix generated based on the multi-mode forecast precipitation data and the auxiliary variable data according to the corresponding precipitation quantitative correction model, and performing state transmission and feature reconstruction in the forward and reverse directions along the time sequence to obtain a quantitative correction result; and integrating the grids based on the zero-value result and the quantitative correction result, reconstructing and outputting target forecast precipitation data. Through the application, the expression capability of local precipitation structure and severe precipitation characteristics is improved.

Description

Technical Field

[0001] This application belongs to the field of artificial intelligence technology, specifically involving deep learning, numerical model precipitation correction and other technical fields, and more specifically, it relates to a precipitation correction method and system with dual strategies to enhance local perception. Background Technology

[0002] Precipitation forecasting plays a vital role in disaster prevention and mitigation, water resource allocation, and agricultural production. With the development of numerical models, multi-model ensemble forecasts of precipitation products have become increasingly abundant, providing an important data foundation for precipitation forecasting. However, due to factors such as model initial field errors and uncertainties in parameterization schemes, numerical model precipitation products exhibit systematic biases. Therefore, developing effective precipitation correction methods is of great significance.

[0003] Existing precipitation correction methods mainly fall into two categories: statistical post-processing and machine learning. Statistical post-processing methods correct biases by establishing a statistical mapping relationship between forecasts and observations, such as quantile mapping and frequency matching. Machine learning methods utilize historical data to train nonlinear mapping models, achieving end-to-end modeling from forecast variables to corrected precipitation. These methods typically treat the study area as a whole for unified modeling, using multi-model forecast precipitation and auxiliary variables as inputs, and outputting corrected precipitation results.

[0004] However, diurnal precipitation exhibits significant spatial heterogeneity and discontinuity in its temporal series, with a high proportion of rainless days. Existing unified modeling methods struggle to distinguish precipitation differences across regions and lack sufficient accuracy in identifying precipitation occurrence, leading to substantial biases in the spatial distribution and intensity forecasts of the correction results. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this application aims to provide a dual-strategy enhanced local perception precipitation correction method and system, which is intended to solve the problem that existing unified modeling methods are unable to characterize the spatial heterogeneity and temporal discontinuity of precipitation, resulting in insufficient accuracy in precipitation location and intensity forecasts.

[0006] The first aspect of this application relates to a dual-strategy enhanced local perception precipitation correction method, the method comprising: acquiring multi-model forecast precipitation data and auxiliary variable data of the study area to be corrected; determining the precipitation sub-region to which each grid in the study area belongs based on a pre-established spatial clustering partitioning model, wherein the spatial clustering partitioning model is generated by extracting features from historical baseline precipitation data and historical auxiliary variable data; classifying and mapping the multi-model forecast precipitation data and the auxiliary variable data, identifying the precipitation occurrence state grid by grid, wherein the precipitation occurrence state includes rainy state and rainless state; for grids in the rainless state, directly outputting zero-value results; for grids in the rainy state, performing local noise suppression processing on the attention score matrix generated based on the multi-model forecast precipitation data and the auxiliary variable data according to the precipitation quantitative correction model corresponding to its precipitation sub-region, and performing state transfer and feature reconstruction along the forward and reverse directions of the time series to calculate quantitative correction results; and performing grid integration based on the zero-value results and the quantitative correction results, reconstructing and outputting target forecast precipitation data.

[0007] This application first determines the precipitation sub-region to which each grid belongs through a spatial clustering model, then distinguishes between rainy and rainless states through classification mapping, and adopts a differentiated output strategy for different states, so that the precipitation correction process can simultaneously adapt to spatial heterogeneity and precipitation discontinuity. This avoids unnecessary intensity regression for rainless samples and reduces the impact of overall regional modeling on the smoothing of local precipitation structure, thereby improving the overall accuracy of precipitation occurrence state identification and target forecast precipitation data correction.

[0008] Furthermore, this application introduces local precipitation noise suppression, temporal forward and reverse state transfer, and feature reconstruction for rainy grids, enabling the model to not only weaken local noise in multi-source meteorological variables and attention distribution, but also extract the bidirectional evolution relationship of precipitation processes in the time dimension, which is beneficial to improving the quantitative correction stability under local heavy precipitation, edge precipitation, and complex weather processes.

[0009] In some implementations, the step of performing local noise suppression processing on the attention score matrix generated based on the multi-model forecast precipitation data and the auxiliary variable data according to the precipitation quantitative correction model corresponding to its precipitation sub-region, and performing state transfer and feature reconstruction along the forward and reverse time series to calculate the quantitative correction result includes: integrating the multi-model forecast precipitation data and the auxiliary variable data into an input sequence; using the encoder of the precipitation quantitative correction model, extracting the spatial interaction features of the input sequence through a self-attention mechanism, calculating local smoothing features based on the spatial interaction features, performing adaptive weighted fusion of the spatial interaction features and the local smoothing features to perform the local precipitation noise suppression processing while preserving the spatial distribution gradient of precipitation, and outputting encoded features; using the decoder of the precipitation quantitative correction model, extracting the bidirectional evolution features of the encoded features in the time dimension through a bidirectional long short-term memory neural network, performing multi-head cross-attention fusion of the extracted features and the encoded features, realizing the feature reconstruction through residual operation and dimension mapping, and calculating the quantitative correction result.

[0010] This scheme extracts spatial interaction features through the encoder and performs local smoothing feature fusion, enabling the model to suppress local noise while preserving precipitation spatial gradient and edge abrupt change information. Through the bidirectional long short-term memory neural network and residual reconstruction mechanism in the decoder, the ability to express the temporal evolution of precipitation is enhanced, thereby improving the accuracy of precipitation intensity correction for rainy samples.

[0011] In some implementations, the encoder utilizing the quantitative precipitation correction model extracts spatial interaction features from the input sequence through a self-attention mechanism, calculates local smoothing features based on these features, and adaptively weights and fuses the spatial interaction features with the local smoothing features to perform local precipitation noise suppression processing while preserving the spatial distribution gradient of precipitation, outputting encoded features. This includes: mapping the input sequence into a query matrix, a key matrix, and a value matrix using the self-attention mechanism, and calculating an unscaled attention score matrix as the spatial interaction features based on the query matrix and the key matrix; for each target bit in the unscaled attention score matrix... The target element is placed, and its neighboring elements are extracted within a local filtering window centered on the target element. The median of the neighboring elements is used to replace the value of the target element, resulting in a filtered attention score matrix as the local smoothing feature. The learnable gating parameters corresponding to the self-attention mechanism are obtained, and the unscaled attention score matrix and the filtered attention score matrix are adaptively weighted and fused using the learnable gating parameters to obtain an optimized attention score matrix. The optimized attention score matrix is ​​scaled and normalized to obtain an attention weight matrix. The attention weight matrix is ​​multiplied by the value matrix and dimension-mapped to obtain the encoded feature.

[0012] This scheme reduces the interference of outlier attention response and meteorological noise on spatial interaction modeling by replacing the unscaled attention score matrix with neighborhood median; at the same time, it adaptively fuses the original score and the filtered score through learnable gating parameters, so that the smoothing process does not weaken key local abrupt changes, thus taking into account both noise suppression and precipitation boundary feature preservation.

[0013] In some implementations, the decoder of the precipitation quantitative correction model extracts the bidirectional evolution features of the encoded features in the time dimension through a bidirectional long short-term memory neural network, and performs multi-head cross-attention fusion with the extracted features and the encoded features. The feature reconstruction is achieved through residual operations and dimension mapping to calculate the quantitative correction result. This includes: inputting the encoded features as decoding input features into the bidirectional long short-term memory neural network, calculating the forward and reverse hidden states respectively, and concatenating the forward and reverse hidden states along the feature dimension to obtain bidirectional temporal features representing the bidirectional evolution features; and then... The decoded input features and the bidirectional temporal features are concatenated along the feature dimension. The concatenated features are then subjected to linear mapping and nonlinear activation to obtain local enhanced features. The local enhanced features are then added to the decoded input features using the residual operation of element-wise addition, followed by layer normalization to obtain fused sequence features. The fused sequence features are used as query features, and the encoded features output by the encoder are used as key and value features to perform multi-head cross-attention operations. The output of the multi-head cross-attention operations is then transformed using a feedforward network to obtain decoded output features. The decoded output features are then input into a fully connected layer for dimensional mapping to complete the feature reconstruction and calculate the quantitative correction result.

