Att-TS-UKAN rainfall deviation correction method based on conditional dynamic region sampling

Abstract

The invention discloses an Att-TS-UKAN rainfall deviation correction method based on conditional dynamic region sampling. The Att-TS-UKAN rainfall deviation correction method comprises the following steps: (1) acquiring global rainfall forecast, satellite data and related meteorological factor data and carrying out preprocessing; (2) constructing a deep learning model of Att-UKAN; comprising a convolution block, an attention module and a self-attention KAN module; (3) designing a weighted mean square error and differentiable weather TS score based on sample data distribution; (4) training the model by using global rainfall forecast data, related factors and rainfall satellite data through a conditional dynamic region sampling data enhancement technology; (5) performing deviation correction on the rainfall forecast data of the target area by using the optimized model to generate a more accurate rainfall forecast result; according to the invention, the spatial distribution and magnitude prediction of rainfall forecast are optimized.

Description

Technical Field

[0001] The invention relates to the technical field of weather forecasting, and in particular to an Att-TS-UKAN precipitation deviation correction method based on conditional dynamic regional sampling. Background Art

[0002] Precipitation forecast is of great significance to agriculture, disaster prevention and mitigation, urban planning, etc. However, due to the complexity of the precipitation process and the computing power of the numerical weather prediction (NWP) model, it is difficult for traditional methods to achieve high-precision prediction of heavy precipitation events. If you want to obtain more accurate precipitation forecasts, especially heavy precipitation forecasts, it is necessary to correct the deviations through post-processing methods.

[0003] In traditional correction products, numerical models are usually combined with observational data, and the deviations are linked to the long-term statistical differences between forecasts and observations, so as to construct statistical-dynamic forecasting methods to characterize local weather characteristics that are missing in the model, such as using the frequency matching method (FMM) to eliminate frequency deviations at different precipitation levels, and the quantile mapping method (QM) to map the distribution of model output to the distribution of observational data, and correct the deviations in the model output. However, these traditional objective methods have limited ability to process massive meteorological data and are difficult to reflect the impact of spatial information.

[0004] Existing deep learning models such as LSTM (Long Short-Term Memory) and CNN (Convolutional Neural Networks) have demonstrated the potential of deep learning in bias correction. However, since heavy precipitation accounts for a small proportion of the total samples, most correction models trained with such unbalanced samples have good correction performance for small-intensity precipitation, but their performance for extreme precipitation events is still not ideal. Moreover, most models have difficulty capturing local small-scale precipitation, resulting in ambiguous correction effects. Summary of the invention

[0005] Purpose of the invention: The purpose of the present invention is to provide an Att-TS-UKAN precipitation bias correction method based on conditional dynamic regional sampling to solve the problems of insufficient accuracy of existing methods in the prediction of heavy precipitation events and local small-scale precipitation and poor adaptability of the model to extreme precipitation events.

[0006] Technical solution: The Att-TS-UKAN precipitation deviation correction method based on conditional dynamic regional sampling described in the present invention comprises the following steps: (1) Obtain global precipitation forecasts, satellite data, and related meteorological factor data and perform preprocessing; (2) Construct the deep learning model of Att-UKAN, including convolutional blocks, attention modules, and self-attention KAN modules. (3) Design weighted mean square error and differentiable meteorological TS score based on sample data distribution; (4) Using the data augmentation technique of conditional dynamic regional sampling to train the model using global precipitation forecast data and related factors and precipitation satellite data; (5) Use the optimized model to correct the deviation of precipitation forecast data in the target area and generate more accurate precipitation forecast results.

[0007] Furthermore, the factor data include humidity, wind speed, and terrain; the preprocessing is as follows: the spatial resolution of the data is unified to 0.25°, and the temporal resolution is unified to 6 hours.

[0008] Furthermore, step (2) of the Att-UKAN deep learning model is as follows: the input image is first processed by the encoder, and the first three layers of the encoder extract features through convolution operations, and then input into two tokenized self-attention KAN modules KAN Block; each encoder block reduces the resolution of the feature map by half; the decoder contains two tokenized self-attention KAN modules and three convolution blocks; each decoder block doubles its resolution; information fusion is achieved between the encoder and the decoder through jump connections; in addition, the decoder part also connects the jump connection information of the previous decoder block with the corresponding encoder block through the attention module to capture precipitation features of different scales.

