Space perception based on deep learning and multi-intensity gate control residual precipitation forecast correction method and system

Abstract

The application discloses a spatial perception and multi-intensity gating residual precipitation forecast correction method and system based on deep learning, comprising: uniformly preprocessing grid precipitation forecast data, reanalysis precipitation true value data and original terrain data; performing targeted conversion, adaptive coding and fusion on multi-modal features affecting precipitation to construct a multi-modal prediction factor library; constructing a spatial perception and multi-intensity gating residual precipitation forecast correction model, training and verifying the model through multi-loss function collaborative optimization, and obtaining a trained model; optimizing model parameters, and performing inverse transformation and physical constraint on model output to obtain a precipitation forecast correction result; and based on the precipitation forecast correction result, taking global reanalysis precipitation true value data as a reference, evaluating the correction effect and verifying the effectiveness of the model. The application significantly improves the accuracy, spatiotemporal consistency and business applicability of precipitation forecast, and has important value for fine precipitation forecast.

Description

Technical Field

[0001] This invention belongs to the field of meteorological forecasting technology, specifically relating to a method and system for correcting spatial perception and multi-intensity gated residual precipitation forecasts based on deep learning. Background Technology

[0002] With the continuous development of forecasting technology, the overall capability of precipitation forecasting has been significantly optimized. As one of the core tasks of meteorological operations, forecast accuracy directly affects the scientific nature of decision-making in key areas such as flood control and drought relief, disaster emergency response, agricultural production, and transportation. With global warming, the frequency, intensity, impact range, and suddenness of extreme precipitation events have significantly increased, placing higher demands on the precision of precipitation forecasts, such as spatial resolution, accuracy of intensity classification, ability to capture extreme precipitation events, and spatiotemporal consistency. Among these, 1-7 day lead time precipitation forecasts, as a "critical window" for disaster prevention and mitigation, require both spatial precision for 1-3 day short-term forecasts and improved trend reliability for 4-7 day medium-term forecasts. This is the core period with the most urgent needs in operational applications. This invention uses 1-7 day precipitation forecasts as a typical application scenario for model training and validation, while also flexibly adapting to correction tasks with longer forecast lead times, possessing broad engineering expansion and application potential.

[0003] Currently, precipitation forecasting mainly relies on numerical weather prediction models. These models simulate atmospheric physical processes based on atmospheric dynamic equations, providing fundamental support for large-scale precipitation trend forecasting. However, they still face insurmountable technical bottlenecks in practical applications. First, the parameterization schemes of physical processes in numerical models suffer from simplification errors, which accumulate gradually with the forecast lead time, leading to insufficient adaptability to complex underlying surfaces. In areas with complex topography, they are prone to problems such as biased precipitation intensity estimation, with the cumulative effect of errors in medium- and long-term forecasts being particularly significant. Second, traditional forecast correction methods are mostly based on linear statistical models, such as bias correction methods and moving average methods. These methods struggle to characterize the nonlinear relationships between precipitation forecasts and various influencing factors such as topography, spatiotemporal characteristics, precipitation intensity levels, and historical precipitation patterns. They lack the ability to differentiate and adapt to error characteristics at different forecast lead times and have limited correction effects for low-probability events such as extreme precipitation. Furthermore, existing deep learning-based correction models often focus on single-dimensional optimization, failing to fully consider the differences in error patterns among precipitation intensities and lacking refined integration of topographic factors. Topography, as a key underlying surface factor influencing precipitation formation and distribution, directly alters the spatiotemporal distribution patterns of precipitation due to its type differences. However, existing methods mostly reflect geographical information indirectly through latitude and longitude. Different topographic types have a significant modulating effect on the errors of multi-time- and multi-level precipitation forecasts, especially in medium- and long-term forecast periods. Summary of the Invention

[0004] Purpose of the invention: The purpose of this invention is to provide a deep learning-based spatial perception and multi-intensity gated residual precipitation forecast correction method that integrates multi-dimensional influencing factors and adapts to different precipitation intensity characteristics and error patterns in different time periods.

[0005] Technical solution: The method described in this invention includes the following steps:

[0006] Unified preprocessing is performed on gridded precipitation forecast data, reanalysis precipitation true data, and raw topographic data;

[0007] Based on the preprocessed data, targeted transformation, adaptive encoding and fusion are performed on the multimodal characteristics affecting precipitation to generate multimodal forecast factor vectors, and a multimodal forecast factor library is further constructed.

[0008] A spatially-aware and multi-intensity-gated residual precipitation forecast correction model was constructed. The model was trained and validated through multi-loss function collaborative optimization to obtain a well-trained model. The model includes a shared feature extraction network, a spatial awareness module, a precipitation existence prediction head, a residual prediction head, and a power-gated unit. The shared feature extraction network takes a multimodal forecast factor library as input and outputs a deep fusion feature vector. The spatial awareness module takes preprocessed latitude and longitude features as input and outputs an enhanced spatial representation feature vector. The output of the spatial awareness module is element-wise added to the output of the shared feature extraction network to enhance the spatial heterogeneity capture capability of the deep fusion features. The precipitation existence prediction head takes a deep fusion feature vector after fusing spatially encoded features as input and outputs the precipitation existence probability. The residual prediction head takes a deep fusion feature vector after fusing spatially encoded features as input and outputs a logarithmic spatial residual correction. The power-gated unit takes the precipitation existence probability and the logarithmic spatial residual correction as input and outputs a gated weighted logarithmic spatial residual correction.

[0009] The model parameters are optimized through iterative training, and a regularization strategy is used to prevent overfitting. At the same time, the logarithmic space residual correction amount after gated weighting of the model output is subjected to inverse transformation and physical constraints to obtain the precipitation forecast correction results.

[0010] Based on the precipitation forecast correction results, and with reference to the true global reanalysis precipitation data, the model correction effect is evaluated from three dimensions: core indicators, inter-regional performance, and stability, thus verifying the model's effectiveness.

[0011] Further preprocessing includes:

[0012] Based on the preset latitude and longitude range slices, extract the grid data of the corresponding specifications to make the spatial grids of the three types of data completely aligned;

[0013] By accurately matching the timestamps of the forecast initialization time and the forecast lead time, a one-to-one correspondence between grid precipitation forecast data and reanalysis precipitation true data is established.

[0014] Identify and remove invalid values ​​from the data, including missing values ​​and outlier values ​​that are outside the physically reasonable range;

[0015] Based on the pre-defined terrain classification results, the terrain classification labels corresponding to each grid point are read from the terrain grid data.

[0016] Furthermore, targeted transformation, adaptive encoding, and fusion are performed on the multimodal features affecting precipitation, including:

[0017] The targeted transformation and adaptive coding of core forecast features include: the core forecast features are gridded precipitation forecast data; a logarithmic transformation is performed on the preprocessed gridded precipitation forecast data to obtain continuous features after targeted transformation; Z-Score normalization coding is performed on the continuous core forecast features after targeted transformation to obtain core forecast modal coding features.

[0018] The targeted transformation and adaptive encoding of spatiotemporal features include: forecast lead time extraction: directly extracting the original numerical values ​​of the matched forecast lead time; longitude and latitude extraction of grid points: directly extracting the original longitude and latitude coordinates of the target area; targeted transformation of the flood season month at the forecast initialization time: classifying the flood season month at the forecast initialization time to clarify the flood season month's affiliation; Z-Score normalization encoding is performed on the forecast lead time and longitude / latitude as continuous features; the flood season month is a discrete feature, and its numerical form is directly preserved; thus, the spatiotemporal modal coding features are obtained.

[0019] The targeted transformation and adaptive encoding of terrain features include: using basic terrain labels: directly adopting the terrain category labels that have been divided in the data preprocessing stage; constructing terrain-derived interaction features: based on the first few categories of terrain category labels, constructing terrain-latitude interaction features by performing element-by-element operations with the original latitude data of the grid, and obtaining terrain-derived interaction features; performing one-hot encoding on discrete terrain category labels, and not encoding or standardizing terrain-latitude interaction features; concatenating the one-hot encoded terrain features with the terrain-latitude interaction features to obtain terrain modal coding features.

[0020] The targeted transformation and adaptive coding of statistical features include: the targeted transformation of statistical features into the historical no-precipitation frequency features of each grid point; and then performing Z-Score normalization coding on the historical no-precipitation frequencies to obtain statistical mode coding features.

[0021] Multimodal feature fusion includes: sequentially concatenating the obtained core forecast modal coding features, spatiotemporal modal coding features, topographic modal coding features, and statistical modal coding features according to their dimensions to form a multimodal forecast factor vector, and generating a multimodal forecast factor library based on the multimodal forecast factor vector.

[0022] Furthermore, the shared feature extraction network includes two fully connected layers, a batch normalization layer, a LeakyReLU activation function layer, and a Dropout regularization layer connected in series; the output is a multi-dimensional deep fusion feature vector.

