A Short-Term Heavy Precipitation Forecasting Method and System Based on Deep Learning and Multi-Model Fusion

Abstract

This application discloses a method and system for short-term heavy precipitation forecasting based on deep learning multi-model fusion, relating to the field of meteorological forecasting. The specific steps are as follows: acquiring multi-dimensional radar product data and regional precipitation data within a preset time period; preprocessing the regional precipitation data and performing temporal and spatial matching between the multi-dimensional radar product data and the preprocessed regional precipitation data to obtain matching data; filtering the matching data corresponding to periods of heavy precipitation and constructing a dataset; training multiple short-term heavy precipitation forecasting sub-models based on the dataset, and fusing the outputs of the multiple short-term heavy precipitation forecasting sub-models based on a preset fusion mechanism to obtain the final short-term heavy precipitation forecast result. The overall forecasting system of this invention exhibits stronger adaptability and stability to various convective intensities and maintains high prediction accuracy even in complex precipitation scenarios, providing a high-precision and highly generalizable technical solution for minute-level short-term heavy precipitation forecasting.

Description

Technical Field

[0001] This application relates to the field of meteorological forecasting, and in particular to a method and system for forecasting short-term heavy precipitation based on deep learning multi-model fusion. Background Technology

[0002] Currently, existing 0-2 hour minute-level short-term forecasts still heavily rely on extrapolation methods based on real-time observations from radar and satellites. These methods focus on linear extrapolation of echo movement, but they are weak in characterizing the dynamic mechanisms of convection formation and dissipation, making it difficult to accurately capture the rapid organization and fine structural changes of severe convection. This results in significant limitations in forecasting the location, intensity, and movement of heavy precipitation. Furthermore, traditional forecasting processes rely on manual feature extraction and thresholding, making it difficult to fully extract detailed information at the small and medium scales within the data. Statistical models based on experience and physical parameters are also limited by subjective perception, making them prone to missed predictions, false alarms, and structural biases in complex severe convective scenarios. Overall, there is still significant room for improvement in forecast performance. To overcome these bottlenecks, deep learning technology has been introduced into the field of precipitation nowcasting in recent years. However, existing single deep learning precipitation forecasting models have significant limitations: their forecasting performance varies significantly across different precipitation levels, and their overall generalization ability is insufficient. They are also weak in identifying severe precipitation events such as rainstorms, easily leading to missed reports and magnitude deviations. Furthermore, precipitation location errors are large, with prominent issues of rain area shifts and center misalignment. False alarms and inaccurate reports frequently occur during periods of weak precipitation, resulting in poor forecast stability and fault tolerance. At the same time, single models struggle to adapt to complex and variable weather systems, lack effective error correction methods, and have limited overall forecast accuracy and operational applicability.

[0003] Therefore, overcoming the multiple shortcomings of a single deep learning model in short-term heavy precipitation forecasting is a technical problem that urgently needs to be solved. Summary of the Invention

[0004] The purpose of this application is to provide a method and system for short-term heavy precipitation forecasting based on deep learning multi-model fusion, which overcomes the above-mentioned shortcomings.

[0005] To achieve the above objectives, this application provides the following solution: Firstly, this application provides a short-term heavy precipitation forecasting method based on deep learning multi-model fusion, the specific steps of which are as follows: Acquire multidimensional radar product data and regional precipitation data within a preset time period; The regional precipitation data is preprocessed, and the multidimensional radar product data is matched temporally and spatially with the preprocessed regional precipitation data to obtain matching data. Filter matching data corresponding to periods of heavy rainfall and construct a dataset; Multiple short-term heavy precipitation forecast sub-models are trained based on the dataset, and the output results of the multiple short-term heavy precipitation forecast sub-models are fused based on a preset fusion mechanism to obtain the final short-term heavy precipitation forecast result.

[0006] Optionally, the multi-dimensional radar product data includes multi-level basic reflectivity, liquid water content, and echo top height.

[0007] Optionally, the preprocessing of the regional precipitation data specifically involves performing temporal resolution accumulation and interpolation on the regional precipitation data.