[0009] Furthermore, step (3) is as follows: the weighted mean square error (WMSE) is used to measure the overall deviation between the predicted value and the observed value, and weights are set for the MSE of different pixels; the meteorological TS score is used to optimize the prediction ability of heavy precipitation; the weighted mean square error (WMSE) formula is as follows: ; ; in, represents the i-th predicted value, represents the i-th true observation, N represents the number of samples, Indicates the weights set for pixels of different values ​​based on the sample data distribution; The meteorological TS scoring formula is as follows: ; ; ; ; in, represents the i-th predicted value, represents the i-th true observation, N represents the number of samples, and thld represents the set precipitation threshold.

[0010] Furthermore, step (4) is as follows: the global precipitation forecast and satellite data are randomly selected to sample the target area size, and when the precipitation value of the selected satellite data sample satisfies that the number of pixels with precipitation values ​​exceeding 50 accounts for more than 1% of the total number of pixels in the sampling area, it is returned for model training; during training normalization, the problem of data imbalance between samples is reduced by logarithmic transformation, and the specific transformation formula is as follows: ; ; Where x represents the original forecast and satellite precipitation data, is a hyperparameter, set to 0.0001, This is the normalized data that is finally input into the model training.

[0011] Furthermore, step (5) is as follows: three scoring indices, namely, hit rate POD, false alarm rate FAR and Heidke SkillScore HSS, are introduced to conduct a deterministic evaluation of the precipitation bias correction capability. The three scoring indices are calculated based on the binary confusion matrix, and the calculation formula is as follows: ; ; ; Among them, TP means that the observation is positive and the prediction is positive; FN means that the observation is positive and the prediction is negative; FP means that the observation is negative and the prediction is positive; TN means that the observation is negative and the prediction is negative.

[0012] The Att-TS-UKAN precipitation deviation correction system based on conditional dynamic regional sampling described in the present invention comprises: Preprocessing module: used to obtain global precipitation forecasts, satellite data and related meteorological factor data and perform preprocessing; Att-UKAN module: used to build the deep learning model of Att-UKAN; including: convolutional blocks, attention modules and self-attention KAN modules; Design module: used to design weighted mean square error and differentiable meteorological TS score based on sample data distribution; Enhancement module: used to train the model using global precipitation forecast data and related factors and precipitation satellite data through data enhancement technology of conditional dynamic area sampling; Prediction module: used to use the optimized model to correct the deviation of precipitation forecast data in the target area and generate more accurate precipitation forecast results.

[0013] An electronic device described in the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded into the processor, the method for correcting Att-TS-UKAN precipitation deviation based on conditional dynamic area sampling according to any one of the items is implemented.

[0014] A storage medium described in the present invention stores a computer program, and when the computer program is executed by a processor, it implements any one of the Att-TS-UKAN precipitation deviation correction methods based on conditional dynamic area sampling.

[0015] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: the data of conditional dynamic area sampling is input into the Att-TS-UKAN model, and the spatial distribution and magnitude prediction of precipitation forecast are optimized by introducing the attention module and the self-attention KAN module and combining the loss function of TS score. The experimental results show that the forecasting performance of the present invention under multiple precipitation thresholds is significantly better than that of the traditional method. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 is a flow chart of the present invention; Figure 2 It is a schematic diagram of the structure of the deep learning model Att-UKAN of the present invention; Figure 3 It is a probability density distribution diagram under different data normalization modes of the present invention; Figure 4 It is a comparison chart of the correction results of the present invention and the scores of other methods. DETAILED DESCRIPTION

[0017] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0018] The embodiment of the present invention provides an Att-TS-UKAN precipitation deviation correction method based on conditional dynamic regional sampling. The research area of ​​the embodiment of the present invention is the Chinese region. The region includes 15°N to 55°N and 70°E to 134°E. Figure 1 As shown, the following steps are included: Step 1: Data acquisition and training and test set division; The forecast precipitation data used in the embodiment of the present invention is the daily precipitation of the Global Forecast System (GFS) of the National Center for Environmental Prediction (NCEP) of the United States, with a spatial resolution of 0.25° and a temporal resolution of 6 hours. The satellite (observation) precipitation data is the Level 3 product IMERG of the NASA Global Precipitation Observation (GPM) program, with a spatial resolution of 0.1° and a temporal resolution of 0.5 hours, which is unified to the resolution of the forecast precipitation data through the upscaling method. The relevant meteorological factor data include the global terrain relief data (earth relief) provided by GMT and the ERA5 reanalysis data of the European Center for Medium-Term Weather Forecasts (ECMWF), including meteorological element products such as humidity and wind speed. The sample data covers from June 2019 to September 2023, and is divided into training set, validation set and test set according to the ratio of 80%, 10% and 10%.