[0023] The spatial perception module maps the standardized latitude and longitude features into a multi-dimensional spatial feature vector through a linear projection layer. After multiplying the multi-dimensional spatial feature vector with the adaptively learned feature fusion weights, it performs element-wise addition and fusion with the deep feature vector output by the shared feature extraction network to enhance spatial perception capabilities.

[0024] The precipitation existence prediction head consists of two fully connected layers and an activation function layer. The outputs of the two fully connected layers are connected to the input of the activation function layer, and the activation function layer outputs the probability of precipitation existence.

[0025] The residual prediction head consists of a single fully connected layer and outputs log-space residual corrections.

[0026] The power-gated unit is used to perform power-weighted calculations. The inputs are the probability of precipitation and the logarithmic space residual correction, and the output is the logarithmic space residual correction after gating and weighting.

[0027] Furthermore, the power-gated unit employs a power-gated strategy to dynamically control the impact of residual correction on the final forecast result, adjusting the residual weights based on the precipitation existence probability, as shown in the formula:

[0028] ;

[0029] in, For the first Logarithmic space residual correction for grid points after power-gated weighting For the first Precipitation at each grid point has a probability. For power-order parameters, For the first Logarithmic space residual correction for each grid point.

[0030] Furthermore, multi-loss function collaborative optimization includes:

[0031] Joint Enhanced Spatial Zero-Value Penalty Loss Logarithmic Huber loss and quantile loss Construct the total loss function The expression is:

[0032] ;

[0033] in, To enhance the spatial zero-value penalty loss weight, For the logarithmic Huber loss weights, Weights for quantile loss. To assign a fixed weight to the penalty for false precipitation, Set a fixed weight for the deviation penalty term. Set a fixed weight for the residual penalty term. For false precipitation penalties, This is a deviation penalty item. This is a residual penalty term.

[0034] Furthermore, an inverse transformation and physical constraints are applied to the log-space residual correction amount after gated weighting of the model output, including:

[0035] Spatial filtering optimization: Based on spatial connectivity verification, during the model training phase, false precipitation regions with low precipitation probability below the first preset continuous grid threshold are removed, as are false precipitation regions with precipitation probability below the first preset threshold. During the verification and inference phases, false precipitation regions with low precipitation probability below the second preset continuous grid threshold are removed, as are false precipitation regions with precipitation probability below the second preset threshold. Regions with an original forecast of 0 use the same probability threshold as the corresponding phase.

[0036] Inverse logarithmic space transformation: This involves correcting the logarithmic space residuals after gated weighting. The inverse transformation formula for converting the corrected result to the actual spatial distribution of precipitation is:

[0037] ;

[0038] ;

[0039] in, For the model after correction The precipitation value of each grid point represents the intensity of the final corrected precipitation forecast for each grid point; For the first The log-space original forecast value of the grid point is the _th _th_ grid point. The result of logarithmic transformation of the original precipitation forecast values ​​for each grid point; For the first Logarithmic space residual correction for grid points after power-gated weighting; For the preprocessed first Precipitation forecast data for each grid point; To obtain An exponential function with base 0. For taking the logarithmic function;

[0040] Physical scope clipping: After model correction, the first Precipitation values ​​at each grid point By constraining the data within a preset range and eliminating outliers that exceed the physical range, the precipitation forecast correction results are obtained.

[0041] Furthermore, the model correction effect was evaluated from three dimensions: core metrics, interval performance, and stability, including:

[0042] The overall error index and precipitation event scoring index are calculated to quantify the improvement of the correction results. The overall error index is the deviation, used to characterize the degree of systematic deviation between the original precipitation forecast value or the precipitation forecast correction result and the true precipitation data of the reanalysis. The precipitation event scoring index includes threat score and hit rate.

[0043] The performance of the interval classification includes: Interval classification assessment: Classifying the intervals according to precipitation standards, focusing on the situation of heavy rain, rainstorm, and torrential rain, calculating the deviation, threat score and hit rate of each interval, and analyzing the correction effect of heavy rain, rainstorm, and torrential rain; Comparative analysis: Statistics on the deviation improvement, threat score improvement and hit rate improvement of each level interval, calculating the proportion of intervals with improvement greater than the set threshold, and quantifying the overall effectiveness of the correction plan;

[0044] Stability analysis: including time-leading stability verification: based on forecast data from multiple time periods before and after the forecast, analyze the correction effect, threat score and hit rate differences at different time periods to ensure that the model performs consistently across the entire time range;

[0045] Spatial adaptability analysis: Combined with terrain features, the model's adaptability to different terrain regions is verified; a comparison map of precipitation spatial distribution is generated simultaneously to assist in the analysis of threat scores and hit rates in different terrain regions, ensuring the model's adaptability to spatial heterogeneity.

[0046] The system corresponding to the method includes:

[0047] The data preprocessing unit is used to perform unified preprocessing on gridded precipitation forecast data, reanalysis precipitation true data, and raw topographic data.

[0048] The multimodal forecast factor library construction unit is used to perform targeted transformation, adaptive encoding and fusion of multimodal features affecting precipitation based on preprocessed data of various types, generate multimodal forecast factor vectors, and further construct a multimodal forecast factor library;

[0049] The model building and training unit is used to construct a spatially aware and multi-intensity gated residual precipitation forecast correction model. Through multi-loss function collaborative optimization, the model is trained and validated to obtain a trained model. The model includes a shared feature extraction network, a spatial awareness module, a precipitation existence prediction head, a residual prediction head, and a power-gated unit. The shared feature extraction network takes a multimodal forecast factor library as input and outputs a deep fusion feature vector. The spatial awareness module takes preprocessed latitude and longitude features as input and outputs an enhanced spatial representation feature vector. The output of the spatial awareness module is element-wise added and fused with the output of the shared feature extraction network to enhance the spatial heterogeneity capture capability of the deep fusion features. The precipitation existence prediction head takes a deep fusion feature vector after fusing spatially encoded features as input and outputs the precipitation existence probability. The residual prediction head takes a deep fusion feature vector after fusing spatially encoded features as input and outputs a logarithmic spatial residual correction. The power-gated unit takes the precipitation existence probability and the logarithmic spatial residual correction as input and outputs a gated weighted logarithmic spatial residual correction.

[0050] The model optimization unit is used to optimize model parameters through iterative training and combine regularization strategies to prevent overfitting. At the same time, it performs inverse transformation and physical constraints on the gated weighted log space residual correction of the model output to obtain precipitation forecast correction results.

[0051] The evaluation unit is used to assess the model correction effect based on the precipitation forecast correction results, with reference to the true precipitation data from reanalysis, from three dimensions: core indicators, inter-regional performance, and stability, and to verify the effectiveness of the model.

[0052] The method corresponds to a computer program product, which includes a computer program / instruction that, when executed by a processor, implements the method.

[0053] Beneficial effects: Compared with the prior art, the present invention has significant technical effects: (1) The present invention is based on a deep learning scheme of spatial perception and multi-intensity gated residuals. By fusing precipitation forecast data, 9 types of terrain features (deep sea, shallow sea, coastal plain, plain, hill, low mountain, medium mountain, high mountain, and extremely high mountain), latitude and longitude spatial features, and historical no precipitation frequency and other multimodal information, it can effectively capture the nonlinear correlation between precipitation and terrain, spatiotemporal location, and intensity level. Compared with traditional single data source or linear correction methods, the data coverage is more comprehensive and the terrain adaptability is stronger, which significantly improves the forecast accuracy of complex underlying surface areas; (2) It proposes to fuse enhanced spatial zero-value penalty loss and log Huber loss. The composite loss function of loss and quantile loss balances the learning priority of different intensities of precipitation through dynamic weight strategy. It not only uses spatial consistency constraints to suppress false precipitation, but also improves the correction accuracy of regular precipitation through logarithmic space robust loss, and strengthens the capture ability of extreme precipitation with the help of quantile loss. It solves the contradiction that traditional methods tend to underestimate extreme precipitation when suppressing false precipitation, and greatly improves the reliability of precipitation forecast. (3) The model introduces a multi-intensity gating mechanism and a spatial perception module. Through the power gating unit, the residual correction weight is dynamically adjusted for different intensities of precipitation such as zero value, trace amount, and rainstorm, so that the model focuses on key errors. The spatial perception module strengthens the integration of geographic information through latitude and longitude coding, and improves the ability to characterize the spatial consistency and spatial heterogeneity of the forecast. It effectively improves the model's learning ability and generalization performance of the spatiotemporal distribution characteristics of precipitation. The model has been verified by 1-7 day lead time data to adapt to the cumulative characteristics of forecast error. The model parameters can be adjusted according to the error law of different forecast lead time and the application scenarios can be expanded. It has strong business application value.