[0008] Optionally, the steps for constructing the dataset are as follows: The preprocessed regional precipitation data and the multidimensional radar product data are standardized to obtain precipitation grayscale maps and multidimensional radar product grayscale maps, respectively. A set of input samples is composed of multi-frame precipitation grayscale images with a preset interval unit time resolution and multi-dimensional radar product grayscale images. The input samples of each group are divided into sample image sequences and control labels according to a preset ratio, and the input samples of each group are aggregated to construct the dataset.

[0009] Optionally, the short-term heavy precipitation forecast sub-model includes a first short-term heavy precipitation forecast sub-model based on SmaAT-Unet, a second short-term heavy precipitation forecast sub-model based on Swin-Transformer, and a third short-term heavy precipitation forecast sub-model based on PhyDnet.

[0010] Optionally, the preset fusion mechanism execution steps are as follows: The TS scores for each of the short-term heavy precipitation forecast sub-models corresponding to different precipitation levels are calculated, and the dynamic fusion weights of each of the short-term heavy precipitation forecast sub-models are calculated based on the TS scores.

[0011] Optionally, the calculation expression for the dynamic fusion weight is: ; In the formula, It is the first The weights of each sub-model; It is the first TS score of individual sub-models; yes The sum of the TS scores of each sub-model.

[0012] Secondly, this application provides a short-term heavy precipitation forecasting system based on deep learning multi-model fusion, comprising: The data acquisition module is used to acquire multidimensional radar product data and regional precipitation data within a preset time period; The data matching module is used to preprocess the regional precipitation data and perform temporal and spatial matching between the multidimensional radar product data and the preprocessed regional precipitation data to obtain matching data. The dataset construction module is used to filter matching data corresponding to periods of heavy rainfall and construct datasets. The precipitation forecast module is used to train multiple short-term heavy precipitation forecast sub-models based on the dataset, and to fuse the output results of the multiple short-term heavy precipitation forecast sub-models based on a preset fusion mechanism to obtain the final short-term heavy precipitation forecast result.

[0013] According to the specific embodiments provided in this application, this application has the following technical effects: 1. This invention uses multidimensional radar product data and minute-level regional precipitation data observed by automatic weather stations as input after spatiotemporal matching, and constructs a dataset, rather than relying solely on a single radar combination reflectivity or a single precipitation data. By introducing multidimensional, high spatiotemporal resolution observation information, it can provide richer initial field information at small and medium scales for forecast models, effectively enhancing the ability to characterize the formation, dissipation, movement, and morphological evolution of strong convective systems, thereby providing a more accurate data foundation for subsequent multi-model fusion forecasts; 2. This invention employs three mainstream algorithms—SmaAT-Unet, Swin-Transformer, and PhyDnet—as baseline forecasting models and introduces a dynamic weight fusion strategy based on TS scores. This strategy adaptively adjusts the fusion weights according to the historical scores of each model at different precipitation levels, achieving real-time complementarity of the advantages of each model and reducing false alarm and missed alarm rates.

[0014] 3. Because the dynamic weight design can flexibly allocate the model contribution based on the actual precipitation level, the overall forecasting system of this invention has stronger adaptability and stability to various convective intensities and can still maintain high prediction accuracy under complex precipitation scenarios, providing a high-precision and highly generalizable technical solution for minute-level short-term heavy precipitation forecasting. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a schematic diagram of a method flow provided in an embodiment of this application. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] To make the above-mentioned objectives, features and advantages of this application more readily apparent, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] This embodiment discloses a short-term heavy precipitation forecasting method based on deep learning multi-model fusion, the specific steps of which are as follows: S1. Obtain multi-dimensional radar product data and regional precipitation data within a preset time period; S2. Preprocess the regional precipitation data and perform temporal and spatial matching between the multidimensional radar product data and the preprocessed regional precipitation data to obtain matching data; S3. Filter the matching data corresponding to the period of heavy precipitation and construct the dataset; S4. Train multiple short-term heavy precipitation forecast sub-models based on the dataset, and fuse the output results of the multiple short-term heavy precipitation forecast sub-models based on the preset fusion mechanism to obtain the final short-term heavy precipitation forecast result.

