A Modeling Method for Objective Forecasting of Short-Term Heavy Precipitation that Integrates Multi-Scale Meteorological Features
By integrating multi-scale meteorological features into a short-term heavy precipitation forecasting method, and utilizing a progressive feature pyramid network and a multi-scale spatiotemporal attention module, the problem of insufficient forecast accuracy and weak extreme event forecasting capability in existing technologies is solved, achieving more efficient multi-dimensional forecasting.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HECHI CITY METEOROLOGICAL BUREAU OF GUANGXI ZHUANG AUTONOMOUS REGION
- Filing Date
- 2026-01-22
- Publication Date
- 2026-06-02
AI Technical Summary
Existing short-term heavy precipitation forecasting technologies suffer from problems such as rigid data fusion methods, low feature extraction efficiency, insufficient model adaptability, and lack of physical constraints, resulting in insufficient forecast accuracy, limited product variety, and weak ability to forecast extreme events.
Employing a progressive multi-scale feature pyramid network, a multi-scale spatiotemporal attention fusion module, and a dynamic weight loss function, this study integrates multi-source meteorological data through a multi-task learning framework, adaptively adjusts model parameters, and predicts precipitation probability, intensity, and path.
It improves the efficiency of multi-scale feature extraction, enhances the model's generalization ability, improves the ability to predict extreme events, reduces unreasonable forecasts, and provides richer forecast products.
Smart Images
Figure CN122132719A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of precipitation forecasting technology, and in particular to a method for objective forecasting and modeling short-term heavy precipitation that integrates multi-scale meteorological features. Background Technology
[0002] Short-duration heavy precipitation is a type of severe weather event characterized by its sudden onset and significant impact, and its forecasting has always been a challenge in the meteorological field. Current technologies, particularly traditional numerical weather prediction models, have limited ability to forecast short-duration heavy precipitation. Radar extrapolation methods primarily focus on echo motion, failing to adequately capture changes such as the generation and dissipation of precipitation. In recent years, deep learning techniques have been applied to precipitation forecasting, but existing methods often use only a single data source (such as radar data) or simply stitch together multi-source data, failing to fully explore the complex correlations between multi-scale meteorological characteristics. Furthermore, existing models typically only predict precipitation intensity, neglecting multi-dimensional information such as precipitation probability and movement paths, resulting in limited forecast products. Simultaneously, the scarcity of heavy precipitation samples in the training dataset makes it difficult for models to learn the patterns of extreme events. Although many existing technical solutions have publicly demonstrated their ability to predict short-duration heavy precipitation, these solutions still have many shortcomings.
[0003] Although existing short-term heavy precipitation forecasting technologies have introduced multi-source data fusion and deep learning techniques, they have the following technical defects: (1) Rigid data fusion methods: Existing methods mostly use simple splicing or weighted averaging to fuse multi-source data, failing to fully consider the reliability differences of different data sources under different meteorological conditions. (2) Low feature extraction efficiency: Multi-scale feature extraction often adopts serial or parallel stacking methods, failing to achieve cross-scale dynamic interaction of features. (3) Insufficient adaptability of forecast models: Existing models mostly adopt fixed architectures, unable to dynamically adjust model parameters according to actual weather types and forecast scenarios. (4) Lack of physical constraints: Pure data-driven models may produce forecast results that do not conform to atmospheric physical laws. Therefore, it is necessary to design a modeling method for objective short-term heavy precipitation forecasting that integrates multi-scale meteorological features, and to quickly achieve objective short-term heavy precipitation forecasting through the constructed model. Summary of the Invention
[0004] The purpose of this invention is to provide a modeling method for objective forecasting of short-term heavy precipitation that integrates multi-scale meteorological features, thereby solving the technical problems of insufficient accuracy, single forecast products, and weak forecasting capability for extreme events in existing technologies.
[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0006] A modeling method for objective forecasting of short-term heavy precipitation that integrates multi-scale meteorological features, the method comprising the following steps:
[0007] Step 1: Collect data at several scales and preprocess the collected data;
[0008] Step 2: Construct a progressive feature pyramid network with multiple scales;
[0009] Step 3: Design spatiotemporal attention fusion modules at several scales;
[0010] Step 4: Construct a task learning framework, including a precipitation probability head, a precipitation intensity head, and an optical flow head;
[0011] Step 5: Construct a dynamic weighted loss function;
[0012] Step 6: Model training and optimization, actual rainfall prediction, and then fine-tuning of the model to complete model construction.