[0054] With significant improvements in core meteorological operational indicators such as deviation, precipitation score, and hit rate, the overall technical effect of this invention demonstrates outstanding practicality and progressiveness. Attached Figure Description

[0055] Figure 1 This is a flowchart of the precipitation forecast correction method of the present invention;

[0056] Figure 2 A schematic diagram illustrating targeted transformation and adaptive encoding of multimodal features;

[0057] Figure 3 A schematic diagram of the spatial perception and multi-intensity gated residual model structure;

[0058] Figure 4 A comparison chart of bias deviation (BIAS) for moderate to heavy rain;

[0059] Figure 5 A comparison chart of TS scores for heavy rain and above;

[0060] Figure 6This is a comparison chart of the hit rate (POD) for heavy rain and above. Detailed Implementation

[0061] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0062] This invention constructs a deep learning-based spatial perception and multi-intensity gated residual precipitation forecast correction method, focusing on the Chinese region. It preprocesses ECMWF (European Centre for Medium-Range Weather Forecasts) precipitation forecast data, ERA5 (European Centre for Medium-Range Weather Forecasts 5th Generation Atmospheric Reanalysis) data, and raw topographic data, then fuses multimodal features. Through joint training and learning of latitude and longitude spatial features, precipitation intensity classification features, nine types of topographic features, and historical no-precipitation frequency features, it can effectively capture the nonlinear correlation between precipitation and spatial location, topographic type, and intensity level. A precipitation forecast correction model integrating a spatial perception module and a multi-intensity gated residual mechanism is employed, and a composite loss function including enhanced spatial zero-value penalty loss, logarithmic Huber loss, and quantile loss is designed. Dynamic adjustment correction strategies are implemented for precipitation of different intensities, enhancing the accurate capture of extreme precipitation and effectively suppressing false precipitation, ultimately resulting in precipitation correction results that combine spatiotemporal consistency and intensity accuracy.

[0063] like Figure 1 As shown, the method described in this invention has a progressive and data-driven relationship between its steps: Step 1 is data preprocessing, providing high-quality raw data for subsequent steps; Step 2, based on the preprocessed data from Step 1, first performs targeted transformation on each modal feature, constructing derived and statistical features with clear physical meaning, then standardizes continuous features individually and performs differential one-hot encoding on discrete features before adaptive fusion, generating a 20-dimensional feature vector that the model can directly input, which is the core bridge connecting the raw data and model training; Step 3, based on the 20-dimensional feature vector from Step 2, constructs and trains... The spatial perception and multi-intensity gated residual precipitation forecast correction model is the core step in achieving precipitation correction. Step 4 optimizes model parameters through iterative training, combines regularization strategies to prevent overfitting, and performs post-processing and physical constraints on the model output results, such as power-gated weighting and numerical pruning, to generate accurate correction results that conform to meteorological laws and improve forecast reliability. Step 5 conducts multi-dimensional effect evaluation based on the final correction results of Step 4 to verify the model's effectiveness. Each step is interconnected, forming a complete technical chain of "data input → feature processing → model training → result optimization → effect verification". The specific steps are as follows:

[0064] Step 1: Data acquisition and preprocessing;

[0065] 1.1 Data Acquisition;

[0066] We will acquire three types of core data, as follows:

[0067] Grid precipitation forecast data: This data is acquired through ECMWF. During the training phase, forecast data within a pre-set forecast lead time is used. In this embodiment, the forecast lead time is selected as 24h~168h (1~7 days). Grid precipitation forecast data characterizes the precipitation intensity at grid points.

[0068] Further analysis of the true precipitation data: This data is taken from ERA5 and serves as the real-time reference data for precipitation forecasting. This data represents the actual precipitation intensity at the grid points;

[0069] Raw terrain data: The data comes from GEBCO and is in grid format. In this example, the target area is China. The preprocessing part classifies and encodes the terrain types, including: deep sea, shallow sea, coastal plain, plain, hills, low mountains, medium mountains, high mountains, and extremely high mountains. The terrain classification criteria can be flexibly adjusted according to different application scenarios, such as regional fine-grained forecasting and nationwide large-scale forecasting.

[0070] 1.2 Data preprocessing;

[0071] The above three types of data undergo unified preprocessing to ensure data quality and format consistency. The specific steps are as follows:

[0072] Target area extraction: Based on the preset latitude and longitude range slices, extract grid data of corresponding specifications to ensure that the spatial grids of the three types of data are completely aligned; in this embodiment, 71×125 grid data is extracted, that is, 71 grid points in the latitude direction and 125 grid points in the longitude direction.

[0073] Data association and matching: By accurately matching the forecast initialization time with the forecast lead time timestamp, a one-to-one correspondence is established between grid precipitation forecast data and reanalysis precipitation true value data; for example, the 72-hour lead time grid precipitation forecast data initialized at 00:00 on June 1, 2022 corresponds to the reanalysis precipitation true value data on June 3, 2022.

[0074] Invalid value removal: Identify and remove invalid values ​​from the data, including missing values ​​and abnormal extreme values ​​that exceed the physical reasonable range, such as negative precipitation values ​​or extreme anomalies with precipitation values ​​greater than 500 mm.

[0075] Topographic classification: Based on the pre-classified topographic data of China's regions into 9 categories (classification criteria refer to geographical features: deep sea (water depth > 2000m), shallow sea (water depth 0~2000m); coastal plain (elevation < 100m, distance from coastline 0~50km), plain (elevation ≤ 200m), hills (elevation 200~500m), low mountains (elevation 500~1000m), medium mountains (elevation 1000~3500m), high mountains (elevation 3500~5000m), and extremely high mountains (elevation > 5000m)), the topographic classification labels corresponding to each grid point are directly read from the extracted topographic grid data of the target area to obtain the topographic classification data.

[0076] Step 2: Adaptive encoding of multimodal features and construction of a multimodal prediction factor library;

[0077] This step, based on the clean raw data preprocessed in Step 1, first performs targeted transformation of multimodal features, then executes differential adaptive encoding and feature fusion to generate a 20-dimensional multimodal forecast factor vector. This is further used to construct a multimodal forecast factor library, providing input data for model training in Step 3. The dataset is divided as follows: data from 2020, 2021, and 2022 are used as the training set; data from 2023 is used as the validation set; and data from 2024 is used as the test set. The spatial grid specifications, data format, and feature dimensions of the three datasets are completely identical, differing only in time range, ensuring data matching for model training, tuning, and evaluation. Figure 2 As shown, the specific steps are as follows:

[0078] The four types of multimodal features are specifically defined as follows: the core forecast feature is gridded precipitation forecast data based on ECMWF, and in this embodiment, the gridded precipitation forecast data for China with a forecast lead time of 24h to 168h is extracted; the spatiotemporal feature is the feature extracted by fusing the time and spatial dimensions of the gridded precipitation forecast data, where the time dimension includes the forecast lead time and the flood season month at the forecast initialization time, and the spatial dimension is the longitude and latitude of the grid points; the topographic feature is the topographic type feature obtained based on the pre-divided topographic classification results, which includes nine topographic types: deep sea, shallow sea, coastal plain, plain, hills, low mountains, medium mountains, high mountains, and extremely high mountains; and the statistical feature is the historical no-precipitation frequency feature of the grid points calculated based on the reanalysis precipitation true value data.

[0079] Before adaptive coding, targeted transformations are performed on the above four types of multimodal features to achieve physical morphology adaptation: logarithmic transformation is performed on the core forecast features to adapt to the physical characteristics of non-uniform precipitation distribution; the continuous original values ​​and the flood season month categories at the initial time of the discrete forecast are extracted from the spatiotemporal features respectively; the terrain features reuse the already divided terrain category labels and construct terrain-latitude interactive features; and the statistical features complete the numerical calculation of the historical no-precipitation frequency of grid points.

[0080] Based on this, a unified adaptive labeling and coding principle is applied to the four types of features that have undergone targeted transformation: For discrete features such as terrain type and the flood season month at the forecast initialization time, one-hot coding is used to complete feature coding; for continuous features such as core forecast data after logarithmic transformation, forecast lead time, longitude and latitude of grid points, and historical no-precipitation frequency, standardized coding is used to complete feature coding to eliminate the influence of dimensional differences and numerical skewness distribution; subsequently, dimensional mapping and feature fusion are performed on each type of feature after adaptive labeling and coding to generate corresponding single-modal feature vectors, which are then concatenated into a unified multimodal forecast factor vector.

[0081] 2.1 Targeted transformation and adaptive encoding of multimodal features;

[0082] For four types of multimodal features—core forecast features, spatiotemporal features, topographic features, and statistical features—targeted transformation is first performed, followed by adaptive coding. The purpose of targeted transformation is to adapt the physical form of the features based on the physical meaning of precipitation and the original attributes of the data. The purpose of adaptive coding is to standardize the transformed features into a format that the model can directly input. This represents feature data after targeted transformation of single-class multimodal features.

[0083] The detailed steps for targeted conversion and adaptive encoding are as follows:

[0084] Targeted conversions corresponding to core forecast features:

[0085] A logarithmic transformation is performed on the gridded precipitation forecast data after the basic preprocessing in Step 1 to enhance its sensitivity to small precipitation values, thereby obtaining the continuous features after targeted transformation. .

[0086] Logarithmic transformation formula: ;in, Characterizing the first The result of logarithmic transformation of precipitation forecast data for each grid point; This indicates the first step after basic preprocessing in step 1. Precipitation forecast data for each grid point.