[0020] Furthermore, this embodiment focuses on deep fusion of multiple models and dynamic intelligent weighting. Specifically, firstly, based on a unified spatiotemporal sequence data standard, three deep learning baseline models with complementary advantages are constructed and trained in parallel: SmaAt-UNet is used to efficiently extract local refined precipitation features, Swin-Transformer is used to capture long-distance dependencies, and PhyDnet, which models the temporal evolution of precipitation systems, is used for physical constraints. Secondly, on independent validation sets, the Threat Score (TS) of each model is evaluated for different precipitation thresholds (such as light rain and heavy rain) to quantify the forecasting skills of each model under specific weather conditions. Based on this, a dynamic weighting algorithm based on the TS score is designed to assign optimal weights to each model at different precipitation levels, achieving adaptive integration that "maximizes strengths and avoids weaknesses." Finally, a comprehensive forecast product is generated through a weighted fusion strategy, effectively overcoming the defects of "false alarms" or "missed alarms" of single models under extreme weather conditions, and improving the forecast accuracy and operational stability under complex meteorological conditions.

[0021] In one embodiment, the multidimensional radar product data and regional precipitation data in S1 are minute-level radar product data (including nine levels of basic reflectivity, liquid water content, echo top height, etc.) with unit time resolution and unit spatial resolution within a preset time period, and precipitation data monitored by regional automatic weather stations.

[0022] In one embodiment, the preprocessing of regional precipitation data specifically involves performing temporal resolution accumulation and interpolation on minute-level regional precipitation data monitored by regional automatic weather stations.

[0023] In one embodiment, the steps for constructing the dataset are as follows: The preprocessed regional precipitation data and multidimensional radar product data were standardized to obtain precipitation grayscale maps and multidimensional radar product grayscale maps. A set of input samples is composed of multi-frame precipitation grayscale images with a preset interval unit time resolution and multi-dimensional radar product grayscale images. The input samples of each group are divided into sample image sequences and control labels according to a preset ratio, and the input samples of each group are aggregated to construct a dataset.

[0024] Furthermore, the dataset is constructed by filtering multidimensional radar product data and regional precipitation data within the period of heavy precipitation (i.e., from start to end). The dataset includes a training set, a validation set, and a test set, specifically: The multidimensional radar product data (including nine levels of basic reflectivity, liquid water content, echo top height, etc.) and the precipitation data monitored by regional automatic weather stations (i.e. regional precipitation data) are standardized to obtain grayscale maps (including precipitation grayscale maps and multidimensional radar product grayscale maps). Multiple frames of radar and precipitation grayscale images with a preset time interval and unit time resolution are taken to form a set of input samples. The time interval between each image in the input sample is a unit time. The multiple frames in the sample are labeled as follows: ; Among them, the former The frame image is a sequence of sample images input to the network; later The frame image serves as the control label for this group; The samples are divided into training set, test set and validation set according to a preset ratio of 8:1:1.

[0025] In one embodiment, the steps for obtaining the final short-term heavy precipitation forecast result are as follows: The first, second, and third short-term heavy precipitation forecast sub-models were established based on three deep learning methods: SmaAT-Unet, Swin-Transformer, and PhyDnet. Dynamic weights are derived based on the graded TS score; The final short-term heavy precipitation forecast result is obtained by fusing the forecast results of three short-term heavy precipitation forecast sub-models with dynamic weights.

[0026] Furthermore, the core of SmaAt-UNet is the encoder-decoder structure of U-Net, and the "SmaAt" in its name represents its two key improvements: Lightweight convolution (Small): Replaces traditional convolution with depthwise separable convolution (DSC). This convolution method decomposes a standard convolution into two steps: depthwise convolution and pointwise convolution, which can reduce the number of model parameters and computational cost.

[0027] Attention Mechanism: A Convolutional Block Attention Module (CBAM) is introduced into the convolutional blocks of the encoder. CBAM weights the feature maps sequentially from both channel and spatial dimensions, enabling the model to focus more on important feature regions while suppressing irrelevant background information.

[0028] Furthermore, the main idea behind PhyDNet temporal network is to attempt to build a physical constraint model using deep networks. This is achieved by using convolution to simulate partial derivatives and moment loss for supervision, learning physical information to supplement existing networks (ConvLSTM). The information captured by existing deep networks is used to model prior physical knowledge (i.e., the information on the left-hand side) through partial differential equations. Finally, the physical information and existing information are combined to obtain better results.

[0029] Furthermore, the Swin Transformer (Shift Window Transformer) is a hierarchical visual Transformer architecture. It was designed to address the bottlenecks of early visual Transformers (such as ViT) in terms of computational efficiency and adaptability to multi-scale tasks, and has now become the mainstream general-purpose backbone network in the field of computer vision.