[0014] This scheme extracts both forward and reverse temporal dependencies simultaneously using a bidirectional long short-term memory neural network, enabling the quantitative correction model to utilize evolutionary information from periods before and after precipitation events. Through feature concatenation, linear mapping, and residual operations, bidirectional temporal features are fused while preserving the decoded input features. Furthermore, a multi-head cross-attention mechanism is used to achieve interactive fusion of fused sequence features and encoded features, which can alleviate information decay in deep sequence modeling and improve the stability of feature reconstruction and the accuracy of precipitation intensity output.

[0015] In some implementations, the steps for establishing the spatial clustering partitioning model include: acquiring historical baseline precipitation data and historical auxiliary variable data, and extracting latitude and longitude, normalized vegetation index, and baseline monthly precipitation statistical characteristics as target clustering samples; calculating the membership degree of each target clustering sample to a preset clustering center, and obtaining the updated target clustering centers; calculating the initial mean, initial covariance, and initial mixture weights based on the membership degree and the target clustering centers to construct an initial Gaussian mixture model; iteratively updating the Gaussian component weights and Gaussian component parameters in the initial Gaussian mixture model based on the expectation-maximization algorithm until convergence, thereby obtaining the spatial clustering partitioning model.

[0016] This scheme uses latitude and longitude, vegetation index, and baseline monthly precipitation statistics as clustering samples, so that the precipitation sub-region division can simultaneously reflect geographical location, underlying surface conditions, and climate precipitation differences. By initializing the Gaussian mixture model with membership degree and iteratively converging, the rationality of the initial clustering parameters and the stability of the regional division can be improved, providing a spatially consistent modeling foundation for subsequent zoning correction.

[0017] In some implementations, calculating the membership degree of each target cluster sample to a preset cluster center includes: performing clustering operations for each candidate partition within a preset range of candidate partition numbers, and calculating multiple clustering evaluation indicators; calculating the objective weights corresponding to each clustering evaluation indicator based on the variability of indicator information and the indicator conflict of each clustering evaluation indicator; using the objective weights to perform a weighted summation of each clustering evaluation indicator to obtain a mixed evaluation index corresponding to each candidate partition number; selecting the candidate partition number corresponding to the maximum value of the mixed evaluation index as the target partition number, and setting the number of preset cluster centers based on the target partition number to calculate the membership degree.

[0018] This scheme evaluates the number of candidate partitions using multiple indicators and determines objective weights based on the variability and conflict of the indicator information, thereby reducing the uncertainty caused by manually setting the number of partitions based on experience. By selecting the target number of partitions through a hybrid evaluation index, a balance can be achieved between the degree of regional subdivision and the quality of clustering, thus improving the adaptability of precipitation subregion division.

[0019] In some implementations, before performing local noise suppression processing on the attention score matrix generated based on the multi-model forecast precipitation data and the auxiliary variable data, the method further includes: calculating the variable importance of each initial feature variable in the multi-model forecast precipitation data and the auxiliary variable data; calculating the maximum mutual information value between each pair of the initial feature variables; when the maximum mutual information value between any two initial feature variables is greater than or equal to a preset mutual information threshold, removing the feature variable with lower variable importance from the two pairs to obtain the retained feature variables; selecting target feature combinations from the retained feature variables in descending order of variable importance, with a cumulative contribution reaching a preset proportion, and performing the local precipitation noise suppression processing based on the target feature combinations.

[0020] This scheme uses a combination of variable importance and maximum mutual information value to screen feature variables, which can retain key meteorological factors that contribute significantly to precipitation correction and remove highly redundant variables with low contributions. By determining the target feature combination through cumulative contribution, the input dimensionality and collinearity interference can be reduced, thereby improving the training efficiency and generalization ability of the subsequent local noise suppression and quantitative correction model.

[0021] In some implementations, the step of classifying and mapping the multi-model forecast precipitation data and the auxiliary variable data to identify the precipitation occurrence state grid by grid includes: acquiring historical baseline precipitation data; marking the historical baseline precipitation data as a rainy state when it is greater than or equal to a preset precipitation threshold, and marking the historical baseline precipitation data as a dry state when it is less than the preset precipitation threshold; pre-training a precipitation state classification model using a training set containing the rainy state label and the dry state label; inputting the multi-model forecast precipitation data and the auxiliary variable data into the precipitation state classification model, and outputting the grid classification probability to identify the rainy state or the dry state grid by grid.

[0022] This scheme constructs rainy and dry state labels using historical baseline precipitation data and trains a precipitation state classification model. This enables the correction process to first identify whether precipitation has occurred before deciding whether to perform quantitative intensity correction. This reduces the zero-value bias caused by dry samples participating in regression and improves the state recognition capability and integrated output reliability in near-zero precipitation scenarios.

[0023] In some implementations, the step of acquiring multi-model forecast precipitation data and auxiliary variable data for the study area to be corrected includes: acquiring ensemble forecast precipitation data containing forecast results from multiple meteorological centers as the multi-model forecast precipitation data; acquiring digital elevation model data, normalized vegetation index data, and meteorological variable data for the study area as the auxiliary variable data, wherein the meteorological variable data includes surface pressure, near-surface temperature, near-surface dew point temperature, and specific humidity, temperature, and wind components at a specified pressure layer.

[0024] This scheme integrates precipitation forecast data from multiple meteorological centers to synthesize the forecast information of different numerical models on precipitation processes. By combining digital elevation models, normalized vegetation index, and multi-layered meteorological variables, the model can utilize topography, underlying surface, water vapor, thermal and dynamic conditions to jointly characterize the precipitation formation environment, thereby improving the completeness of the precipitation correction input information.

[0025] The second aspect of this application relates to a dual-strategy enhanced local perception precipitation correction system, the system comprising: a data acquisition unit for acquiring multi-model forecast precipitation data and auxiliary variable data of the study area to be corrected; a sub-region determination unit for determining the precipitation sub-regions to which each grid belongs within the study area based on a pre-established spatial clustering partitioning model, wherein the spatial clustering partitioning model is generated by extracting features and clustering from historical baseline precipitation data and historical auxiliary variable data; and a state identification unit for classifying and mapping the multi-model forecast precipitation data and the auxiliary variable data, and identifying the precipitation occurrence state grid by grid. The precipitation occurrence states include rainy and rainless states; the quantitative correction unit is used to directly output zero-value results for grids in the rainless state; for grids in the rainy state, according to the precipitation quantitative correction model corresponding to their precipitation sub-region, local noise suppression processing is performed on the attention score matrix generated based on the multi-model forecast precipitation data and the auxiliary variable data, and state transfer and feature reconstruction are performed along the forward and reverse time series to calculate the quantitative correction result; the data output unit is used to perform grid integration based on the zero-value result and the quantitative correction result, reconstruct and output the target forecast precipitation data.

[0026] Thirdly, this application provides an electronic device, comprising: at least one memory for storing a program; and at least one processor for executing the program stored in the memory, wherein when the program stored in the memory is executed, the processor is configured to execute the method described in the first aspect or any possible implementation thereof.

[0027] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when run on a processor, causes the processor to perform the method described in the first aspect or any possible implementation thereof.

[0028] Fifthly, this application provides a computer program product that, when run on a processor, causes the processor to perform the method described in the first aspect or any possible implementation thereof.

[0029] It is understood that the beneficial effects of the second to fifth aspects mentioned above can be found in the relevant descriptions in the first aspect mentioned above, and will not be repeated here.

[0030] Overall, the technical solutions conceived in this application have the following beneficial effects compared with the prior art: (1) By using historical baseline precipitation data and historical auxiliary variable data to establish a spatial clustering partitioning model, the study area is divided into precipitation sub-regions with consistent precipitation characteristics. The corresponding quantitative precipitation correction model is called for different sub-regions, which can reduce the spatial smoothing error caused by the overall regional modeling and improve the ability to express the local precipitation structure and heavy precipitation characteristics.