[0087] Targeted transformations corresponding to spatiotemporal features:

[0088] The process comprises three parts: forecast lead time extraction, latitude and longitude extraction, and flood season month classification. Specifically: Forecast lead time extraction: directly extracting the original forecast lead time values ​​after matching in step 1; Latitude and longitude extraction: directly extracting the original latitude and longitude coordinates of the target area; Flood season month classification: classifying the flood season months (May, June, and July) at the forecast initialization time to clarify the flood season month affiliation; after the above processing is completed, the spatiotemporal features after targeted transformation are obtained. .

[0089] Targeted conversion based on terrain features:

[0090] The process comprises two parts: the reuse of basic terrain labels and the construction of terrain-derived interaction features. The basic terrain labels are reused directly from the nine terrain category labels already defined during the data preprocessing stage. The construction of terrain-derived interaction features involves selecting the first three categories from the terrain category labels and performing element-by-element operations with the original latitude data to construct terrain-latitude interaction features, generating 3D terrain-derived interaction features. Integrating the nine-dimensional terrain category labels and the 3D terrain-derived interaction features yields a complete 12-dimensional terrain feature set. This results in the terrain features after targeted transformation. .

[0091] Targeted transformations corresponding to statistical characteristics:

[0092] Based on reanalysis of true precipitation data, the historical frequency of no precipitation at each grid point is calculated, and the statistical features are numerically constructed; thus, continuous features after targeted transformation are obtained. .

[0093] Using the four types of features after targeted transformation as input, an adaptive encoding operation is performed. The adaptive encoding performs differentiated processing according to the continuous and discrete forms of features. The core includes two types of operations: Z-Score normalization of continuous features and differential one-hot encoding of discrete features, as detailed below:

[0094] Z-Score standardization is performed on continuous features: Z-Score standardization is performed on all continuous features after targeted transformation to eliminate dimensional differences; including:

[0095] Z-Score standardization was performed on the targeted transformed grid precipitation forecast data to obtain the core forecast mode coding features;

[0096] Z-Score standardization was applied to the original forecast lead time and original latitude and longitude.

[0097] Z-Score normalization was performed on historical no-precipitation frequencies to obtain statistical mode coding features.

[0098] Discrete feature processing: The original discrete features are first encoded using one-hot encoding. Specifically, the terrain category labels after targeted transformation are encoded using one-hot encoding to obtain terrain one-hot features. The terrain-latitude interaction features are not encoded or standardized. The flood season months are encoded using one-hot encoding to obtain flood season month one-hot features. Neither the terrain one-hot features nor the flood season month one-hot features are standardized; they are directly retained as 0 / 1 values. The terrain one-hot features are concatenated with the terrain-latitude interaction features to finally form the terrain modal coding features.

[0099] The spatiotemporal modal coding features are composed of the original forecast lead time, original latitude and longitude, and unique thermal characteristics of the flood season months after Z-Score standardization;

[0100] Z-Score standardization is used to eliminate dimensional differences in continuous features. The calculation method is as follows: ;in, This refers to the feature data after targeted transformation of a single type of continuous feature, and broadly refers to the various types of continuous features mentioned above. ; This represents the average value of this type of feature on the training set; This represents the standard deviation of this type of feature on the training set; This represents the data after Z-Score standardization for this type of feature.

[0101] The aforementioned continuous feature standardization and discrete feature differential one-hot encoding together constitute a complete adaptive encoding process. The core logic of the overall feature processing is as follows: After targeted transformation, the multimodal features are scaled and formatted according to the aforementioned classification and categorical encoding rules to obtain standardized features, which are then input into the model to complete subsequent deep nonlinear mapping and association mining. This adaptive encoding process can be represented by a unified mathematical expression:

[0102] ;

[0103] in, This represents the adaptive coding mapping function, which uniformly covers all the operational logic of Z-Score standardization, discrete one-hot coding, and direct retention of terrain-derived features mentioned above, and provides a standardized mapping relationship for achieving unified multimodal feature scale. Representation and Input The relevant adaptive coding parameters are determined by Obtained through statistical analysis of its own numerical distribution; This represents the feature data after targeted transformation of single-class multimodal features, which serves as the input basis for adaptive encoding, as mentioned above. , respectively represent: ,Right now , representing the core forecast feature, is a one-dimensional feature obtained by targeted logarithmic transformation of the original grid precipitation forecast data; 2, that is , representing spatiotemporal characteristics, is a 6-dimensional combination feature obtained after forecast lead time, targeted extraction of latitude and longitude, and classification of flood season months; 3, that is , representing terrain features, is a 12-dimensional combined feature obtained by specifically constructing terrain label reuse and terrain-latitude interaction features; 4, that is , representing statistical characteristics, is a one-dimensional feature obtained by targeted statistical calculation of the historical frequency of no precipitation in the grid.

[0104] It is a set of feature vectors encoded by the feature mapping function for four modes. This represents the core forecast modality coding features. Represents spatiotemporal modal coding features. Represents terrain modal coding features. This represents statistical mode coding features; the feature mapping function is the adaptive coding function, and the goal of the feature mapping function is to adjust the parameters statistically. , so that the encoded Adapt to the model data distribution and complete the final conversion from the original transformed features to the model-recognizable coded features.

[0105] 2.2 Multimodal feature fusion;

[0106] After obtaining the multimodal coding features (i.e., the collective term for the aforementioned core forecast modality coding features, spatiotemporal modality coding features, topographic modality coding features, and statistical modality coding features), the four types of coding features are concatenated sequentially according to their dimensions and fused into a unified 20-dimensional feature vector. This vector is defined as the multimodal forecast factor vector. The splicing method is as follows:

[0107] ;

[0108] in, This represents the multimodal forecast factor vector; The core forecast modality coding feature is a 1-dimensional feature; It is a spatiotemporal modal coding feature, which is a 6-dimensional feature, including 1-dimensional time, 2-dimensional latitude and longitude, and 3-dimensional month; The terrain modality encoding features are 12-dimensional features, including 9-dimensional terrain and 3-dimensional terrain interaction; The features are 1-dimensional statistical modal coding features; after concatenation, the feature vector dimensions satisfy 1+6+12+1=20 dimensions, and the multimodal forecast factor vectors of all samples together constitute the multimodal forecast factor library.

[0109] 2.3 Division of the multimodal forecasting factor library;

[0110] Multimodal forecast factor library The dataset is divided into training, validation, and test sets according to the year (training set corresponds to 2020-2022, validation set corresponds to 2023, and test set corresponds to 2024) to ensure that the data in the model training, tuning, and evaluation stages are independent and reasonably matched.

[0111] Step 3: Construct a spatial perception and multi-intensity gated residual precipitation forecast correction model, and use a multimodal forecast factor library to train and validate the model to obtain a trained model;

[0112] This step, based on the multimodal forecasting factor library constructed in step 2, designs a spatially perceptive and multi-intensity gated residual precipitation forecast correction model. Through collaborative optimization of multiple loss functions, it achieves accurate correction of the original precipitation forecast. The direct input data of the model is the multimodal forecasting factor library output in step 2. The data is directly output as the probability of precipitation. Correction amount for the original logarithmic space residual , Corresponding to the The two direct outputs are weighted by the power-gated unit built into the model to obtain the gated weighted logarithmic space residual correction. This step only completes the static definition of the model structure and core strategies, such as power-law gating and multiple loss function rules. The model parameters are randomly initialized, and no actual training operation is performed. Figure 3 As shown, the specific steps are as follows:

[0113] 3.1 Overall structure and module composition of the model;

[0114] The spatial perception and multi-intensity gated residual precipitation forecasting correction model consists of five core modules: a shared feature extraction network, a spatial perception module, a precipitation existence prediction head, a residual prediction head, and a power-gated unit. The input-output connections between these modules are as follows:

[0115] The input to the shared feature extraction network is the multimodal forecasting factor library constructed in step 2. The output is a deep fusion feature vector; the spatial perception module takes Z-score-normalized latitude and longitude features as input and outputs an enhanced spatial representation feature vector. The output of the spatial perception module is element-wise added and fused with the output of the shared feature extraction network to enhance the spatial heterogeneity capture capability of the deep fusion features; the precipitation existence prediction head takes the deep fusion feature vector after fusing spatial coding features as input and outputs the precipitation existence probability. The input to the residual prediction head is the deep fused feature vector after fusing spatial encoded features, and the output is the log-space residual correction. The input to the power-order gating unit is the probability of precipitation. With log space residual correction The output is the corrected logarithmic space residual after gated weighting. .