[0013] Further, in step 1, radar data, satellite data, numerical model output, reanalysis data and ground observation data are collected. The collected data are preprocessed, including spatiotemporal matching, quality control and feature engineering. In feature engineering, several derived variables are calculated, including vertical integral liquid water content, combined reflectivity, convective effective potential energy and severe weather threat index.
[0014] Furthermore, in step 2, a progressive feature pyramid with several scales is inserted between the encoder and decoder of the U-Net network. The encoder part adopts a branch structure, and each branch uses convolutional kernels of different scales or patches of different sizes to extract features of different scales. The features of different scales are gradually fused from coarse to fine through a progressive fusion module to form a feature pyramid with several levels. The decoder part uses features of several levels for upsampling and prediction.
[0015] Furthermore, in step 2, a bidirectional attention mechanism is set in the feature pyramid network, and a cross-scale bidirectional attention module is designed to achieve mutual enhancement of features at different scales. This includes guiding attention from coarse to fine scale, coarse-scale features guiding fine-scale feature extraction, fine-grained attention from fine to coarse scale, and fine-scale features correcting coarse-scale features.
[0016] Furthermore, in step 3, the spatiotemporal attention fusion module of several scales is the MST-AFM module, which is used to fuse features from radar, satellite, numerical model, reanalysis data and ground observation data. For each data source, a spatiotemporal attention module is first used to capture its important spatial and temporal features respectively. Then, a cross-attention mechanism is used to allow the features from different data sources to interact and adaptively learn the fusion weights. Finally, the weighted fused features are fed into the subsequent network.
[0017] Furthermore, in step 4, the results of three tasks are output simultaneously: the probability of short-term heavy precipitation, the precipitation intensity, and the precipitation movement path. The three tasks share the feature extraction layer, but each has an independent output head. Through multi-task learning, richer representations are learned, and the generalization ability is improved.
[0018] Furthermore, in step 5, the dynamic weight loss function consists of three parts: a weighted cross-entropy loss for heavy precipitation samples, a weighted mean square error loss for precipitation intensity, and a smoothing L1 loss for optical flow. The weights are dynamically adjusted according to the difficulty of the samples. At the same time, a hard sample mining mechanism is introduced to focus on those samples that the model predicts incorrectly.
[0019] Furthermore, in step 6, historical data is used to train the model, a sliding window is used to construct training samples, and early stopping and cross-validation are used to prevent overfitting. During the training process, gradient accumulation and mixed precision training techniques are used to accelerate the training. The trained model is then applied to the corresponding area for rainfall prediction. Then, rainfall data for the corresponding time period is collected, and the collected rainfall is compared with the predicted amount using time as the label. Finally, the influencing factors within the model are adjusted to obtain the final model.
[0020] The present invention, by adopting the above-described technical solution, has the following beneficial effects:
[0021] This invention utilizes a Progressive Multi-Scale Feature Pyramid Network (PM-FPN) to more effectively extract and fuse multi-scale features, capturing information at different scales, from local convection to weather systems. Through a Multi-Scale Spatiotemporal Attention Fusion Module (MST-AFM), it adaptively fuses multi-source data, fully leveraging the complementarity of radar, satellite, and numerical model data. A multi-task learning framework simultaneously predicts precipitation probability, intensity, and path, providing richer forecast products and improving the model's generalization ability. A dynamic weight loss function alleviates the sample imbalance problem, enabling the model to pay more attention to heavy precipitation and difficult samples, thus enhancing its ability to forecast extreme events. Attached Figure Description
[0022] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation
[0023] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and preferred embodiments. However, it should be noted that many details listed in the specification are merely to provide the reader with a thorough understanding of one or more aspects of the present invention, and these aspects of the invention can be implemented even without these specific details.