[0031] (2) By classifying and mapping each grid, rainy and no-rain states are distinguished. Zero values ​​are directly output for no-rain states, and quantitative corrections are performed for rainy states. This enables the model training and inference process to distinguish between precipitation occurrence and precipitation intensity correction tasks, reduce the interference of frequent zero values ​​on regression learning, and improve the reliability of results in no-rain, near-no-rain, and rainy boundary scenarios.

[0032] (3) By suppressing local precipitation noise, adaptively fusing spatial interaction features and local smoothing features, bidirectional temporal state transfer and residual feature reconstruction, the model’s ability to comprehensively model multi-source meteorological variables, local precipitation abrupt changes and temporal evolution relationships is enhanced, thereby improving the correction performance of target forecast precipitation data under continuity index and different precipitation intensity levels. Attached Figure Description

[0033] Figure 1 This is a flowchart illustrating the precipitation correction method with dual-strategy enhanced local perception provided in the embodiments of this application.

[0034] Figure 2 The present invention provides clustering performance evaluation results and optimal precipitation zone feature maps under different numbers of precipitation zones, wherein (a) shows the changes of CH, DB, DI, SC and CI indices under different numbers of zones, (b) shows the monthly average precipitation box plots of different precipitation zones, and (c) shows the spatial distribution map of grid points in each precipitation zone.

[0035] Figure 3 The embodiments of this application provide the optimal feature combination and its feature correlation and importance results determined by the feature selection module at different precipitation zones and the whole watershed scale. Among them, (a)-(d) are the correlation of each feature at the scale of Zone I, Zone II, Zone III and the whole watershed, respectively, and (e)-(h) are the optimal feature combination determined at the scale of Zone I, Zone II, Zone III and the whole watershed, respectively.

[0036] Figure 4 This is a comparison chart of performance indicators of precipitation correction results in different precipitation zones provided in the embodiments of this application. (a) is a comparison chart of continuity indicators R, MAE and RMSE in different precipitation zones, and (b)-(d) are comparison charts of TS, POD and FAR indicators under different precipitation intensity levels.

[0037] Figure 5These are comparison charts of continuity indices for different numerical precipitation forecast products and different quantitative precipitation correction models provided in the embodiments of this application. (a) is a comparison chart of index R, (b) is a comparison chart of index MAE, and (c) is a comparison chart of index RMSE.

[0038] Figure 6 These are performance comparison charts of different numerical precipitation forecast products and different quantitative precipitation correction models provided in the embodiments of this application under different precipitation intensities. (a) is a comparison chart of the TS index, (b) is a comparison chart of the POD index, and (c) is a comparison chart of the FAR index.

[0039] Figure 7 This is a comparison chart of the impact of the zoning strategy, classification strategy and zoning-classification dual strategy on precipitation correction performance provided in the embodiments of this application. Among them, (a) is a comparison chart of the continuity index R, MAE and RMSE corresponding to different strategies, and (b)-(d) are comparison charts of the TS, POD and FAR corresponding to different strategies. Detailed Implementation

[0040] 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.

[0041] In this application, the term "and / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent three cases: A existing alone, A and B existing simultaneously, and B existing alone. In this application, the symbol " / " indicates that the related objects are in an "or" relationship, for example, A / B means A or B.

[0042] In this application, the terms "first" and "second," etc., are used to distinguish different objects, not to describe a specific order of objects. For example, "first response message" and "second response message," etc., are used to distinguish different response messages, not to describe a specific order of response messages.

[0043] In the embodiments of this application, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design that is described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design. Specifically, the use of the terms "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0044] In the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more, for example, multiple processing units means two or more processing units, multiple elements means two or more elements, etc.

[0045] The embodiments of this application are described below with reference to the accompanying drawings. Figure 1 As shown, this embodiment provides a dual-strategy enhanced local perception precipitation correction method. This method automatically generates target forecast precipitation data for the study area to be corrected through spatial clustering partitioning, precipitation state classification, local precipitation noise suppression, and bidirectional temporal feature reconstruction. The core process of this method includes steps S100-S500.

[0046] Step S100: Obtain multi-model forecast precipitation data and auxiliary variable data for the study area to be corrected.

[0047] Specifically, the study area to be corrected can be a watershed, administrative region, climate zone, or any spatial region composed of multiple regular grids. Multi-model forecast precipitation data can include ensemble forecast precipitation products from meteorological centers such as ECMWF (European Centre for Medium-Range Weather Forecasts), JMA (Japan Meteorological Agency), KMA (Korea Meteorological Agency), and NCEP (National Center for Environmental Prediction), as well as precipitation forecast results output from other numerical weather prediction models.

[0048] Ancillary variable data can include digital elevation model (DEM) data, normalized difference vegetation index (NDV) data, and meteorological variable data. Meteorological variable data includes, but is not limited to, surface pressure, near-surface temperature, near-surface dew point temperature, and specific humidity, temperature, and wind components at a specified pressure layer. DEM data can characterize the impact of topographic relief on the spatial distribution of precipitation; the NEV can characterize differences in underlying vegetation cover and surface hydrothermal conditions; and meteorological variable data can characterize the precipitation formation environment from the perspectives of water vapor, thermal, and dynamic conditions. By introducing multi-model forecast precipitation data and multi-source ancillary variable data, this step provides a complete data foundation for subsequent spatial zoning, state identification, and quantitative correction.

[0049] Step S200: Based on the pre-established spatial clustering partitioning model, determine the precipitation sub-regions to which each grid belongs within the study area. The spatial clustering partitioning model is generated by extracting features and clustering using historical baseline precipitation data and historical auxiliary variable data.

[0050] Specifically, a precipitation sub-region refers to several sub-regions divided from the study area based on the spatial heterogeneity of historical precipitation. Precipitation characteristics within each sub-region are relatively consistent, but significant differences exist between different sub-regions. Historical baseline precipitation data can be MSWEP (Multi-Source Weighted-Ensemble Precipitation) precipitation products, surface observation fusion precipitation products, or other highly reliable precipitation baseline data. The spatial clustering model is generated by extracting clustering features such as latitude and longitude, normalized vegetation index, and monthly precipitation statistics from historical baseline precipitation data and historical auxiliary variable data, and then iteratively optimizing them using fuzzy clustering and Gaussian mixture models. In this step, the latitude and longitude, normalized vegetation index, and monthly precipitation statistics of the grid to be corrected are input into the trained spatial clustering model, and the model outputs the precipitation sub-region identifier for each grid.

[0051] Given the significant spatial non-stationarity of diurnal precipitation, applying a uniform model to the entire study area for correction can easily lead to over-smoothing of localized heavy precipitation, orographic precipitation, or marginal precipitation. This embodiment establishes precipitation sub-regions first, allowing the subsequent quantitative precipitation correction model to be trained and inferred within a more spatially consistent sample range, thereby improving its ability to represent local precipitation structure.

[0052] Step S300: Classify and map the multi-model forecast precipitation data and auxiliary variable data, and identify the precipitation occurrence status grid by grid. The precipitation occurrence status includes rainy status and no rainy status.

[0053] Specifically, precipitation occurrence status refers to the binary judgment of rain or no rain made by a precipitation status classification model for each grid cell, used to determine whether the grid cell directly outputs a zero value or enters the quantitative correction process. Classification mapping involves inputting multi-model forecast precipitation data and auxiliary variable data into a pre-trained precipitation status classification model, which outputs a precipitation occurrence status identifier for each grid cell. In this embodiment, the dual strategy refers to a precipitation correction framework combining a zoning strategy and a classification strategy. The zoning strategy divides the area into sub-regions through spatial clustering and establishes correction models for each sub-region. The classification strategy identifies the precipitation occurrence status grid by grid cell to achieve separate processing of rainy and no-rain samples.

[0054] In step S400, for grids in a rainless state, zero-value results are directly output; for grids in a rainy state, according to the precipitation quantitative correction model corresponding to the precipitation sub-region to which it belongs, local noise suppression processing is performed on the attention score matrix generated based on multi-model forecast precipitation data and auxiliary variable data, and state transfer and feature reconstruction are performed along the forward and reverse directions of the time series to calculate the quantitative correction results.