[0030] Windowed Self-Attention (W-MSA). To reduce computational cost, the Swin Transformer no longer computes global attention. Instead, it divides the image into multiple non-overlapping local windows (e.g., 7×7 size) and performs self-attention computation only within each window. This change reduces the computational complexity from quadratic (O(N)) to quadratic (O(N)) time complexity. 2 The speed was reduced to linear level (O(N)), enabling it to efficiently process high-resolution images.

[0031] Shifted Window Mechanism (SW-MSA). Simple windowing attention can limit the receptive field of the model, preventing information exchange between different windows. To address this, the Swing Transformer alternates between regular window partitioning and shifted window partitioning within consecutive Transformer blocks. By shifting the window by half its size in the next computation, regions that originally belonged to different windows can be connected within the new window, thus enabling cross-window information exchange and enhancing the model's modeling capabilities.

[0032] Hierarchical Feature Architecture. Borrowing from the design principles of CNNs, the Swin Transformer uses an operation called "PatchMerging" to progressively merge adjacent image patches deep within the network. This process continuously reduces the spatial resolution of the feature maps while increasing the number of channels, ultimately constructing a hierarchical feature representation similar to a Feature Pyramid Network (FPN). This allows the Swin Transformer to adapt naturally to various visual tasks such as object detection and semantic segmentation, just like CNNs.

[0033] Furthermore, each sample is divided into 40 frames according to the dataset, loaded into each sub-model by a loader, and then training begins.

[0034] During sub-model training, the batch size is 10, the number of iterations is 50, the loss function is weighted MAE, the learning rate is 0.001, and the optimization function is Adam. A timestep is incorporated into the sub-model, assuming each radar image represents one timestep. Utilizing a many-to-one approach, Timesteps=10 is used as input, meaning 10 radar echo images and precipitation time-series images are simultaneously input, resulting in a 3-dimensional array with a resolution of 480×480, forming an input of shape [10,10,12,480,480]. This input is then sequentially passed through SmaAT-Unet, PhyDNet, and Swin-Transformer models to obtain 30 consecutive frames of precipitation forecast images with a resolution of 480×480 at future time points, also with a shape of [10,10,1,480,480], thus enabling the predictions of each sub-model.

[0035] Furthermore, a dynamic weight design based on TS scores is adopted. The core idea is that the better a sub-model performs on a specific task (the higher its TS score), the greater the weight its prediction should carry in the fusion process. Design principle: On the validation set, the TS score of each sub-model is calculated separately for different precipitation thresholds, i.e., precipitation levels (such as light rain, heavy rain, and torrential rain).

[0036] The dynamic fusion weights are specifically designed as follows: the TS scores of each sub-model can be normalized and directly used as their weights. For example, for rainstorm forecasting, if the TS scores of sub-models A, B, and C are 0.5, 0.3, and 0.2 respectively, then their fusion weights can be designed as 0.5, 0.3, and 0.2.

[0037] This method can adaptively select the sub-model that is "best" in the current scenario, thereby improving the accuracy of forecasts for key weather conditions (such as heavy rain).

[0038] Furthermore, the specific processing steps for the final short-term heavy precipitation forecast results are as follows: Step 1: Assuming there are N sub-models, for a specific precipitation threshold (e.g., precipitation ≥ 5 mm in 6 minutes), the first sub-model... The TS score of each sub-model is .

[0039] Calculate the TS score for each sub-model: First, the TS score for each sub-model is calculated on an independent validation set.

[0040] :TS score of sub-model 1; :TS score of sub-model 2; ... :TS score of submodel N.

[0041] Step 2: Calculate the normalized weights: Divide the TS score of each sub-model by the sum of the TS scores of all sub-models to obtain the normalized weight of that sub-model. Weight calculation formula: ; In the formula, It is the first The weights of each sub-model; It is the first TS score of individual sub-models; yes The sum of the TS scores of each sub-model.

[0042] Step 3: Generate fusion prediction results: Final fusion prediction results These are the prediction results of all sub-models. The weighted sum and fusion prediction formula are as follows: ; In the formula, It is the final fused forecast field (e.g., a fused precipitation distribution map). It is the first The original forecast field of each sub-model; It is the first The weights calculated by each sub-model.