[0019] Step 2: Build a deep learning model based on Att-UKAN, and use convolutional blocks, attention modules, and self-attention KAN modules to optimize the encoder and decoder structures to capture precipitation features at different scales; Figure 2 As shown, the Att-UKAN deep learning model architecture in the present invention adopts an encoder-decoder architecture based on U-Net, combining convolutional layers, self-attention KAN modules (KAN Block) and attention modules (Attention Block), aiming to capture precipitation features of different scales through multi-level feature extraction and fusion.

[0020] The encoder on the left side consists of two parts. First, three convolutional layers are used to extract different features of the input image in sequence. The specific composition and operation of each convolution block are as follows: use a 3*3 convolution kernel with a stride and padding of 1 to perform convolution operations; use a 2*2 pooling window for maximum pooling downsampling to reduce the image size. Through the feature extraction of these convolutional layers, local features and detail information in the precipitation forecast image can be effectively captured, helping the model to identify potential sources of precipitation deviations, such as local climate patterns or regional weather changes. These low-level features will provide important basic information for subsequent deviation corrections, ensuring that the detailed features of the input data are fully retained during the correction process.

[0021] After the three-layer convolution operation, the image enters the two-layer labeled self-attention KAN module (KAN Block). The specific composition and operation include: cutting and reshaping the image into 3*3 small image blocks (patches) with sequence features, inputting several patches with position labels into the three-layer KAN layer, and simulating complex high-dimensional function mapping through learnable activation functions. This module can capture the global dependencies and regional patterns in precipitation data. In this way, the KAN module can not only learn local features, but also help the model identify possible regional differences in precipitation deviations. Next, a separable convolution layer (DwConv) is connected to the original labeled patch layer residual, and the output features are passed to the next module through the normalization layer, ensuring that the local and global features of precipitation deviation can be balanced and optimized during the model training process, reducing the transmission and accumulation of deviations. Each encoder block halves the resolution of the feature map. After the five-layer encoder structure, the image resolution is reduced to 1 / 32 of the original input.

[0022] The decoder on the right corresponds to the encoder structure, including two tokenized self-attention KAN modules and three convolution blocks, which restore the image resolution and size layer by layer through maximum pooling upsampling. In this process, the decoder can accurately restore the local details and spatial structure in the precipitation forecast, ensuring that the forecast results after deviation correction are consistent with the actual precipitation distribution.

[0023] In addition, information fusion is achieved between the encoder and decoder of each layer through jump connections. In particular, the right encoder part connects the jump connection information of the decoder block of the previous layer with the encoder block corresponding to this layer through the attention module (Attention Block), so as to achieve cross-level feature integration. Through this information fusion, the model can dynamically adjust the feature weights and effectively combine feature information at different levels to further improve the correction ability of precipitation deviations and ensure global consistency and local accuracy in the deviation correction process, that is, it can maintain the model's accurate prediction of large-scale precipitation patterns and effectively adjust regional deviations.

[0024] Step 3: Design the weighted mean square error (WMSE) and differentiable meteorological TS score (DTS) based on the sample data distribution to enhance the model's prediction performance for heavy precipitation events; adopt a combination of weighted mean square error (WMSE) and differentiable TS score (DTS).

[0025] It is used to measure the overall deviation between the predicted value and the observed value. Different weights are set for pixels with different values ​​to increase the importance of heavy precipitation samples in the model training process. The specific calculation formula is as follows: ; ; in, represents the i-th predicted value, represents the i-th true observation, N represents the number of samples, Indicates the weights set for pixels of different values ​​based on the sample data distribution; The meteorological TS score (Threat Score) is introduced and converted into a differentiable form through the sigmoid function to optimize the prediction ability of heavy precipitation. The calculation formula is as follows: ; ; ; ; in, represents the i-th predicted value, represents the i-th true observation, N represents the number of samples, and thld represents the set precipitation threshold.