[0116] The specific implementation of each module is as follows:

[0117] The shared feature extraction network consists of two fully connected layers, a batch normalization layer, a LeakyReLU activation function layer, and a Dropout regularization layer connected in series. The first fully connected layer has an input dimension of 20 and an output dimension of 24, meaning it has 24 hidden units (neurons). It is then connected to a batch normalization layer with a momentum of 0.9, a LeakyReLU activation function layer with a negative slope of 0.1, and a Dropout regularization layer. The random deactivation probability of the Dropout regularization layer is set to 0.7 to reduce the risk of model overfitting. The second fully connected layer has an input dimension of 24 and an output dimension of 12. It is then connected to a batch normalization layer and a LeakyReLU activation function layer. No Dropout regularization layer is set, and the final output is a 12-dimensional deep fused feature vector.

[0118] Spatial perception module: The standardized latitude and longitude features from step 2 are mapped into a 12-dimensional spatial feature vector through a linear projection layer; after multiplying the vector by a feature fusion weight of 0.1, it is fused element-wise with the 12-dimensional deep fusion feature vector output by the shared feature extraction network to enhance spatial perception capabilities; this spatial perception module is embedded in the feature extraction process and has a feature fusion relationship with the shared feature extraction network.

[0119] Precipitation Prediction Header: Constructed from two fully connected layers and multiple activation function layers; the first fully connected layer has an input dimension of 12 and an output dimension of 6, followed by a LeakyReLU activation function layer; the second fully connected layer has an input dimension of 6 and an output dimension of 1, followed by a Sigmoid activation function layer, outputting the precipitation probability with a value ranging from 0 to 1. ;

[0120] Residual prediction head: a single fully connected layer with 12 input dimensions and 1 output dimension, no activation function, outputting log-space residual correction. First, it is clipped to the initial range of [-8.0, 8.0], then after power-gated weighting, it is clipped a second time to the final range of [-4.0, 4.0].

[0121] Power-gated unit: No additional network layers, only performs power-weighted calculations, the input is the probability of precipitation. Correction to the original logarithmic space residual The output is the corrected logarithmic space residual after gated weighting. .

[0122] To ensure the numerical stability of the model training, the input multimodal forecast factor library is... The precipitation probability values ​​output by the model are clipped to the range of [-5,5] to avoid interference from extreme values ​​during model training.

[0123] 3.2 Implementation of Power-Order Gating Strategy

[0124] The power-gating strategy is the core algorithm strategy for implementing the power-gating unit function in step 3.1. That is, it uses power-weighted mathematical logic to determine the probability of precipitation. Correction to the original logarithmic space residual Dynamic weighting is performed, and the final output is the logarithmic space residual correction after gated weighting. This is the core implementation method for the power-gated unit to achieve dynamic adaptation of "precipitation existence - residual correction". The power-gated strategy dynamically controls the impact of the residual correction on the final forecast result. The core is to adjust the residual weights through the precipitation existence probability, as shown in the formula:

[0125] ;

[0126] in, For the first Logarithmic space residual correction for grid points after power-gated weighting For the first The probability of precipitation occurring at each grid point is represented by a value between 0 and 1, indicating the magnitude of the probability of precipitation occurring at that grid point. For power-order parameters, The empirically optimal value is kept constant during both the model training and inference phases, without dynamic adjustments. For the first The logarithmic space residual correction for each grid point is the unweighted residual correction directly output by the model residual prediction head.

[0127] 3.3 Collaborative optimization using multiple loss functions;

[0128] A total loss function is constructed by combining three core loss functions, unifying and optimizing model parameters to balance the capabilities of precipitation existence assessment, precipitation intensity prediction, and extreme precipitation correction. Total Loss Function The expression is:

[0129] ;

[0130] The specific implementations of each loss function are as follows:

[0131] Enhanced spatial zero-value penalty loss The expression is:

[0132] ;

[0133] Enhanced spatial zero-value penalty loss For the precipitation existence determination task, and considering the Focal loss handling of class imbalance, a false precipitation penalty item is added. and spatial smoothing loss Optimize the accuracy of precipitation existence assessment; among which Weighting for penalties related to false precipitation. For spatial smoothing constraint weights, =0.6, =0.01 are all fixed empirical optimal values; This represents a dynamic zero-value weight. A dynamic zero-value weight adjustment strategy is introduced, with the initial value of the dynamic zero-value weight set to 1.3. As the model trains, the weight linearly decays to 1.0, thus strengthening the model's priority in fitting samples without precipitation during the initial training phase. Indicates Focal loss. For the first The probability of precipitation at each grid point is denoted by a value between 0 and 1, representing the magnitude of the probability of precipitation at each grid point. Indicates the first The existence and true labeling of precipitation at each grid point; Generated based on reanalysis of true precipitation data: when the actual precipitation value at a grid point is >0.1 mm, =1 indicates that there is precipitation; when the actual precipitation value is ≤0.1mm, =0 indicates no precipitation; As a monitoring signal, it can optimize the probability prediction accuracy of precipitation;

[0134] Logarithmic Huber loss For precipitation intensity prediction tasks, the residual correction is optimized in logarithmic space to address small errors ( ) uses squared loss to address large errors ( Linear loss is used to reduce extreme value interference, where As the loss threshold, in this embodiment, we take... The expression is:

[0135] ;

[0136] in, Indicates the first Logarithmic space residual correction for each grid point.

[0137] Quantile loss Designed for extreme value prediction tasks across the entire precipitation range, this involves correcting the logarithmic space residual of the i-th grid point. The asymmetric loss calculation enhances the model's ability to fit heavy and extreme precipitation events; its physical meaning is the loss value characterizing the correction bias for extreme precipitation, expressed as:

[0138]

[0139] in, For quantile loss parameters, This indicates that positive biases, i.e., underestimating extreme precipitation, are given a higher penalty weight, which effectively alleviates the problem of underestimation in extreme precipitation forecasts. This indicates that the expected value of the loss term within the parentheses is taken, which is the average of the loss values ​​of all training samples. It is used to quantify the average extreme precipitation correction bias of the model on the entire sample. For the model after correction Precipitation values ​​for each grid point For the first Reanalyze the true precipitation data at each grid point.

[0140] Fixed weight for false precipitation penalty: =0.08;

[0141] Deviation penalty term with fixed weight: =0.04;

[0142] Fixed weight for residual penalty term: =0.002;

[0143] The weights mentioned above are all empirically optimal values ​​to ensure that the addition of penalty terms does not disrupt the balance of the core loss function.

[0144] The false precipitation penalty term is a sub-term of the enhanced spatial zero-value penalty loss. It is specifically used to quantify the severity of false precipitation events such as "no precipitation in the original forecast but precipitation in the model prediction." By applying a targeted penalty to misjudged precipitation in areas with no precipitation, it enhances the accuracy of the model's judgment on the existence of precipitation. The specific formula is as follows:

[0145] ;

[0146] in, The total number of samples in the training batch, dimensionless; For the first The original forecast zero-value mask for each sample is set to 1 if the original forecast value is ≤0.1 mm, and 0 otherwise. It is used to locate areas with no precipitation. For the first The probability of precipitation existing in each sample, with a value ranging from 0 to 1; For the model after correction Precipitation values ​​for each grid point; The SoftMask activation function is used to strengthen the penalty for false precipitation above the threshold. This loss term is related to the enhanced spatial zero-value penalty loss. This creates a "dual constraint": the former focuses on suppressing spurious precipitation during model training, while the latter emphasizes sample fitting optimization. Their combined effect effectively reduces the generation of spurious precipitation in areas without precipitation grids, improving the spatial plausibility of forecast results. This penalty term only applies during model training.

[0147] This is a dimensionless bias penalty term, serving as a supplementary constraint to the logarithmic Huber loss. It quantifies the systematic bias of precipitation values ​​after model correction relative to the true reanalysis precipitation values. By penalizing this global bias trend, it prevents the model from deviating from its overall direction due to excessive focus on local errors. The specific formula is as follows:

[0148] ;

[0149] Used to constrain systematic biases in correction results; The model after correction Precipitation values ​​for each grid point; Indicates the first Reanalysis of true precipitation data for each grid point.

[0150] In practical applications, when the model underestimates the intensity of heavy rainfall or overestimates the intensity of light rainfall, the bias penalty term will drive the adjustment of model parameters, making the mean of the corrected result closer to the true value, thereby improving the reliability of long-term forecasts. The code logic involves calculating the mean difference between the corrected precipitation value and the true value, quantifying the bias using squared loss or linear loss, and then applying a fixed weight. =0.04 is included in the total loss. The larger the absolute value of the deviation, the stronger the penalty, and the weight... =0.04 is the empirically optimal value, ensuring that the statistical unbiasedness of the correction results is effectively optimized without disrupting the balance of the core loss function.

[0151] This is a residual penalty term, a supplementary constraint to the logarithmic Huber loss. It is used to constrain the numerical fluctuation range of the logarithmic space residual correction, preventing excessive residual anomalies from causing the correction result to exceed the reasonable range of meteorological and physical conditions, such as in the event of extreme anomalies in heavy precipitation or negative precipitation. The specific formula is as follows:

[0152] ;

[0153] The loss term is related to the "in the model" The cropping to [-4.0, 4.0] forms a double safeguard, jointly ensuring the stability and physical rationality of the residual correction, and avoiding extreme anomalies in residuals that could lead to correction results that deviate from actual meteorological patterns (such as extreme anomalies where single-grid precipitation values ​​exceed 500mm). In the code implementation, the residual correction amount... The loss is calculated using L2 regularization with fixed weights. =0.002 is included in the total loss to suppress excessive fluctuations in the residuals. This weight value has been verified by a large number of experiments. It can effectively constrain the range of residuals without affecting the optimization priority of the core loss function for judging the existence and intensity of precipitation. It only plays a role in the training phase to improve the stability of the model.

