A precipitation forecast correction method based on a cascaded Attention-U-Net
By using a cascaded Attention-U-Net model to perform grid point correction on precipitation forecasts, the problem of large errors in existing precipitation forecasts has been solved, achieving higher accuracy and reliability, and improving the precision of weather forecasts.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHONGQING INST OF GREEN & INTELLIGENT TECH CHINESE ACAD OF SCI
- Filing Date
- 2023-05-09
- Publication Date
- 2026-07-21
Smart Images
Figure CN116679355B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of artificial intelligence and relates to a precipitation forecast correction method based on cascaded Attention-U-Net. Background Technology
[0002] Weather forecasting has always been a focus of scientists because it contributes to improving human living standards in various aspects, including agriculture, transportation, social and economic decision-making, and disaster prevention. Precipitation forecasting is one of the most important tasks in weather forecasting, as precipitation is the primary meteorological factor causing droughts and floods and a key technology for severe weather warnings. However, numerical weather prediction models are susceptible to the influence of initial atmospheric conditions and boundary conditions when forecasting precipitation, often leading to errors in the prediction results. Therefore, conducting research on precipitation forecast correction techniques is crucial, as it is of great significance for promoting sustainable socio-economic development and disaster prevention and mitigation.
[0003] Current techniques for precipitation forecasting employ a spatiotemporal neighborhood probability approach, converting quantitative precipitation forecasts into probabilistic forecasts to address the spatiotemporal uncertainties inherent in model-based precipitation forecasts. However, this approach only provides a tiered probabilistic precipitation forecast field and does not yield a deterministic precipitation correction field. It also fails to reflect extreme precipitation values, thus failing to meet the demands for refined quantitative precipitation forecasting services. Furthermore, the gridded correction model for precipitation is influenced by factors such as the initial field, physical process parameterization, and computational scheme. Numerical forecasts suffer from systematic and random errors, which are also significant factors affecting the accuracy of numerical forecasts for precipitation and other meteorological elements. Summary of the Invention
[0004] To address the aforementioned technical problems, this invention proposes a precipitation forecast correction method based on cascaded Attention-U-Net, comprising the following steps:
[0005] S1: Using precipitation, temperature and time from numerical forecasts as prediction factors, the observed actual precipitation data is divided into grid points with time intervals of n.
[0006] S2: Establish a cascaded Attention-U-Net precipitation forecast correction model; the cascaded Attention-U-Net precipitation forecast correction model includes: a precipitation classification network and a correction network;
[0007] S3: Train a precipitation classification network based on precipitation data, and use the trained precipitation classification network to classify the precipitation at grid points;
[0008] S4: For areas with precipitation, train a correction network using the corresponding actual precipitation data within the area. The trained correction network is used to correct the precipitation in the precipitation area, and the TS score is calculated to evaluate the precipitation correction effect in the corrected precipitation area.
[0009] The beneficial effects of this invention are:
[0010] This invention addresses the characteristics of precipitation forecasting by employing a cascaded Attention-U-Net model to perform grid point corrections on numerical forecasts, reducing the error between models and observations, improving the accuracy and reliability of precipitation forecasts, and also contributing to enhancing the reliability and application value of weather forecast products. Attached Figure Description
[0011] Figure 1 This is a flowchart of a precipitation forecast correction method based on cascaded Attention-U-Net according to the present invention. Detailed Implementation
[0012] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0013] A precipitation forecast correction method based on cascaded Attention-U-Net includes:
[0014] S1: Using precipitation, temperature and time from numerical forecasts as prediction factors, the observed actual precipitation data is divided into grid points with time intervals of n.
[0015] S2: Establish a cascaded Attention-U-Net precipitation forecast correction model; the cascaded Attention-U-Net precipitation forecast correction model includes: a precipitation classification network and a correction network;
[0016] S3: Train a precipitation classification network based on precipitation data, and use the trained precipitation classification network to classify the precipitation at grid points;
[0017] S4: For areas with precipitation, train a correction network using the corresponding actual precipitation data within the area. The trained correction network is used to correct the precipitation in the precipitation area, and the TS score is calculated to evaluate the precipitation correction effect in the corrected precipitation area.