[0055] Specifically, because zero-value samples frequently appear in daily precipitation data, directly inputting both rainless and rainy samples into the regression model can easily lead to the model being dominated by zero-value samples, resulting in insufficient learning of precipitation intensity for rainy samples. Therefore, this embodiment first identifies the precipitation occurrence state grid by grid through classification mapping, directly outputting zero-value results for rainless states, and then performing quantitative correction for rainy states.

[0056] For each precipitation sub-region, a corresponding quantitative precipitation correction model is trained using samples within that region; during inference, the model instance corresponding to the sub-region to which the grid belongs is called.

[0057] For grids with rainy conditions, multi-model forecast precipitation data and auxiliary variable data are input into the quantitative precipitation correction model corresponding to the precipitation sub-region of that grid. This model extracts the spatial interaction features between the multi-model forecast precipitation data and auxiliary variable data through an encoder and performs robust smoothing on local noise. Subsequently, the decoder extracts precipitation evolution features along the forward and reverse temporal sequences, and reconstructs the features through residual operations and dimensional mapping. Thus, the model can simultaneously take into account spatial heterogeneity, precipitation discontinuity, and temporal evolution relationships.

[0058] Step S500: Based on the zero-value results and quantitative correction results, perform grid integration, reconstruct and output the target forecast precipitation data.

[0059] Specifically, for grids identified as being in a rainless state, the zero-value result is used as the correction output for that grid; for grids identified as being in a rainy state, the output of the quantitative precipitation correction model for the corresponding precipitation sub-region is used as the correction output for that grid. Subsequently, all grid results are integrated according to the original grid structure of the study area to reconstruct the target forecast precipitation data.

[0060] This step involves re-stitching the classification and quantitative correction results onto a spatial grid, allowing the target forecast precipitation data to simultaneously retain the zero-value characteristics of the rainless grid and the intensity correction results of the rainy grid. Compared to the unified regression correction method, this approach reduces near-zero precipitation errors and improves the stability of local precipitation intensity corrections.

[0061] Based on the above embodiments, this embodiment further explains the internal structure of the precipitation quantitative correction model in step S400. The precipitation quantitative correction model includes an encoder and a decoder. This model is a partition-classification dual-strategy enhanced local perception Informer (DELPInformer), that is, it introduces a median filtering optimization module and a residual bidirectional long short-term memory neural network module on the basis of the Informer model. The median filtering optimization module is used to suppress local noise in the attention score matrix and retain key mutation information, and the residual bidirectional long short-term memory neural network module is used to enhance the ability to extract temporal features in the decoding stage.

[0062] In step S400, multi-model forecast precipitation data and auxiliary variable data are integrated into an input sequence. The input sequence is a feature matrix formed by arranging multi-model forecast precipitation values ​​and auxiliary variable values ​​at different time steps within the same grid along the time dimension. This matrix is ​​used by the model to capture the spatiotemporal evolution characteristics of the precipitation process. The input sequence is fed into the Informer encoder to learn the long- and short-term dependency features in the sequence. The encoder mainly consists of a multi-head probabilistic sparse self-attention mechanism (PPSA) and a self-attention distillation layer. The PPSA is calculated as follows:

[0063] In the formula, Q , K and V These are the input feature vector matrices, for Q The sparsified probability sparse matrix, Softmax This is the activation function.

[0064] To avoid redundant calculations, after updating the Q matrix using PPSA, the input sequence is distilled, starting from the... j layer to the first j The distillation operation of the +1 layer is expressed as:

[0065] In the formula, MaxPool is the maximum pooling operation, and ELU is the maximum pooling operation. Conv1d is the activation function. This is a one-dimensional convolution operation. This is a self-attention module.

[0066] However, PPSA typically focuses its attention on uneven regions in the probability distribution, treating outliers as the primary factor and thus preserving incorrect features. Furthermore, the inherent noise in meteorological data further interferes with the modeling process. To address this, this application introduces a median filtering optimization module into the Informer encoder to perform locally robust smoothing on the attention score matrix in the multi-head probabilistic sparse self-attention mechanism.

[0067] The encoder of the quantitative precipitation correction model extracts spatial interaction features of the input sequence through a self-attention mechanism, and calculates local smoothing features based on these features. Specifically, the self-attention mechanism maps the input sequence into a query matrix, a key matrix, and a value matrix. Q , K and V The unscaled attention score matrix is ​​calculated based on the query matrix and key matrix as spatial interaction features. For the h-th attention head, the unscaled attention score matrix can be represented as:

[0068] In the formula, For the first h The unscaled attention score matrix of each attention head. and The first h The query matrix and key matrix corresponding to each attention head.

[0069] To reduce the interference of local noise on the attention distribution, the unscaled attention score matrix is... The elements in the matrix are subjected to two-dimensional median filtering. For positions in the matrix... For each element at a given location, its neighborhood elements are extracted within a local window centered on that element. The median of these neighborhood elements is then used to replace the original score value, resulting in the filtered attention score matrix. Its calculation expression is as follows:

[0070] In the formula, for Middle position The median filtering result for the element at position 1. r It is the radius of the filter window ( k= 2 r+ 1). Preferably, k =3 and r =1.

[0071] Furthermore, to avoid excessive smoothing of the attention score due to median filtering, which weakens key local mutation information, this application introduces a learnable gating parameter to adaptively fuse the original attention score matrix and the filtered attention score matrix, obtaining an optimized unscaled attention score matrix, the expression of which is:

[0072] In the formula, For the first h The unscaled attention score matrix after attention head optimization For the first h The learnable gating parameters corresponding to each attention head are constrained to the [0,1] interval by the Sigmoid function to ensure the stability of the fusion process.

[0073] The optimized unscaled attention score matrix is ​​scaled and Softmax normalized to obtain the optimized attention weight matrix. ,in For the first h The feature dimensions of the attention head are then calculated. Finally, the optimized attention weight matrix is ​​multiplied by the value matrix to obtain the first attention head.h The output of each attention head The outputs of all attention heads are concatenated and linearly mapped to obtain the final output of the encoder module, i.e., the encoded features, which are expressed as:

[0074] In the formula, H To focus on the number of heads, This is the output mapping matrix. Through the aforementioned median filtering and learnable gating fusion mechanism, the model can suppress local noise in the attention score while retaining key abrupt changes and local precipitation edge features in the original attention matrix, thereby improving the feature extraction capability and precipitation correction accuracy for complex precipitation sequences.

[0075] While the encoder efficiently handles long sequence dependencies using PPSA, it may have limitations in extracting fine features from short-term dynamic information. Therefore, a ResBiLSTM module is introduced into the decoding layer to address these shortcomings. BiLSTM comprises a forward long short-term memory neural network and a backward long short-term memory neural network, capable of capturing both forward and backward dependencies in the sequence, leading to a more comprehensive understanding of precipitation sequence features.

[0076] The encoded features output by the encoder are used as the decoding input features in a bidirectional long short-term memory neural network. The forward and reverse hidden states are calculated separately, and then concatenated along the feature dimension to obtain the bidirectional temporal features representing the bidirectional evolution. The expression is shown below:

[0077]

[0078]

[0079] In the formula, and These are the hidden states of the forward and reverse LSTM layers, respectively. for t Input of time period, and They are respectively t -1 time and t +1 time-period memory unit's state variable, The hidden state of BiLSTM is represented by a bidirectional temporal feature.

[0080] Forward long short-term memory neural networks can transmit precipitation evolution information from historical moments to the present moment, while inverse long short-term memory neural networks can supplement future context information from subsequent moments to the present moment. The combination of the two can more comprehensively characterize the temporal dependence of precipitation processes, and is especially suitable for the sequence features of diurnal precipitation that are persistent, intermittent, and abrupt.

[0081] Next, the decoded input features and bidirectional temporal features are concatenated along the feature dimension. The concatenated features are then subjected to linear mapping and nonlinear activation to obtain locally enhanced features. These locally enhanced features are then added element-wise with the original code input features using residual operations, followed by layer normalization to obtain fused sequence features. The residual operation preserves the decoded input features while fusing the bidirectional temporal dependency information extracted by the decoder. This design helps alleviate the gradient vanishing or information decay problems during the training of deep sequence models, improving model stability in long-sequence precipitation correction tasks.