[0154] Dynamic loss weight adjustment: Weights of augmented space zero-value penalty loss during training. Initially 0.35 and gradually decreasing, logarithmic Huber loss weights Initially 0.8 and gradually increased to 1.0, quantile loss weights. The total loss is obtained by weighted summation of the three types of losses, with the value fixed at 0.02.

[0155] Step 4: Model training optimization and prediction result post-processing;

[0156] This step is the engineering implementation of the model design in step 3. The core is to optimize the model parameters built in step 3 through iterative training, and to combine this with a regularization strategy to prevent overfitting. Simultaneously, post-processing is performed on the model output to improve its practical application performance. The specific steps are as follows:

[0157] 4.1 Model training optimization strategies;

[0158] Optimizer selection: SGD (Stochastic Gradient Descent) optimizer is adopted, with an initial learning rate of 0.0003, momentum of 0.9, weight decay coefficient of 0.005, corresponding to L2 regularization;

[0159] Learning rate decay: The ReduceLROnPlateau strategy is adopted. When the validation set loss stagnates continuously, the learning rate is reduced by a decay factor of 0.8, with the minimum learning rate being 5e-7.

[0160] Weighted sampling: Referring to the rainfall levels in the national standard, WeightedRandomSampler is used to assign differentiated weights to rainfall samples of different intensities, with a focus on heavy rainfall. Light rain (0.1-9.9mm), moderate rain (10-24.9mm), heavy rain (25-49.9mm), and torrential rain (50-99.9mm) all have a weight of 1.1. Rainfall of 100mm or more has a weight of 2.0, rainfall of 200mm or more has a weight of 1.5, and extreme rainfall of 300mm or more has a weight of 1.2. The basic weights of various rainfall intensities need to be further multiplied element-wise with the original sample weights and the rainfall mask to obtain the final sample training weights, thereby achieving accurate sample imbalance correction.

[0161] Regularization techniques: Implement gradient pruning and data augmentation specific to the training phase to enhance the model's generalization ability.

[0162] 4.2 Early stopping strategy and optimal model saving;

[0163] The model training employs a multi-dimensional metric to construct an early stopping detection mechanism, with core evaluation metrics including validation set loss. The system considers precipitation forecast errors and systematic deviations between corrected precipitation forecasts and true global reanalysis precipitation values. If the above core indicators fail to reach the preset improvement range after 15 consecutive iterations, an early stop strategy is triggered, terminating model training and saving the model parameters with the best performance on the validation set. The optimal model is determined by the lowest total loss on the validation set and the best forecast accuracy for key precipitation areas.

[0164] 4.3 Post-processing of prediction results;

[0165] Correction amount of log-space residuals after gated weighting of model output Perform inverse transformation and physical constraints to obtain the final corrected precipitation forecast value. The specific steps are as follows:

[0166] Spatial filtering optimization: Based on spatial connectivity verification, during the model training phase, false precipitation regions with low precipitation probability below the first preset continuous grid threshold are removed, as are false precipitation regions with precipitation probability below the first preset threshold. During the verification and inference phases, false precipitation regions with low precipitation probability below the second preset continuous grid threshold are removed, as are false precipitation regions with precipitation probability below the second preset threshold. Regions with an original forecast of 0 use the same probability threshold as the corresponding phase, without any additional increase.

[0167] In this embodiment, based on spatial connectivity verification, during the model training phase, false precipitation regions with low precipitation probability (less than 3 consecutive grids) are removed; during the verification and inference phase, false precipitation regions with low precipitation probability (less than 6 consecutive grids) are removed. During the training phase, false precipitation regions with low precipitation probability are also removed. Areas with false precipitation less than 0.2 were eliminated during the verification and inference stages. Areas with false precipitation of less than 0.25.

[0168] Inverse logarithmic space transformation: This involves correcting the logarithmic space residuals after gated weighting. The inverse transformation formula for converting the corrected result to the actual spatial distribution of precipitation is:

[0169] ;

[0170] ;

[0171] in, For the model after correction The precipitation value of each grid point represents the intensity of the final corrected precipitation forecast for each grid point; For the first Logarithmic space residual correction for grid points after power-gated weighting; For the first The result of logarithmic transformation of precipitation forecast data for each grid point; The first step after basic preprocessing in step 1 Precipitation forecast data for each grid point; To obtain An exponential function with base 0. This is a logarithmic function.

[0172] Physical rationality trimming: After model correction, the first Precipitation values ​​at each grid point By constraining the precipitation forecast within a reasonable range of 0-500mm and eliminating outliers that exceed the physical range, a precipitation forecast correction result that can be used for operational applications is finally obtained.

[0173] Step 5: Multi-dimensional evaluation of model correction effectiveness;

[0174] This step, based on the precipitation forecast correction results from step 4, uses the true global reanalysis precipitation data as a reference to evaluate the model correction effect from three dimensions: core indicators, inter-regional performance, and stability, thus verifying the model's effectiveness. The specific steps are as follows:

[0175] 5.1 Calculation of core evaluation indicators;

[0176] The overall error index and precipitation event score index are calculated to quantify the improvement in correction effectiveness. The definitions and formulas for each index are as follows:

[0177] Bias: Used to characterize either the original precipitation forecast or the corrected final precipitation forecast. The degree of systematic deviation relative to the true precipitation data from reanalysis;

[0178] Threat Score (TS): ,in The hit count represents the number of grid points where both the forecast and actual data show precipitation, and these must correspond to the same precipitation level. The false alarm count represents the number of grid points where precipitation was predicted but the actual value was not. The number of missed reports indicates the number of grid points where the forecast predicted no precipitation but the actual value showed precipitation. Threat score, dimensionless, ranges from 0 to 1, with values ​​closer to 1 indicating higher forecast accuracy;

[0179] Hit Rate (POD): ,in Hit rate, dimensionless, ranging from 0 to 1, characterizing the ability to capture precipitation events;

[0180] 5.2 Interval division and comparative analysis;

[0181] Interval Assessment: Based on the China Meteorological Administration's 24-hour precipitation standards, precipitation is classified into levels (trace rainfall <0.1mm, light rain 0.1-9.9mm, moderate rain 10-24.9mm, heavy rain 25-49.9mm, torrential rain 50-99.9mm, extremely heavy rain 100-249.9mm, and exceptionally heavy rain ≥250mm). The focus is on the heavy rain, torrential rain, and extremely heavy rain intervals. The deviation, TS score, and POD hit rate of the key intervals are calculated, and the correction effect of heavy rain, torrential rain, and extremely heavy rain is analyzed in detail.

[0182] Comparative analysis: The improvement in deviation, the increase in TS score, and the increase in POD hit rate for each level interval are statistically analyzed. The proportion of intervals with significant improvement is calculated to quantify the overall effectiveness of the correction plan.

[0183] 5.3 Stability analysis;

[0184] Time-of-Lead Stability Validation: Based on forecast data from the first three time periods (24h, 48h, 72h) and the last four time periods (96h, 120h, 144h, 168h), the correction effects, TS scores, and POD hit rates were analyzed to ensure that the model performs consistently across the entire time-of-lead time range.

[0185] Spatial Adaptability Analysis: Taking China as the research scope, and combining nine types of terrain features, the model's adaptability to different terrain regions was verified; a comparison map of precipitation spatial distribution including a base map of China was generated simultaneously to assist in the analysis of TS scores and POD hit rates in different terrain regions, ensuring the model's adaptability to spatial heterogeneity.

[0186] Through the above steps, the model achieves a complete process from multimodal feature learning to accurate correction of precipitation forecasts, taking into account precipitation existence judgment, intensity prediction, and extreme value correction, significantly improving the accuracy of precipitation forecasts. The specific results are as follows:

[0187] (1) Deviation convergence: For forecast deviations of moderate to heavy rainfall, the correction scheme of this invention achieves significant optimization across the entire forecast period; such as Figure 4 As shown in the comparison chart of deviations for moderate to heavy rain, the average absolute value of the deviation has improved by more than 40%.

[0188] (2) Improved scoring system: The TS score has been significantly improved in the levels of heavy rain and torrential rain. The 96-hour lead time for torrential rain has increased from 0.019 to 0.047. Figure 5 The comparison chart of TS scores for heavy rain and above is shown below;

[0189] (3) Improved hit rate (POD): The hit rate for heavy rain, torrential rain, and extremely heavy rain levels has significantly improved. The hit rate for torrential rain has increased from 0.128 to 0.215 within 120 hours. Figure 6 As shown in the comparison chart of hit rates for heavy rain and above, this invention demonstrates a significant enhancement in its ability to capture precipitation events.