[0018] Precipitation is a discontinuous variable, and its simulation error is affected by terrain but is not evenly distributed. If grid points are modeled separately, there may be cases where precipitation samples in some areas are too few, making model training difficult. This invention first divides the grid to determine whether a certain area has precipitation; then, it performs precipitation regression correction on the areas with precipitation.
[0019] Two Attention-U-Net base models are concatenated to obtain a concatenated Attention-U-Net precipitation forecast correction model, where the two Attention-U-Net base models represent the precipitation classification network and the correction network, respectively.
[0020] The Attention U-Net network structure effectively utilizes both low-level and high-level information. Borrowing from the U-Net model, its network structure comprises two symmetrical parts: the first part operates similarly to a regular convolutional network, employing 3x3 convolutions and pooling downsampling to capture contextual information (i.e., relationships between pixels); the second part is largely symmetrical, using 3x3 convolutions and upsampling for image segmentation. Unlike U-Net, in the feature fusion stage, Attention U-Net adds an attention gate to weight the feature maps generated during downsampling before incorporating them into the upsampling features. This increases the weight of task-relevant regions, resulting in better segmentation performance.
[0021] The classification model uses Focal Loss as the loss function, while the regression model uses Weighted Loss and Perceptual Loss as the loss functions.
[0022] Focal Loss: This loss function can adaptively reduce the weight of "easy-to-distinguish" samples and increase the weight of "hard-to-distinguish" samples, thereby improving classification accuracy.
[0023] Weighted Loss: Primarily used to assign different weights to different precipitation conditions, allowing the model to focus more on key conditions (precipitation).
[0024] Perceptual Loss: This loss function makes the output of the regression model look closer to the actual precipitation distribution to the naked eye, avoiding the model mechanically reducing the loss and making the results look distorted.
[0025] Meteorological elements are input into the first Attention U-Net to obtain the probability of rain / no rain at each grid point. The threshold is set to 0.5, where a probability greater than 0.5 indicates rain and a probability less than 0.5 indicates no rain. Then, the classification features are combined with the precipitation correction factor and input into the second Attention U-Net structure for precipitation regression. The output layer uses a 1x1 convolutional layer to obtain the corrected precipitation.
[0026] The temperature data from the measured stations is decoded and interpolated onto the model's latitude and longitude. Temperature-related variables at each altitude level are extracted, including temperature, relative humidity, water vapor density, wind speed and direction, air drag coefficient, friction velocity, convective potential energy, convective suppression coefficient, horizontal water dispersion, and derived radar reflectivity. The extracted data are used as precipitation correction factors. The correction factors and the forecast data for the precipitation area are input into the correction network to perform precipitation regression and output the corrected precipitation data.
[0027] The TS score for precipitation reflects the ability of forecast models to capture precipitation events of different intensities and extents, and is a comprehensive scoring index.
[0028] The calculation of the TS score includes:
[0029] TS = H / (H + M + F)
[0030] Where TS represents the correction effect TS score, H represents the number of grids where both observations and forecasts indicate precipitation, M represents the number of grids where observations indicate precipitation but forecasts indicate no precipitation, and F represents the number of grids where observations indicate no precipitation but forecasts indicate precipitation.
[0031] The effectiveness of precipitation correction in the corrected precipitation area was assessed based on the TS score, including:
[0032] The TS score ranges between 0 and 1. When the predicted precipitation level matches the observed precipitation level, TS = 1, resulting in the best correction effect. The smaller the TS score, the worse the correction effect.