[0082] Subsequently, multi-head cross-attention operation is performed using the fused sequence features as query features and the encoded features output by the encoder as key and value features. The output of the multi-head cross-attention operation is then transformed by a feedforward network to obtain the decoded output features.

[0083] Finally, the decoded output features are input into a fully connected layer for dimension mapping, and quantitative correction results are calculated.

[0084] Through the collaborative processing of the encoder and decoder described above, this embodiment can suppress local noise while preserving the spatial distribution gradient of precipitation, and extract precipitation evolution features along the forward and reverse time series, thereby improving the precipitation intensity correction accuracy of the rainy grid.

[0085] Based on the above embodiments, this embodiment further explains the process of establishing the spatial clustering partitioning model in step S200. The establishment of the spatial clustering partitioning model includes steps such as cluster sample extraction, fuzzy clustering initialization, Gaussian mixture model iteration, and determination of the optimal number of partitions.

[0086] First, historical baseline precipitation data and historical auxiliary variable data are obtained, and latitude and longitude, normalized vegetation index, and statistical characteristics of baseline monthly precipitation are extracted as target clustering samples.

[0087] Specifically, historical baseline precipitation data refers to high-precision observational precipitation data products used for training and validation, such as the Multi-Source Weighted Precipitation Estimation Product (MSWEP), with spatiotemporal resolutions of 1 day and 0.1 degrees, respectively. Baseline monthly precipitation statistical characteristics include monthly average precipitation, monthly average number of precipitation days, monthly average precipitation intensity, and daily maximum precipitation. These statistics characterize the precipitation climate features of each grid point from different dimensions. The target clustering sample consists of latitude and longitude, normalized difference vegetation index, and baseline monthly precipitation statistical characteristics, used to characterize the spatial location, underlying surface features, and precipitation climate features of each grid point.

[0088] Secondly, the membership degree of each target cluster sample to the preset cluster centers is calculated, and the updated target cluster centers are obtained. The fuzzy C-means algorithm is used to provide initial parameters for the Gaussian mixture model. The membership degree of each sample point to all cluster centers is calculated by optimizing the objective function. The expression of the objective function is:

[0089]

[0090]

[0091] In the formula, n For the sample size, c The number of cluster centers. For the first i The nth sample pair k Membership degree of a class m As a weighted index, it is generally taken as m =2, = For the first i Sample To the j Cluster centers The Euclidean distance. The fuzzy C-means algorithm allows each sample to belong to multiple clusters to varying degrees through fuzzy membership, which, compared to hard clustering methods, can more reasonably characterize the gradual transition of precipitation features between regions and provide more reliable initial parameter estimates for Gaussian mixture models.

[0092] The initial mean, initial covariance, and initial mixture weights are calculated based on the membership degree and the target cluster center, respectively, to construct the initial Gaussian mixture model, as shown in the following formula:

[0093]

[0094] In the formula, and The first k The mean vector and covariance matrix of each Gaussian component.

[0095] Then, the Gaussian component weights and parameters in the initial Gaussian mixture model are iteratively updated using the expectation-maximization algorithm until convergence, resulting in the spatial clustering partitioning model. The probability density function of the Gaussian mixture model is expressed as:

[0096] In the formula, c This indicates the number of Gaussian components. For the first k The weighting coefficients of each Gaussian component. For the first k The probability density function of Gaussian components. For the parameters to be estimated in the GMM model, the maximum likelihood function is used to estimate them. The solution is:

[0097] In the formula, For the likelihood function, the EM algorithm is usually used to solve the parameter estimation problem. By alternately executing the E-step and M-step, the parameters meet the convergence condition and the final model parameter values ​​are obtained.

[0098] Finally, within the preset number of candidate zones, clustering operations were performed for each candidate zone, and multiple clustering evaluation indices were calculated. Four indices were selected to evaluate the precipitation zoning effect: the variance ratio criterion (Calinski-Harabasz, CH), the Davies-Bouldin index (DB), the Dunn index (DI), and the silhouette coefficient (SC). DB is a negative indicator, while the others are positive. The CRITIC method was used to assign weights to the indices, objectively determining their weights. Construct a hybrid evaluation index with different numbers of partitions. The optimal number of precipitation zones for the model is adaptively determined. The expression is as follows:

[0099]

[0100]

[0101] In the formula, For the first j The weight of each indicator, and The first j The variability and conflict of individual indicators For the first j The amount of information in each indicator Number of partitions i The j Each indicator has a value. For the first j The average of the indicators, For the first j The first indicator and the first k The correlation coefficients of the indicators.

[0102] Figure 2 The clustering performance and optimal precipitation zoning results are presented under different numbers of precipitation zones. The number of zones, k, is set from 2 to 18. As the value of k increases, the Mixture Evaluation Index (CI) exhibits a clear unimodal distribution. When k=3, CI reaches its maximum value of 0.834, an increase of 11.52% compared to the CI value when k=2. When k continues to increase from 3 to 18, CI generally shows a downward trend, indicating that excessive subdivision reduces clustering quality. Therefore, the three-zone scheme can better characterize the spatial heterogeneity of precipitation in the study area, providing a reasonable spatial basis for subsequent zoning modeling.

[0103] Spatially, the three precipitation zones are mainly concentrated in the northwest, central, and southeast of the study area, respectively. Zone I is mainly distributed in the northwest, Zone II in the central region, and Zone III in the southeast. There are significant differences in monthly average precipitation among the different zones, with Zone III having the highest monthly average precipitation at 85.13 mm, followed by Zone I at 81.13 mm, and Zone II having the lowest at 71.89 mm. The relatively even distribution of grid points within each zone indicates that the precipitation zoning method proposed in this application can effectively distinguish regional precipitation differences while maintaining spatial continuity.

[0104] Based on the above embodiments, before performing local precipitation noise suppression processing, it is necessary to screen the feature variables in the multi-model forecast precipitation data and auxiliary variable data to reduce the interference of redundant variables on model training.

[0105] First, the importance of each initial feature variable in the multi-model forecast precipitation data and auxiliary variable data is calculated. Variable importance refers to the contribution of each feature variable to the precipitation correction target, measured by the absolute value of its weight coefficient in the elastic network regression model. The elastic network controls model complexity through L1 and L2 regularization terms and can effectively handle features with multicollinearity. Its loss function expression is:

[0106]

[0107] In the formula, n For the sample size, For the first i The target value for each sample For the first i The feature vector of each sample This is the weight matrix. and These are the L1 and L2 norms of the weight matrix, respectively. As a parameter, when At this time, the elastic network is called Lasso. This is known as ridge regression. Parameters are optimized using 5-fold cross-validation, and the importance I of the feature variables is calculated. EN , and press I EN The values ​​are sorted in descending order.

[0108] Elastic networks combine the feature selection capability of L1 regularization with the stability advantage of L2 regularization, enabling them to effectively identify the feature variables that contribute most to precipitation corrections among meteorological forecast factors with multicollinearity.

[0109] Secondly, calculate the maximum mutual information value between each pair of initial feature variables. When the maximum mutual information value between any two initial feature variables is greater than or equal to a preset mutual information threshold, remove the feature variable with lower importance from the two pairs, obtaining the retained feature variables. The maximum mutual information value is a statistic that measures the non-linear dependency between two feature variables; a larger value indicates a higher degree of information redundancy between the two variables. The formula for calculating the maximum mutual information value is:

[0110]

[0111] In the formula, and They are the characteristic variables. x and y Marginal probability distribution, for x and y The joint probability distribution, a and b They are respectively x and y The number of grids divided in the direction, B This is the upper limit for the number of grid cells. The preset mutual information threshold is preferably set to 0.7. When the maximum mutual information value between two feature variables is greater than or equal to 0.7, the two variables are considered to have highly overlapping information. The variable with higher importance is retained, and the variable with lower importance is removed.

[0112] Considering that relying solely on variable importance ranking cannot identify information redundancy, and relying solely on maximum mutual information for redundancy removal cannot distinguish contribution, this embodiment combines variable importance ranking with maximum mutual information for redundancy removal, achieving the optimal feature combination screening with high contribution and low redundancy.