[0190] Multiple forecast start times from June to August 2019 were selected, and three forecast data with forecast lead times of 24h, 48h, and 72h were extracted. The model forecast effect was verified and the performance was tested. The original forecast had problems with false precipitation areas, namely small-area isolated precipitation and underestimation of intensity. This invention makes the spatial distribution of precipitation more closely match the true value by enhancing false precipitation filtering and gating residual correction. This is completely consistent with the post-processing logic in step 4.3, namely spatial filtering + inverse transformation + physical constraints.

[0191] Multiple forecast start times from June to August 2019 were selected, and four forecast data points with lead times of 96h, 120h, 144h, and 168h were extracted for model forecast effectiveness verification and performance testing, forming full-lead-time coverage with the previous three time periods. Even with a mid-term lead time of 4-7 days, this invention can still effectively suppress false precipitation and correct intensity deviations, making the spatial continuity and intensity accuracy of precipitation closer to the true value. This is completely consistent with the conclusion of "consistent optimization across the entire lead time" in the lead-time stability verification in step 5.3, verifying the model's adaptability to the accumulation of mid-term forecast errors.

[0192] This invention leverages the strong nonlinear fitting capabilities of deep learning, combined with spatial perception, multi-intensity hierarchical optimization, and quantitative fusion of topographic factors, to balance the requirements for refinement and reliability across different timeframes. It can effectively improve the accuracy, spatiotemporal consistency, and operational applicability of precipitation forecasts. It has been fully validated using 1-7 day timeframe data and can be extended to other timeframes through parameter fine-tuning.

[0193] In summary, the method of this invention first acquires gridded precipitation forecast data for the target area, then analyzes the true precipitation data and raw topographic data, and performs preprocessing such as data cleaning and invalid value removal; referring to the national standard rainfall level, precipitation intensity levels are classified, and a multi-intensity supervision label system is established; topographic classification data is integrated into features in the form of 9 uniquely hot codes, including deep sea, shallow sea, coastal plain, plain, hills, low mountains, medium mountains, high mountains, and extremely high mountains; based on the preprocessed data, a comprehensive multimodal forecast factor library containing core forecast features, spatiotemporal features, topographic features, and statistical features is generated; and a system integrating spatial perception and multimodal forecasting is constructed. A deep learning model for intensity-gated residuals is proposed. This model comprises a shared feature extraction network, a precipitation presence prediction head, and a residual prediction head. The shared feature extraction network enhances feature robustness through batch normalization (BatchNorm) and dropout. The spatial awareness module incorporates geospatial information through latitude and longitude standardization and linear projection. The multi-intensity gating mechanism dynamically adjusts the residual correction weights based on the precipitation presence probability output by the model, according to different precipitation intensity levels. An improved composite loss function combination strategy is employed, including EnhancedSpatial loss function for precipitation presence identification. Zero Penalty Loss (enhanced spatial zero-value penalty loss function), LogHuber Loss (log space robust loss) for residual correction regression, and Quantile Loss (quantile loss) for extreme precipitation fitting optimization are used, and the weight coefficients of each loss function are dynamically adjusted based on the distribution characteristics of precipitation samples. The proportion of precipitation samples of different intensities is balanced by weighted random sampling, and the model is iteratively trained using a dynamic learning rate scheduling strategy and an early stopping mechanism. For the target forecast time, based on the precipitation forecast data, the log space gated residual correction is obtained through model inference. After being restored to the actual precipitation physical quantity through inverse logarithmic transformation, the false precipitation filtering strategy is combined to remove false precipitation areas and generate the final precipitation correction forecast.

[0194] This invention effectively captures the spatiotemporal distribution characteristics of precipitation and its coupling with terrain by using spatial feature fusion and a multi-intensity gated residual mechanism. In particular, it enhances the accurate fitting of extreme precipitation and the effective suppression of false precipitation. The composite loss function and dynamic weight strategy employed balance the learning priorities of precipitation existence identification, regular precipitation intensity correction, and extreme precipitation optimization, exhibiting stronger nonlinear fitting capabilities and meteorological-physical consistency than traditional methods. Through an end-to-end deep learning framework, it achieves direct mapping from raw numerical forecasts to accurate correction results without the need for manual feature selection, thereby improving the accuracy of precipitation forecasts and the efficiency of operational applications. It has significant practical application value for scenarios such as refined precipitation forecasting and meteorological support for flood control and drought relief in meteorological operations.

[0195] The present invention also provides a system corresponding to the method, comprising:

[0196] The data preprocessing unit is used to perform unified preprocessing on gridded precipitation forecast data, reanalysis precipitation true data, and topographic classification data.

[0197] The multimodal forecast factor library construction unit is used to perform targeted transformation, adaptive encoding and fusion of multimodal features affecting precipitation based on preprocessed data of various types, generate multimodal forecast factor vectors, and further construct a multimodal forecast factor library;

[0198] The model building and training unit is used to construct a spatially aware and multi-intensity gated residual precipitation forecast correction model. Through multi-loss function collaborative optimization, the model is trained and validated to obtain a trained model. The model includes a shared feature extraction network, a spatial awareness module, a precipitation existence prediction head, a residual prediction head, and a power-gated unit. The shared feature extraction network takes a multimodal forecast factor library as input and outputs a deep fusion feature vector. The spatial awareness module takes preprocessed latitude and longitude features as input and outputs an enhanced spatial representation feature vector. The output of the spatial awareness module is element-wise added and fused with the output of the shared feature extraction network to enhance the spatial heterogeneity capture capability of the deep fusion features. The precipitation existence prediction head takes a deep fusion feature vector after fusing spatially encoded features as input and outputs the precipitation existence probability. The residual prediction head takes a deep fusion feature vector after fusing spatially encoded features as input and outputs a logarithmic spatial residual correction. The power-gated unit takes the precipitation existence probability and the logarithmic spatial residual correction as input and outputs a gated weighted logarithmic spatial residual correction.

[0199] The model optimization unit is used to optimize model parameters through iterative training and combine regularization strategies to prevent overfitting. At the same time, it performs inverse transformation and physical constraints on the gated weighted log space residual correction of the model output to obtain precipitation forecast correction results.

[0200] The evaluation unit is used to assess the model correction effect based on the precipitation forecast correction results, with reference to the true precipitation data from reanalysis, from three dimensions: core indicators, inter-regional performance, and stability, and to verify the effectiveness of the model.

[0201] The present invention also provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the method described.

Claims

1. A method for spatial perception and multi-intensity gated residual precipitation forecast revision based on deep learning, characterized in that, Includes the following steps: Unified preprocessing is performed on gridded precipitation forecast data, reanalysis precipitation true data, and raw topographic data; Based on the preprocessed data, targeted transformation, adaptive encoding and fusion are performed on the multimodal characteristics affecting precipitation to generate multimodal forecast factor vectors, and a multimodal forecast factor library is further constructed. A spatially-aware and multi-intensity-gated residual precipitation forecast correction model was constructed. The model was trained and validated through multi-loss function collaborative optimization to obtain a well-trained model. The model includes a shared feature extraction network, a spatial awareness module, a precipitation existence prediction head, a residual prediction head, and a power-gated unit. The shared feature extraction network takes a multimodal forecast factor library as input and outputs a deep fusion feature vector. The spatial awareness module takes preprocessed latitude and longitude features as input and outputs an enhanced spatial representation feature vector. The output of the spatial awareness module is element-wise added to the output of the shared feature extraction network to enhance the spatial heterogeneity capture capability of the deep fusion features. The precipitation existence prediction head takes a deep fusion feature vector after fusing spatially encoded features as input and outputs the precipitation existence probability. The residual prediction head takes a deep fusion feature vector after fusing spatially encoded features as input and outputs a logarithmic spatial residual correction. The power-gated unit takes the precipitation existence probability and the logarithmic spatial residual correction as input and outputs a gated weighted logarithmic spatial residual correction. The model parameters are optimized through iterative training, and a regularization strategy is used to prevent overfitting. At the same time, the logarithmic space residual correction amount after gated weighting of the model output is subjected to inverse transformation and physical constraints to obtain the precipitation forecast correction results. Based on the precipitation forecast correction results, and with reference to the true global reanalysis precipitation data, the model correction effect is evaluated from three dimensions: core indicators, inter-regional performance, and stability, thus verifying the model's effectiveness.

2. The method according to claim 1, characterized in that, Preprocessing includes: Based on the preset latitude and longitude range slices, extract the grid data of the corresponding specifications to make the spatial grids of the three types of data completely aligned; By accurately matching the timestamps of the forecast initialization time and the forecast lead time, a one-to-one correspondence between grid precipitation forecast data and reanalysis precipitation true data is established. Identify and remove invalid values ​​from the data, including missing values ​​and outlier values ​​that are outside the physically reasonable range; Based on the pre-defined terrain classification results, the terrain classification labels corresponding to each grid point are read from the terrain grid data.