[0024] like Figure 1As shown, a modeling method for objective forecasting of short-term heavy precipitation that integrates multi-scale meteorological features is presented. The method includes the following steps:
[0025] Step 1: Collect historical data from South China over the past 5 years, including radar reflectivity factors (6 minutes, 1km×1km), Fengyun-4 satellite multi-channel data (infrared, water vapor, etc., 15 minutes, 1km×1km), CMA-MESO 3KM model output (hourly, 3km×3km), ERA5 reanalysis data (hourly, 0.25°×0.25°), and precipitation observations from automatic ground stations (hourly, station data). Interpolate all this data to a uniform 0.01°×0.01° latitude and longitude grid, aligning the time to 5-minute intervals. Calculate radar-derived variables (VIL, combined reflectivity, etc.) and atmospheric instability indices (CAPE, SWEAT, etc.). Organize the data into samples, each sample including data from the past hour (10 time intervals) as input, and precipitation intensity for the next hour (10 time intervals) as a label (for regression tasks). Simultaneously generate the probability of heavy precipitation in the next hour (binary label, threshold 20mm / h) and optical flow field label (calculated from radar echo sequences).
[0026] Step 2: The network input consists of multi-channel, multi-source data. First, a preprocessing convolutional layer adjusts the number of channels to 256. The encoder has four downsampling stages, each consisting of two convolutional blocks and one downsampling layer. In each stage, we use three parallel branches, employing 3x3, 5x5, and 7x7 convolutional kernels (or dilated convolutions with different dilation rates) to extract features at different scales. Then, a progressive fusion module fuses the feature maps from these three branches: first, the smallest-scale feature map is upsampled and added to the medium-scale feature map; then, it is upsampled again and added to the largest-scale feature map to obtain the multi-scale fused feature for that stage. Thus, each stage outputs a multi-scale fused feature, forming a feature pyramid. The decoder progressively restores spatial resolution through upsampling and skip connections (connected to the multi-scale fused features of the corresponding stage in the encoder). Finally, prediction results are generated through three independent output heads (corresponding to three tasks).
[0027] Step 3: This module is inserted after each stage of the encoder to fuse multi-source data. Assume we have three data sources: radar, satellite, and numerical models, which, after passing through the corresponding stages of the encoder, yield feature maps F_r, F_s, and F_n. First, spatial attention and temporal attention are calculated for each feature map (for temporal attention, we use a self-attention mechanism to calculate along the time dimension). Then, the three feature maps are concatenated along the channel dimension, and a cross-attention module is used to calculate the weights of each data source. Specifically, we use a convolutional layer to generate the query (Q), key (K), and value (V), then calculate the weights using scaled dot product attention, and finally sum the weighted feature maps to obtain the fused feature. This fused feature will serve as the input for the next stage.
[0028] Step 4: The headers for the three tasks are as follows: Precipitation Probability Header: A convolutional layer with a sigmoid activation function, outputting the probability of heavy precipitation occurring at each grid point within the next hour. Precipitation Intensity Header: A convolutional layer, outputting the precipitation intensity at each grid point every 6 minutes within the next hour. Optical Flow Header: A convolutional layer, outputting the echo movement vector (two components, u and v) at each grid point every 6 minutes within the next hour.
[0029] Step 5: The total loss function is the weighted sum of the losses from the three tasks: L = λ1 * L_prob + λ2 * L_intensity + λ3 * L_flow. Where L_prob is the weighted cross-entropy loss, with weights dynamically adjusted based on the ratio of positive to negative samples and precipitation intensity; L_intensity is the weighted mean squared error loss, with weights proportional to precipitation intensity; and L_flow is the smoothed L1 loss. λ1, λ2, and λ3 are hyperparameters used to balance the three tasks.
[0030] Step 6: Use the Adam optimizer with an initial learning rate of 0.001, and adjust the learning rate using cosine annealing. Use a batch size of 8 and train for 100 epochs. Use 5-fold cross-validation to select the best model. Deploy on a meteorological operational server. Real-time data is pushed via a message queue, model inference results are stored in a database, and visualized through a WebGIS platform. Warnings are also issued via SMS, app push notifications, and other methods.
[0031] It can automatically adjust the data fusion strategy according to the actual meteorological conditions, perform stably under different weather scenarios, and has strong generalization ability. Through the two-way cross-scale attention mechanism, the feature utilization efficiency is improved by more than 30%, and the introduction of physical constraints reduces unreasonable forecasts by more than 40%.