[0026] Specifically, the thresholds set in the present invention are 0.1, 10, 25, 50, and 100. and Each threshold adopts the strategy of automatically adjusting weights during training.

[0027] Step 4: Use the global precipitation forecast data and related factors and precipitation satellite data to train the target area model through the data augmentation technology of conditional dynamic regional sampling; Since the total number of samples in the data set is 6155, the sample size is relatively limited for the deep learning model. In addition, the data set contains precipitation forecasts and satellite data for the global range (60°N to 60°S, 180°E to 180°W), with a size of 480×1440, while the size of the target area (15°N to 55°N, 70°E to 134°E) is 160×256. Therefore, the present invention adopts a random sampling method for global data, and effectively expands the number of training samples by randomly selecting areas of different sizes in the target area for sampling. In this way, the present invention expands the number of samples input to the model to 16 times the original sample size, achieving the best effect. In addition, the present invention also performs conditional screening on the precipitation data in the sample, specifically: only when the number of pixels with precipitation values ​​exceeding 50 in the sample accounts for more than 1% of the total number of pixels in the sampling area, the sample is returned for model training. This strategy ensures that the data input to the model contains enough valid precipitation samples, thereby improving the quality and effect of model training.

[0028] On the other hand, Figure 3 As shown, the simplest approach of normalizing by dividing by the maximum value results in a density distribution of the data input to the model ( Figure 3 The blue line) is extremely unbalanced. This imbalance is mainly due to the fact that the majority of samples have no precipitation or little precipitation, while samples with moderate rain or above (i.e., precipitation greater than 10 mm / day) account for only 8.22% (as shown in Table 1). This distribution causes the deep learning model to tend to optimize the performance of a large number of samples with no precipitation or little precipitation during training, while ignoring a small number of samples with heavy precipitation, ultimately achieving better average performance overall, but with poor prediction results for heavy precipitation samples. Therefore, the present invention first reduces the data imbalance problem between samples through logarithmic transformation, and the specific transformation formula is as follows: ; Where x represents the original forecast and satellite precipitation data, is a hyperparameter, set to 0.0001, This is the normalized data that is finally input into the model training.

[0029] Next, the transformed data is normalized: ; The probability density distribution after logarithmic transformation ( Figure 3 Green line) has a more reasonable distribution compared to the ordinary maximum normalization. On this basis, the present invention sets different weights for pixels with different precipitation values ​​when calculating the loss function ( Figure 3 Gray line), further enhancing the influence of heavy precipitation samples on the training effect, thereby improving the model’s prediction ability for heavy precipitation events.

[0030] Table 1 Distribution ratios of different rainfall amounts in the test set data ; Step 5: Use the optimized model to correct the deviation of precipitation forecast data in the target area to generate more accurate precipitation forecast results.

[0031] The present invention introduces three scoring indices, namely, hit rate (POD), false alarm rate (FAR) and Heidke Skill Score (HSS), to perform deterministic evaluation on its correction ability. The three scoring indices are calculated based on the binary confusion matrix, and the calculation formula is as follows: ; ; ; Among them, TP means that the observation is positive and the prediction is positive; FN means that the observation is positive and the prediction is negative; FP means that the observation is negative and the prediction is positive; TN means that the observation is negative and the prediction is negative.

[0032] like Figure 4 As shown in the bar graphs of each model score, the revised results of the present invention ( Figure 4 blue bars) and other methods ( Figure 4 Yellow and green histograms) and the original forecast ( Figure 4 Gray bar graph) shows that this method can improve the overall scoring performance at each threshold, which is specifically manifested in the improvement of HSS score ( Figure 4 Left), hit rate improved ( Figure 4 ), false alarm rate reduction ( Figure 4 Right). Compared with the original forecast, the HSS score can be improved by 17%-23% at the thresholds of 0.1, 10, 25, 50, and 100 mm / day; in particular, at the large thresholds of 50 and 100 (heavy rain and above), the method of the present invention is significantly better than the traditional U-Net method and the U-Net method with attention mechanism, which shows that the combination of the self-attention KAN module and the attention module can better capture large threshold precipitation and improve the hit rate of heavy precipitation forecast.