[0113] Finally, based on the descending order of variable importance, target feature combinations with a cumulative contribution rate reaching a preset proportion are selected from the retained feature variables. Local precipitation noise suppression processing is then performed based on these target feature combinations. The cumulative contribution rate refers to the proportion of the sum of the importance values ​​of the top k feature variables (arranged in descending order of importance) to the total importance value of all retained feature variables. The preset proportion is preferably set to 90%, meaning that feature variables with a cumulative contribution rate reaching 90% are selected as the target feature combinations.

[0114] Figure 3 The optimal feature combinations determined by the feature selection module are presented at different precipitation zones and the whole watershed scale. The results show that the feature selection module selected 3, 5, and 6 feature variables in Zones I, II, and III, respectively, while retaining 4 feature variables at the whole watershed scale. Differences exist in the selected variables across different zones, and near-surface thermal variables, such as 2m temperature and 2m dew point temperature, were excluded in several regions. This may be because their information has already been partially characterized by mid-level atmospheric variables, leading to a reduced marginal contribution to precipitation correction. In contrast, variables such as specific humidity and temperature at the 500hPa and 850hPa pressure layers maintained high stability in most regions, indicating that mid-to-lower-level water vapor and thermal conditions play a crucial role in regional precipitation formation.

[0115] Based on the above embodiments, this embodiment further explains the data acquisition in step S100 and the precipitation status classification in step S300.

[0116] First, ensemble forecast precipitation data, incorporating forecasts from multiple meteorological centers, is acquired as multi-model forecast precipitation data. Ensemble forecast precipitation data refers to a collection of 24-hour precipitation forecast products generated by numerical models independently operated by multiple meteorological centers, including but not limited to forecast products issued by institutions such as the European Centre for Medium-Range Weather Forecasts (ECMWF), the Japan Meteorological Agency (JMA), the Korea Meteorological Administration (KMA), and the U.S. National Center for Environmental Prediction (NCEPA). All products share the same start time of 00:00 UTC and a unified spatial resolution of 0.1 degrees. The technical significance of multi-model ensembles lies in the fact that numerical models from different meteorological centers differ in model structure, parameterization schemes, and initial fields. Ensemble multi-model forecasts can provide richer forecast information, which is beneficial for models to learn the nonlinear mapping relationship between multi-source forecasts and baseline precipitation.

[0117] Digital elevation model (DEM) data, normalized difference in vegetation index (NDV) data, and meteorological variable data for the study area are acquired as auxiliary variables. The spatial resolution of the DEM data is, for example, 90 m, and the spatiotemporal resolution of the NEV data is, for example, 1 month and 1 km. Meteorological variable data includes surface pressure, near-surface temperature, near-surface dew point temperature, and specific humidity, temperature, and wind components at specified pressure layers, such as 500 hPa and 850 hPa. The spatiotemporal resolution of the meteorological variable data is, for example, 1 day and 0.25 degrees. Topographic and vegetation data characterize the influence of underlying surface features on precipitation, while meteorological variable data provide information on atmospheric circulation and water vapor conditions, forming the model input together with forecasted precipitation.

[0118] Data preprocessing was performed on multi-model forecast precipitation data and auxiliary variable data, including time range pruning, outlier and missing value handling, spatial resampling, grid registration, and standardization. Spatially, the multi-model forecast precipitation data and auxiliary variable data were unified to the same resolution grid. Temporally, daily-scale precipitation correction samples and monthly-scale precipitation zoning features were constructed according to the variable applications. The daily-scale precipitation correction samples included base day precipitation, multi-model forecast daily precipitation, and daily meteorological variables. The monthly-scale precipitation zoning features included monthly normalized vegetation index (NZD) and monthly average precipitation, monthly average number of precipitation days, monthly average precipitation intensity, and daily maximum precipitation, all statistically derived from the base day precipitation.

[0119] Then, historical baseline precipitation data is acquired. Data with historical baseline precipitation data greater than or equal to a preset precipitation threshold is labeled as "rainy," and data with historical baseline precipitation data less than the preset precipitation threshold is labeled as "no rain." The preset precipitation threshold is a critical value for precipitation intensity used to distinguish between rainy and no rain conditions, preferably set to 0.1 mm / d, consistent with the criteria for determining trace precipitation in meteorological operations. The rainy and no rain labels are binary classification labels generated based on historical baseline precipitation data and the preset precipitation threshold, used to train the precipitation status classification model.

[0120] A precipitation status classification model is pre-trained using a training set containing labels for rainy and dry states. This model takes multi-model forecast precipitation and auxiliary variables as input, and rainy and dry state labels as output targets. The binary classification model, trained using machine learning algorithms, can automatically determine the precipitation status of each grid cell based on the input data.

[0121] Multi-model forecast precipitation data and auxiliary variable data are input into the precipitation status classification model, and the grid classification probability is output to identify the rainy or rainless status grid by grid.

[0122] Daily precipitation often shows zero values, with rainless days typically accounting for over 70%. Directly using rainless samples in the training of the quantitative correction model would lead to the model being dominated by a large number of zero-value samples, making it difficult to effectively learn the nonlinear characteristics of rainy precipitation. The classification strategy directly outputs zero values ​​for rainless samples, allowing the quantitative correction model to focus solely on rainy samples, reducing training interference, and improving the ability to identify precipitation occurrence states and the accuracy of capturing the transition process between sunny and rainy weather.

[0123] Based on the above embodiments, this embodiment experimentally verifies the precipitation correction method for enhancing local perception with dual strategies provided in this application. Figure 4 The precipitation correction performance under different precipitation zones is demonstrated. The results show that the proposed zoning modeling method exhibits differentiated correction effects in different regions. Zone II shows good overall precipitation intensity correction, with the lowest error level (MAE=1.48mm, RMSE=2.61mm). However, in the heavy precipitation level [20,+∞)mm, Zone II's POD is only 0.02, indicating significant underreporting. This is mainly related to the insufficient number of heavy precipitation samples in this region, which only account for about 0.60% of the test set, making it difficult for the model to fully learn extreme precipitation characteristics. Zone III's continuity index is slightly lower than Zone II's, but it performs better in moderate to heavy precipitation events. For precipitation levels [10,20)mm and [20,+∞)mm, the average TS and POD in Zone III reach 0.26 and 0.41, respectively, indicating that the zoning model has better identification and correction capabilities for heavier precipitation events. The overall correction performance in Region I is relatively weak, especially with a significant increase in error during precipitation events exceeding 10 mm. However, it achieves the highest TS value of 0.31 for precipitation levels in the range of [5, 10) mm, indicating that the model still possesses a certain advantage in identifying this intensity level of precipitation. Where R is the correlation coefficient, MAE is the mean absolute error, RMSE is the root mean square error, TS is the threat score, POD is the probability of detection, and FAR is the false alarm rate.

[0124] Figure 5This chart compares the continuity indices of different numerical weather prediction (NMR) precipitation products and different quantitative precipitation correction models. Among the original NMR products, KMA has the highest correlation coefficient (R) at 0.636, while JMA has the lowest error (MAE = 2.28 mm, RMSE = 4.30 mm), and NCEP performs the worst overall. After correction using deep learning models, the R values ​​of all models improve by at least 8.26%, and the error is significantly reduced, indicating that deep learning correction can significantly improve the accuracy of multi-model NMR precipitation forecasts. Among the different correction models, the performance improvement of Transformer and Informer is relatively limited. Compared to RInformer using only the ResBiLSTM module, MInformer using the median filter optimization module performs slightly better in terms of R and RMSE, while RInformer performs better in terms of MAE (MAE = 1.64 mm). This result suggests that the median filter module is more effective in smoothing local extreme noise, while the ResBiLSTM module is more helpful in correcting the overall average bias. After fusing the median filter optimization module and the ResBiLSTM module, the MAE and RMSE of ELPInformer are reduced by at least 0.04 mm compared to the single improved module, indicating that module fusion can further improve the precipitation correction performance. Building on this, the DELPInformer proposed in this application further introduces a dual strategy of partitioning and classification, achieving the best precipitation correction effect. Compared to ELPInformer without the dual strategy, DELPInformer's R-value is increased by 1.12%, and MAE and RMSE are reduced by 1.22% and 1.01%, respectively.