3. The method according to claim 1, characterized in that, Targeted transformation, adaptive encoding, and fusion are performed on the multimodal features affecting precipitation, including: The targeted transformation and adaptive coding of core forecast features include: the core forecast features are gridded precipitation forecast data; a logarithmic transformation is performed on the preprocessed gridded precipitation forecast data to obtain continuous features after targeted transformation; Z-Score normalization coding is performed on the continuous core forecast features after targeted transformation to obtain core forecast modal coding features. The targeted transformation and adaptive encoding of spatiotemporal features include: forecast lead time extraction: directly extracting the original numerical values ​​of the matched forecast lead time; longitude and latitude extraction of grid points: directly extracting the original longitude and latitude coordinates of the target area; targeted transformation of the flood season month at the forecast initialization time: classifying the flood season month at the forecast initialization time to clarify the flood season month's affiliation; Z-Score normalization encoding is performed on the forecast lead time and longitude / latitude as continuous features; the flood season month is a discrete feature, and its numerical form is directly preserved; thus, the spatiotemporal modal coding features are obtained. The targeted transformation and adaptive encoding of terrain features include: reuse of basic terrain labels: directly using the terrain category labels already defined in the data preprocessing stage; construction of terrain-derived interactive features: based on the first few categories of terrain category labels, element-by-element operations are performed with the original latitude data of the grid to construct terrain-latitude interactive features, thus obtaining terrain-derived interactive features; one-hot encoding is performed on discrete terrain category labels, while terrain-latitude interactive features are not encoded or standardized; the one-hot encoded terrain features are concatenated with the terrain-latitude interactive features to obtain terrain modal coding features. The targeted transformation and adaptive coding of statistical features include: the targeted transformation of statistical features into the historical no-precipitation frequency features of each grid point; and then performing Z-Score normalization coding on the historical no-precipitation frequencies to obtain statistical mode coding features. Multimodal feature fusion includes: sequentially concatenating the obtained core forecast modal coding features, spatiotemporal modal coding features, topographic modal coding features, and statistical modal coding features according to their dimensions to form a multimodal forecast factor vector, and generating a multimodal forecast factor library based on the multimodal forecast factor vector.

4. The method according to claim 1, characterized in that, The shared feature extraction network consists of two fully connected layers, a batch normalization layer, a LeakyReLU activation function layer, and a Dropout regularization layer connected in series. Output a multi-dimensional deep fusion feature vector; The spatial perception module maps the standardized latitude and longitude features into a multi-dimensional spatial feature vector through a linear projection layer. After multiplying the multi-dimensional spatial feature vector with the adaptively learned feature fusion weights, it performs element-wise addition and fusion with the deep feature vector output by the shared feature extraction network to enhance spatial perception capabilities. The precipitation existence prediction head consists of two fully connected layers and an activation function layer. The outputs of the two fully connected layers are connected to the input of the activation function layer, and the activation function layer outputs the probability of precipitation existence. The residual prediction head consists of a single fully connected layer and outputs log-space residual corrections. The power-gated unit is used to perform power-weighted calculations. The inputs are the probability of precipitation and the logarithmic space residual correction, and the output is the logarithmic space residual correction after gating and weighting.

5. The method according to claim 4, characterized in that, The power-gated unit employs a power-gated strategy to dynamically control the impact of residual corrections on the final forecast results. The residual weights are adjusted based on the probability of precipitation occurrence, as shown in the formula: ； in, For the first Logarithmic space residual correction for grid points after power-gated weighting For the first Precipitation at each grid point has a probability. For power-order parameters, For the first Logarithmic space residual correction for each grid point.

6. The method according to claim 1, characterized in that, Multi-loss function collaborative optimization includes: Joint Enhanced Spatial Zero-Value Penalty Loss Logarithmic Huber loss and quantile loss Construct the total loss function The expression is: ； in, To enhance the spatial zero-value penalty loss weight, For the logarithmic Huber loss weights, Weights for quantile loss. To assign a fixed weight to the penalty for false precipitation, Set a fixed weight for the deviation penalty term. Set a fixed weight for the residual penalty term. For false precipitation penalties, This is a deviation penalty item. This is a residual penalty term.

7. The method according to claim 1, characterized in that, Perform inverse transformation and physical constraints on the log-space residual correction of the model output after gated weighting, including: Spatial filtering optimization: Based on spatial connectivity verification, during the model training phase, false precipitation regions with low precipitation probability below the first preset continuous grid threshold are removed, as are false precipitation regions with precipitation probability below the first preset threshold. During the verification and inference phases, false precipitation regions with low precipitation probability below the second preset continuous grid threshold are removed, as are false precipitation regions with precipitation probability below the second preset threshold. Regions with an original forecast of 0 use the same probability threshold as the corresponding phase. Inverse logarithmic space transformation: This involves correcting the logarithmic space residuals after gated weighting. The inverse transformation formula for converting the corrected result to the actual spatial distribution of precipitation is: ； ； in, For the model after correction The precipitation value of each grid point represents the intensity of the final corrected precipitation forecast for each grid point; For the first The log-space original forecast value of the grid point is the _th _th_ grid point. The result of logarithmic transformation of the original precipitation forecast values ​​for each grid point; For the first Logarithmic space residual correction for grid points after power-gated weighting; For the preprocessed first Precipitation forecast data for each grid point; To obtain An exponential function with base 0. For taking the logarithmic function; Physical scope clipping: After model correction, the first... Precipitation values ​​at each grid point By constraining the data within a preset range and eliminating outliers that exceed the physical range, the precipitation forecast correction results are obtained.

8. The method according to claim 1, characterized in that, The model correction effectiveness was evaluated from three dimensions: core metrics, interval performance, and stability, including: The overall error index and precipitation event scoring index are calculated to quantify the improvement of the correction results. The overall error index is the deviation, used to characterize the degree of systematic deviation between the original precipitation forecast value or the precipitation forecast correction result and the true precipitation data of the reanalysis. The precipitation event scoring index includes threat score and hit rate. The performance of the interval classification includes: Interval classification assessment: Classifying the intervals according to precipitation standards, focusing on the situation of heavy rain, rainstorm, and torrential rain, calculating the deviation, threat score and hit rate of each interval, and analyzing the correction effect of heavy rain, rainstorm, and torrential rain; Comparative analysis: Statistics on the deviation improvement, threat score improvement and hit rate improvement of each level interval, calculating the proportion of intervals with improvement greater than the set threshold, and quantifying the overall effectiveness of the correction plan; Stability analysis: including time-leading stability verification: based on forecast data from multiple time periods before and after the forecast, analyze the correction effect, threat score and hit rate differences at different time periods to ensure that the model performs consistently across the entire time range; Spatial adaptability analysis: Combined with terrain features, the model's adaptability to different terrain regions is verified; a comparison map of precipitation spatial distribution is generated simultaneously to assist in the analysis of threat scores and hit rates in different terrain regions, ensuring the model's adaptability to spatial heterogeneity.

9. A deep learning-based spatial perception and multi-intensity gated residual precipitation forecasting correction system, characterized in that, include: The data preprocessing unit is used to perform unified preprocessing on gridded precipitation forecast data, reanalysis precipitation true data, and raw topographic data. The multimodal forecast factor library construction unit is used to perform targeted transformation, adaptive encoding and fusion of multimodal features affecting precipitation based on preprocessed data of various types, generate multimodal forecast factor vectors, and further construct a multimodal forecast factor library; The model building and training unit is used to construct a spatially aware and multi-intensity gated residual precipitation forecast correction model. Through multi-loss function collaborative optimization, the model is trained and validated to obtain a trained model. The model includes a shared feature extraction network, a spatial awareness module, a precipitation existence prediction head, a residual prediction head, and a power-gated unit. The shared feature extraction network takes a multimodal forecast factor library as input and outputs a deep fusion feature vector. The spatial awareness module takes preprocessed latitude and longitude features as input and outputs an enhanced spatial representation feature vector. The output of the spatial awareness module is element-wise added and fused with the output of the shared feature extraction network to enhance the spatial heterogeneity capture capability of the deep fusion features. The precipitation existence prediction head takes a deep fusion feature vector after fusing spatially encoded features as input and outputs the precipitation existence probability. The residual prediction head takes a deep fusion feature vector after fusing spatially encoded features as input and outputs a logarithmic spatial residual correction. The power-gated unit takes the precipitation existence probability and the logarithmic spatial residual correction as input and outputs a gated weighted logarithmic spatial residual correction. The model optimization unit is used to optimize model parameters through iterative training and combine regularization strategies to prevent overfitting. At the same time, it performs inverse transformation and physical constraints on the gated weighted log space residual correction of the model output to obtain precipitation forecast correction results. The evaluation unit is used to assess the model correction effect based on the precipitation forecast correction results, with reference to the true precipitation data from reanalysis, from three dimensions: core indicators, inter-regional performance, and stability, and to verify the effectiveness of the model.

10. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the method described in any one of claims 1-8.

Images