[0043] This embodiment also discloses a short-term heavy precipitation forecasting system based on deep learning multi-model fusion, including: The data acquisition module is used to acquire multidimensional radar product data and regional precipitation data within a preset time period; The data matching module is used to preprocess regional precipitation data and perform temporal and spatial matching between multidimensional radar product data and preprocessed regional precipitation data to obtain matching data. The dataset construction module is used to filter matching data corresponding to periods of heavy rainfall and construct datasets. The precipitation forecast module is used to train multiple short-term heavy precipitation forecast sub-models based on the dataset, and to fuse the output results of the multiple short-term heavy precipitation forecast sub-models based on the preset fusion mechanism to obtain the final short-term heavy precipitation forecast result.

[0044] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0045] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for short-term heavy precipitation forecasting based on deep learning multi-model fusion, characterized in that, The specific steps are as follows: Acquire multidimensional radar product data and regional precipitation data within a preset time period; The regional precipitation data is preprocessed, and the multidimensional radar product data is matched temporally and spatially with the preprocessed regional precipitation data to obtain matching data; Filter matching data corresponding to periods of heavy rainfall and construct a dataset; Multiple short-term heavy precipitation forecast sub-models are trained based on the dataset, and the output results of the multiple short-term heavy precipitation forecast sub-models are fused based on a preset fusion mechanism to obtain the final short-term heavy precipitation forecast result.

2. The short-term heavy precipitation forecasting method based on deep learning multi-model fusion according to claim 1, characterized in that, The multi-dimensional radar product data includes multi-level basic reflectivity, liquid water content, and echo top height.

3. The short-term heavy precipitation forecasting method based on deep learning multi-model fusion according to claim 1, characterized in that, The preprocessing of the regional precipitation data specifically involves performing temporal resolution accumulation and interpolation on the regional precipitation data.

4. The short-term heavy precipitation forecasting method based on deep learning multi-model fusion according to claim 1, characterized in that, The steps for constructing the dataset are as follows: The preprocessed regional precipitation data and the multidimensional radar product data are standardized to obtain precipitation grayscale maps and multidimensional radar product grayscale maps, respectively. A set of input samples is composed of multi-frame precipitation grayscale images with a preset interval unit time resolution and multi-dimensional radar product grayscale images. The input samples of each group are divided into sample image sequences and control labels according to a preset ratio, and the input samples of each group are aggregated to construct the dataset.

5. The short-term heavy precipitation forecasting method based on deep learning multi-model fusion according to claim 1, characterized in that, The short-term heavy precipitation forecast sub-models include a first short-term heavy precipitation forecast sub-model based on SmaAT-Unet, a second short-term heavy precipitation forecast sub-model based on Swin-Transformer, and a third short-term heavy precipitation forecast sub-model based on PhyDnet.

6. The short-term heavy precipitation forecasting method based on deep learning multi-model fusion according to claim 1, characterized in that, The preset fusion mechanism execution steps are as follows: The TS scores for each of the short-term heavy precipitation forecast sub-models corresponding to different precipitation levels are calculated, and the dynamic fusion weights of each of the short-term heavy precipitation forecast sub-models are calculated based on the TS scores.

7. The short-term heavy precipitation forecasting method based on deep learning multi-model fusion according to claim 6, characterized in that, The calculation expression for the dynamic fusion weight is as follows: ； In the formula, It is the first The weights of each sub-model; It is the first TS score of individual sub-models; yes The sum of the TS scores of each sub-model.

8. A short-term heavy precipitation forecasting system based on deep learning multi-model fusion, characterized in that, include: The data acquisition module is used to acquire multidimensional radar product data and regional precipitation data within a preset time period; The data matching module is used to preprocess the regional precipitation data and perform temporal and spatial matching between the multidimensional radar product data and the preprocessed regional precipitation data to obtain matching data. The dataset construction module is used to filter matching data corresponding to periods of heavy rainfall and construct datasets. The precipitation forecast module is used to train multiple short-term heavy precipitation forecast sub-models based on the dataset, and to fuse the output results of the multiple short-term heavy precipitation forecast sub-models based on a preset fusion mechanism to obtain the final short-term heavy precipitation forecast result.

Images