[0033] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A precipitation forecast correction method based on cascaded Attention-U-Net, characterized in that, include: S1: Using precipitation, temperature and time from numerical forecasts as prediction factors, the observed actual precipitation data is divided into grid points with time interval n. S2: Establish a cascaded Attention-U-Net precipitation forecast correction model; the cascaded Attention-U-Net precipitation forecast correction model includes: a precipitation classification network and a correction network; Establish a cascaded Attention-U-Net precipitation forecast correction model, including: Two Attention-U-Net base models are concatenated to obtain a concatenated Attention-U-Net precipitation forecast correction model, where the two Attention-U-Net base models represent the precipitation classification network and the correction network, respectively. The Attention-U-Net basic model includes a downsampling part, an upsampling part, and an attention gate. The downsampling part includes convolutional layers, batch normalization layers, ReLU layers, and max pooling layers. The downsampling part is structurally symmetrical with the upsampling part. Before incorporating the feature maps generated during the downsampling process into the upsampling features, an attention gate is added to perform weighted processing to increase the proportion of task-related regions. S3: Train a precipitation classification network based on precipitation data, and use the trained precipitation classification network to classify the precipitation at grid points; S4: For areas with precipitation, train a correction network using the corresponding actual precipitation data within the area. Use the trained correction network to correct the precipitation in the precipitation area and calculate the TS score to evaluate the precipitation correction effect in the corrected precipitation area.
2. The precipitation forecast correction method based on cascaded Attention-U-Net according to claim 1, characterized in that, A precipitation classification network was trained based on forecast data and precipitation classification data, including: Forecast data and precipitation classification data are used as training labels and input into the precipitation classification network. Focal Loss is used as the loss function of the precipitation classification network for training. When the Focal Loss loss function of the precipitation classification network is minimized, the network parameters are fixed to complete the training of the precipitation classification network.
3. The precipitation forecast correction method based on cascaded Attention-U-Net according to claim 1, characterized in that, The trained precipitation classification network is used to classify gridded precipitation conditions, including: Meteorological elements are input into the trained precipitation classification network to obtain the probability of rain / no rain at each grid point within the grid area. The threshold value is set to 0.
5. Grid points with a probability greater than 0.5 are marked as rainy, and grid points with a probability less than 0.5 are marked as no rainy.
4. The precipitation forecast correction method based on cascaded Attention-U-Net according to claim 1, characterized in that, The correction network was trained using real-time precipitation data, including: The actual precipitation data is used as the training label and input into the correction network. The joint loss of Weighted Loss and Perceptual Loss is used as the loss function of the precipitation classification network for training. When the joint loss function of the precipitation classification network is minimized, the network parameters are fixed and the correction network training is completed.
5. The precipitation forecast correction method based on cascaded Attention-U-Net according to claim 1, characterized in that, The trained correction network is used to correct precipitation in precipitation areas, including: The temperature data from the measured stations is decoded and interpolated onto the model's latitude and longitude. Temperature-related variables at each altitude level are extracted, including temperature, relative humidity, water vapor density, wind speed and direction, air drag coefficient, friction velocity, convective potential energy, convective suppression coefficient, horizontal water dispersion, and derived radar reflectivity. The extracted data are used as precipitation correction factors. The correction factors and the forecast data for the precipitation area are input into the correction network to perform precipitation regression and output the corrected precipitation data.
6. The precipitation forecast correction method based on cascaded Attention-U-Net according to claim 1, characterized in that, The calculation of the TS score includes: in, The TS score indicates the effectiveness of the correction. This indicates the number of grid cells for both observation and forecasting of precipitation. This indicates the number of grid cells where precipitation was observed but no precipitation was predicted. This indicates the number of grid cells where no precipitation was observed but precipitation was predicted.
7. The precipitation forecast correction method based on cascaded Attention-U-Net according to claim 1, characterized in that, The effectiveness of precipitation correction in the corrected precipitation area was assessed based on the TS score, including: The TS score ranges between 0 and 1. When the predicted precipitation level matches the observed precipitation level, TS=1, resulting in the best correction effect. The smaller the TS score, the worse the correction effect.