[0125] Figure 6This study demonstrates the detection performance of different numerical weather prediction products and correction models under varying precipitation intensity levels. Overall, the models perform well under light rain conditions, but their performance declines significantly under moderate rain and above. While the original numerical weather prediction products maintain high POD and TS during heavy precipitation events, their FAR generally exceeds 0.80, indicating a significant tendency for false alarms during heavy precipitation. After correction using deep learning models, the average FAR of all models decreased by 19.04%, demonstrating that the correction models effectively mitigate the false alarm problem of the original numerical weather prediction products. Compared to Transformer and Informer models, the improved Informer model shows certain advantages under different precipitation levels, especially in the [0.1, 5) mm precipitation range, where the TS value generally exceeds 0.50. However, under heavy precipitation conditions of [20, +∞) mm, the average TS and POD of the single-structure improved model are only 0.10 and 0.13, respectively, indicating that single-module improvements are still insufficient to effectively address the problem of missed extreme precipitation events. The proposed DELPInformer performs more consistently across different precipitation levels, showing a significant advantage, particularly in periods of no rainfall and heavy rainfall. For no rainfall events ([0, 0.1) mm), DELPInformer achieves the highest TS and POD, at 0.64 and 0.80 respectively, indicating that the classification strategy effectively improves the ability to identify precipitation occurrence states. For heavy rainfall events ([20, +∞) mm), DELPInformer reduces FAR by 0.12 compared to ELPInformer, while increasing TS from 0.11 to 0.16 and POD from 0.13 to 0.20, representing improvements of 52.15% and 49.77% respectively. These results demonstrate that the zoning strategy can mitigate the spatial smoothing effect of uniform basin-wide modeling by independently modeling spatially homogeneous regions, thereby improving the model's ability to capture local heavy rainfall characteristics.

[0126] Figure 7Further comparisons were made of the correction results of Z-ELPInformer (using only a partitioning strategy), C-ELPInformer (using only a classification strategy), and DELPInformer (using both partitioning and classification strategies). The results show that both partitioning and classification strategies can further improve the correction performance based on ELPInformer, but their focuses differ. The classification strategy mainly improves the identification of rainless or near-rainless events. Compared with ELPInformer, C-ELPInformer reduced MAE by 0.05 mm, a reduction of 3.37%; in the [0, 0.1) mm level, TS and POD improved by 17.77% and 41.97%, respectively. This indicates that the classification strategy can effectively reduce errors under near-zero precipitation conditions. Compared with ELPInformer, Z-ELPInformer improved the average POD by 20.64% in precipitation events greater than or equal to 10 mm, indicating that partitioning modeling can enhance the model's ability to express local precipitation structures and alleviate the spatial smoothing problem caused by uniform modeling across the entire watershed. However, in the [0, 0.1) mm level, Z-ELPInformer's TS and POD are lower than C-ELPInformer's, indicating that using only the partitioning strategy has limited improvement in identifying rainless conditions. DELPInformer, by integrating both partitioning and classification strategies, achieves more balanced performance across different precipitation levels. Compared to the single-strategy model, DELPInformer improves TS and POD by at least 15.49% and 14.34%, respectively, and reduces the average FAR by 8.45%. This result validates the synergistic advantages of the partitioning-classification dual strategy, and the combination of the two can further improve the accuracy of precipitation forecast corrections across the entire watershed.

[0127] In summary, the DELPInformer precipitation correction method proposed in this application effectively improves precipitation forecast performance and exhibits the best results. This is mainly because the method fully considers the spatial heterogeneity of precipitation and effectively identifies precipitation states, making it more conducive to capturing local precipitation details.

[0128] Compared with existing technologies, this application has the following advantages: It proposes a dual-strategy precipitation correction framework based on regionalization and classification. This framework characterizes the spatial heterogeneity of precipitation through hybrid precipitation clustering and addresses the issues of discontinuous daily precipitation and frequent zero values ​​through a rainy / no-rain classification strategy. This reduces spatial smoothing errors caused by unified modeling across the entire watershed and improves the ability to identify precipitation occurrence states and the accuracy of local precipitation correction. Furthermore, by constructing an enhanced local perception DELPInformer model, which integrates feature selection, median filtering optimization, and ResBiLSTM modules on top of the Informer model, it reduces the interference of redundant variables and local noise on model training. This enhances the model's ability to extract key meteorological factors, precipitation abrupt changes, and local features, thereby improving the accuracy of predicted precipitation intensity correction and providing reliable support for refined meteorological forecasting and hydrological applications.

[0129] Based on the above embodiments, a precipitation correction system with dual-strategy enhanced local sensing is provided, the system comprising: The data acquisition unit is used to acquire multi-model forecast precipitation data and auxiliary variable data for the study area to be corrected. The sub-region determination unit is used to determine the precipitation sub-region to which each grid belongs within the study area based on a pre-established spatial clustering partitioning model. The spatial clustering partitioning model is generated by extracting features and clustering using historical baseline precipitation data and historical auxiliary variable data. The state identification unit is used to classify and map the multi-model forecast precipitation data and the auxiliary variable data, and identify the precipitation occurrence state grid by grid. The precipitation occurrence state includes rainy state and no rainy state. The quantitative correction unit is used to directly output zero-value results for grids in the rainless state; for grids in the rainy state, according to the precipitation quantitative correction model corresponding to the precipitation sub-region to which it belongs, it performs local noise suppression processing on the attention score matrix generated based on the multi-model forecast precipitation data and the auxiliary variable data, and performs state transfer and feature reconstruction along the forward and reverse directions of the time series to calculate the quantitative correction result. The data output unit is used to perform grid integration based on the zero-value results and the quantitative correction results, reconstruct and output the target forecast precipitation data.

[0130] It is understood that the detailed functional implementation of each of the above units can be found in the description of the aforementioned method embodiments, and will not be repeated here.

[0131] Based on the methods in the above embodiments, this application provides an electronic device that may include a processor, a communications interface, a memory, and a communication bus, wherein the processor, communications interface, and memory communicate with each other via the communication bus. The processor may invoke logical instructions stored in the memory to execute the methods in the above embodiments.

[0132] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application.

[0133] Based on the methods in the above embodiments, this application provides a computer-readable storage medium storing a computer program that, when run on a processor, causes the processor to execute the methods in the above embodiments.

[0134] Based on the methods in the above embodiments, this application provides a computer program product that, when run on a processor, causes the processor to execute the methods in the above embodiments.

[0135] It is understood that the processor in the embodiments of this application can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. A general-purpose processor can be a microprocessor or any conventional processor.

[0136] The method steps in this application embodiment can be implemented in hardware or by a processor executing software instructions. The software instructions can consist of corresponding software modules, which can be stored in random access memory (RAM), flash memory, read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), registers, hard disks, portable hard disks, CD-ROMs, or any other form of storage medium known in the art. An exemplary storage medium is coupled to the processor, enabling the processor to read information from and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and the storage medium can reside in an ASIC.

[0137] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted through the computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state disk (SSD)).

[0138] It is understood that the various numerical designations used in the embodiments of this application are merely for the convenience of description and are not intended to limit the scope of the embodiments of this application.

[0139] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A precipitation correction method with dual strategies to enhance local perception, characterized in that, include: Acquire multi-model forecast precipitation data and auxiliary variable data for the study area to be corrected; Based on a pre-established spatial clustering model, the precipitation sub-regions to which each grid belongs within the study area are determined. The spatial clustering model is generated by extracting features and clustering data using historical baseline precipitation data and historical auxiliary variable data. The multi-model forecast precipitation data and the auxiliary variable data are classified and mapped, and the precipitation occurrence status is identified grid by grid. The precipitation occurrence status includes rainy status and no rainy status. For grids in the rainless state, zero-value results are directly output; for grids in the rainy state, according to the precipitation quantitative correction model corresponding to the precipitation sub-region to which it belongs, local noise suppression processing is performed on the attention score matrix generated based on the multi-model forecast precipitation data and the auxiliary variable data, and state transfer and feature reconstruction are performed along the forward and reverse directions of the time series to calculate the quantitative correction result. Based on the zero-value results and the quantitative correction results, grid integration is performed to reconstruct and output the target forecast precipitation data.