[0032] Matters not covered in this invention are common knowledge.
[0033] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A modeling method for objective forecasting of short-term heavy precipitation that integrates multi-scale meteorological features, characterized by: The method includes the following steps: Step 1: Collect data at several scales and preprocess the collected data; Step 2: Construct a progressive feature pyramid network with multiple scales; Step 3: Design spatiotemporal attention fusion modules at several scales; Step 4: Construct a task learning framework, including a precipitation probability head, a precipitation intensity head, and an optical flow head; Step 5: Construct a dynamic weighted loss function; Step 6: Model training and optimization, actual rainfall prediction, and then fine-tuning of the model to complete model construction.
2. The method for modeling short-term heavy precipitation objective forecasting by integrating multi-scale meteorological features according to claim 1, characterized in that: In step 1, radar data, satellite data, numerical model output, reanalysis data, and ground observation data are collected. The collected data are preprocessed, including spatiotemporal matching, quality control, and feature engineering. In feature engineering, several derived variables are calculated, including vertical integral liquid water content, combined reflectivity, convective effective potential energy, and severe weather threat index.
3. The method for modeling short-term heavy precipitation objective forecasting by integrating multi-scale meteorological features according to claim 1, characterized in that: In step 2, a progressive feature pyramid of several scales is inserted between the encoder and decoder of the U-Net network. The encoder part adopts a branch structure, and each branch uses convolutional kernels of different scales or patches of different sizes to extract features of different scales. The progressive fusion module gradually fuses the features of different scales from coarse to fine to form a feature pyramid of several levels. The decoder part uses the features of several levels for upsampling and prediction.
4. The method for modeling short-term heavy precipitation objective forecasting by integrating multi-scale meteorological features according to claim 1, characterized in that: In step 2, a bidirectional attention mechanism is set up in the feature pyramid network, and a cross-scale bidirectional attention module is designed to achieve mutual enhancement of features at different scales. This includes guiding attention from coarse scale to fine scale, coarse scale features guiding fine scale feature extraction, fine-grained attention from fine scale to coarse scale, and fine-scale features correcting coarse scale features.
5. The method for modeling short-term heavy precipitation objective forecasting by integrating multi-scale meteorological features according to claim 1, characterized in that: In step 3, the spatiotemporal attention fusion module of several scales is the MST-AFM module, which is used to fuse features from radar, satellite, numerical model, reanalysis data and ground observation data. For each data source, a spatiotemporal attention module is first used to capture its important spatial and temporal features respectively. Then, a cross-attention mechanism is used to allow the features from different data sources to interact and adaptively learn the fusion weights. Finally, the weighted fused features are fed into the subsequent network.
6. The method for modeling short-term heavy precipitation objective forecasting by integrating multi-scale meteorological features according to claim 1, characterized in that: In step 4, the results of three tasks are output simultaneously: the probability of short-term heavy precipitation, the precipitation intensity, and the precipitation movement path. The three tasks share the feature extraction layer, but each has an independent output head. Through multi-task learning, richer representations are learned, and the generalization ability is improved.
7. The method for modeling short-term heavy precipitation objective forecasting by integrating multi-scale meteorological features according to claim 1, characterized in that: In step 5, the dynamic weight loss function consists of three parts: weighted cross-entropy loss for heavy precipitation samples, weighted mean square error loss for precipitation intensity, and smoothing L1 loss for optical flow field. The weights are dynamically adjusted according to the difficulty of the samples. At the same time, a hard sample mining mechanism is introduced to focus on those samples that the model predicts incorrectly.
8. The method for modeling short-term heavy precipitation objective forecasting by integrating multi-scale meteorological features according to claim 1, characterized in that: In step 6, historical data is used to train the model, a sliding window is used to construct training samples, and early stopping and cross-validation are used to prevent overfitting. During the training process, gradient accumulation and mixed precision training techniques are used to accelerate the training. The trained model is then used to predict rainfall in the corresponding area. Then, rainfall data for the corresponding time period is collected, and the collected rainfall is compared with the predicted amount using time as the label. Finally, the influencing factors in the model are adjusted to obtain the final model.