2. The method according to claim 1, characterized in that, The quantitative correction model based on the precipitation sub-region to which it belongs performs local noise suppression processing on the attention score matrix generated based on the multi-model forecast precipitation data and the auxiliary variable data, and performs state transfer and feature reconstruction along the forward and reverse time series to calculate the quantitative correction result, including: The multi-model forecast precipitation data and the auxiliary variable data are integrated into an input sequence; The encoder of the precipitation quantitative correction model extracts the spatial interaction features of the input sequence through a self-attention mechanism, and calculates local smoothing features based on the spatial interaction features. The spatial interaction features and the local smoothing features are then adaptively weighted and fused to perform the local precipitation noise suppression processing while preserving the spatial distribution gradient of precipitation, and output the encoded features. Using the decoder of the precipitation quantitative correction model, the bidirectional evolution features of the encoded features in the time dimension are extracted through a bidirectional long short-term memory neural network. The extracted features are then fused with the encoded features through multi-head cross-attention. The features are reconstructed through residual operation and dimension mapping, and the quantitative correction result is calculated.

3. The method according to claim 2, characterized in that, The encoder utilizing the precipitation quantitative correction model extracts spatial interaction features from the input sequence through a self-attention mechanism, calculates local smoothing features based on these features, and adaptively weights and fuses the spatial interaction features and the local smoothing features to perform local precipitation noise suppression processing while preserving the spatial distribution gradient of precipitation, outputting encoded features, including: The input sequence is mapped into a query matrix, a key matrix, and a value matrix through the self-attention mechanism, and an unscaled attention score matrix is ​​calculated based on the query matrix and the key matrix as the spatial interaction feature. For each target location element in the unscaled attention score matrix, neighborhood elements are extracted within a local filtering window centered on the target location element, and the value of the target location element is replaced by the median of the neighborhood elements to obtain the filtered attention score matrix as the local smoothing feature. Obtain the learnable gating parameters corresponding to the self-attention mechanism, and use the learnable gating parameters to perform the adaptive weighted fusion of the unscaled attention score matrix and the filtered attention score matrix to obtain the optimized attention score matrix; The optimized attention score matrix is ​​scaled and normalized to obtain the attention weight matrix. The attention weight matrix is ​​then multiplied by the value matrix and dimension-mapped to obtain the encoded features.

4. The method according to claim 2, characterized in that, The decoder of the precipitation quantitative correction model extracts the bidirectional evolution features of the encoded features in the time dimension through a bidirectional long short-term memory neural network, and performs multi-head cross-attention fusion between the extracted features and the encoded features. The feature reconstruction is achieved through residual operations and dimension mapping, and the quantitative correction result is calculated, including: The encoded features are input into the bidirectional long short-term memory neural network as decoding input features. The forward hidden state and the reverse hidden state are calculated respectively. The forward hidden state and the reverse hidden state are concatenated along the feature dimension to obtain the bidirectional temporal features that characterize the bidirectional evolution features. The decoded input features and the bidirectional temporal features are concatenated along the feature dimension. The concatenated features are then subjected to linear mapping and nonlinear activation in sequence to obtain local enhanced features. The residual operation, which involves adding corresponding elements of the local enhanced features and the decoded input features, is then subjected to layer normalization to obtain the fused sequence features. Using the fused sequence features as query features, and the encoded features output by the encoder as key and value features, a multi-head cross-attention operation is performed, and the output of the multi-head cross-attention operation is transformed by a feedforward network to obtain the decoded output features. The decoded output features are input into a fully connected layer for dimension mapping to complete the feature reconstruction and calculate the quantitative correction result.

5. The method according to claim 1, characterized in that, The steps for establishing the spatial clustering partitioning model include: Historical baseline precipitation data and historical auxiliary variable data are obtained, and latitude and longitude, normalized vegetation index and baseline monthly precipitation statistical characteristics are extracted from them as target clustering samples. Calculate the membership degree of each target cluster sample to the preset cluster center, and obtain the updated target cluster center; The initial mean, initial covariance, and initial mixture weights are calculated based on the membership degree and the target cluster center, respectively, to construct the initial Gaussian mixture model; The Gaussian component weights and parameters in the initial Gaussian mixture model are iteratively updated using the expectation-maximization algorithm until convergence, thus obtaining the spatial clustering partitioning model.

6. The method according to claim 5, characterized in that, The calculation of the membership degree of each target cluster sample to the preset cluster center includes: Within a preset range of candidate partition numbers, clustering operations are performed for each candidate partition number, and multiple clustering evaluation metrics are calculated. Based on the variability of information content and the conflict of each clustering evaluation index, calculate the objective weight corresponding to each clustering evaluation index. The objective weights are used to perform a weighted summation of each clustering evaluation index to obtain a mixed evaluation index corresponding to the number of each candidate partition. The number of candidate partitions corresponding to the maximum value of the mixed evaluation index is selected as the target number of partitions, and the number of preset cluster centers is set based on the target number of partitions to calculate the membership degree.

7. The method according to claim 1, characterized in that, Before performing local noise suppression processing on the attention score matrix generated based on the multi-model forecast precipitation data and the auxiliary variable data, the method further includes: Calculate the importance of each initial characteristic variable in the multi-model forecast precipitation data and the auxiliary variable data; Calculate the maximum mutual information value between each pair of the initial feature variables. When the maximum mutual information value between any two initial feature variables is greater than or equal to a preset mutual information threshold, remove the feature variable with lower importance between the two to obtain the retained feature variables. According to the order of the importance of the variables from largest to smallest, target feature combinations with a cumulative contribution of a preset ratio are selected from the retained feature variables, and the local precipitation noise suppression processing is performed based on the target feature combinations.

8. The method according to any one of claims 1 to 7, characterized in that, The process of classifying and mapping the multi-model forecast precipitation data and the auxiliary variable data, and identifying the precipitation occurrence status grid by grid, includes: Historical baseline precipitation data is obtained. When the historical baseline precipitation data is greater than or equal to a preset precipitation threshold, it is marked as a rainy state label. When the historical baseline precipitation data is less than the preset precipitation threshold, it is marked as a rainless state label. A precipitation state classification model is pre-trained using a training set containing the rainy state label and the no-rain state label. The multi-model forecast precipitation data and the auxiliary variable data are input into the precipitation state classification model, and the grid classification probability is output to identify the rainy state or the rainless state grid by grid.

9. The method according to any one of claims 1 to 7, characterized in that, The acquisition of multi-model forecast precipitation data and auxiliary variable data for the study area to be corrected includes: Acquire ensemble forecast precipitation data containing forecasts from multiple meteorological centers as the multi-model forecast precipitation data; The digital elevation model data, normalized vegetation index data, and meteorological variable data of the study area are obtained as auxiliary variable data. The meteorological variable data includes surface air pressure, near-surface temperature, near-surface dew point temperature, and specific humidity, temperature, and wind components at a specified pressure layer.

10. A precipitation correction system with dual strategies to enhance local perception, characterized in that, include: The data acquisition unit is used to acquire multi-model forecast precipitation data and auxiliary variable data for the study area to be corrected. The sub-region determination unit is used to determine the precipitation sub-region to which each grid belongs within the study area based on a pre-established spatial clustering partitioning model. The spatial clustering partitioning model is generated by extracting features and clustering using historical baseline precipitation data and historical auxiliary variable data. The state identification unit is used to classify and map the multi-model forecast precipitation data and the auxiliary variable data, and identify the precipitation occurrence state grid by grid. The precipitation occurrence state includes rainy state and no rainy state. The quantitative correction unit is used to directly output zero-value results for grids in the rainless state; for grids in the rainy state, according to the precipitation quantitative correction model corresponding to the precipitation sub-region to which it belongs, it performs local noise suppression processing on the attention score matrix generated based on the multi-model forecast precipitation data and the auxiliary variable data, and performs state transfer and feature reconstruction along the forward and reverse directions of the time series to calculate the quantitative correction result. The data output unit is used to perform grid integration based on the zero-value results and the quantitative correction results, reconstruct and output the target forecast precipitation data.

Images