[0033] The embodiment of the present invention also provides an Att-TS-UKAN precipitation deviation correction system based on conditional dynamic regional sampling, including: Preprocessing module: used to obtain global precipitation forecasts, satellite data and related meteorological factor data and perform preprocessing; Att-UKAN module: used to build the deep learning model of Att-UKAN; including: convolutional blocks, attention modules and self-attention KAN modules; Design module: used to design weighted mean square error and differentiable meteorological TS score based on sample data distribution; Enhancement module: used to train the model using global precipitation forecast data and related factors and precipitation satellite data through data enhancement technology of conditional dynamic area sampling; Prediction module: used to use the optimized model to correct the deviation of precipitation forecast data in the target area and generate more accurate precipitation forecast results.

[0034] An embodiment of the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the computer program is loaded into the processor, the method for correcting Att-TS-UKAN precipitation deviation based on conditional dynamic area sampling according to any one of the items is implemented.

[0035] An embodiment of the present invention also provides a storage medium storing a computer program, and when the computer program is executed by a processor, it implements any one of the Att-TS-UKAN precipitation deviation correction methods based on conditional dynamic area sampling.

Claims

1. A precipitation bias correction method based on conditional dynamic regional sampling based on Att-TS-UKAN, characterized in that: The following steps are involved: (1) Obtain global precipitation forecasts, satellite data, and related meteorological factor data and perform preprocessing; (2) Construct the deep learning model of Att-UKAN, including convolutional blocks, attention modules, and self-attention KAN modules. (3) Design weighted mean square error and differentiable meteorological TS score based on sample data distribution; (4) Using the data augmentation technique of conditional dynamic regional sampling to train the model using global precipitation forecast data and related factors and precipitation satellite data; (5) Use the optimized model to correct the deviation of precipitation forecast data in the target area and generate more accurate precipitation forecast results.

2. The Att-TS-UKAN precipitation deviation correction method based on conditional dynamic regional sampling according to claim 1 is characterized in that: The factor data include humidity, wind speed, and terrain; the preprocessing is as follows: the spatial resolution of the data is unified to 0.25°, and the temporal resolution is unified to 6 hours.

3. The Att-TS-UKAN precipitation deviation correction method based on conditional dynamic regional sampling according to claim 1 is characterized in that: Step (2) of the Att-UKAN deep learning model is as follows: the input image is first processed by an encoder, the first three layers of the encoder extract features through convolution operations, and then input into two tokenized self-attention KAN blocks; each encoder block reduces the resolution of the feature map by half; The decoder contains two tokenized self-attention KAN modules and three convolution blocks; each decoder block doubles its resolution; information fusion is achieved between the encoder and decoder through jump connections; in addition, the decoder part also connects the jump connection information of the previous decoder block with the corresponding encoder block through the attention module to capture precipitation characteristics of different scales.

4. The Att-TS-UKAN precipitation deviation correction method based on conditional dynamic regional sampling according to claim 1 is characterized in that: Step (3) is as follows: The weighted mean square error (WMSE) is used to measure the overall deviation between the predicted value and the observed value, and weights are set for the MSE of different pixels; the meteorological TS score is used to optimize the prediction ability of heavy precipitation; the weighted mean square error (WMSE) formula is as follows: ； ； in, represents the i-th predicted value, represents the i-th true observation, N represents the number of samples, Indicates the weights set for pixels of different values ​​based on the sample data distribution; The meteorological TS scoring formula is as follows: ； ； ； ； in, represents the i-th predicted value, represents the i-th true observation, N represents the number of samples, and thld represents the set precipitation threshold.

5. The Att-TS-UKAN precipitation deviation correction method based on conditional dynamic regional sampling according to claim 1 is characterized in that: Step (4) is as follows: the global precipitation forecast and satellite data are randomly sampled in a target area. When the precipitation value of the selected satellite data sample satisfies that the number of pixels with a precipitation value exceeding 50 accounts for more than 1% of the total number of pixels in the sampling area, it is returned for model training. During training normalization, the problem of data imbalance between samples is reduced through logarithmic transformation. The specific transformation formula is as follows: ； ； Where x represents the original forecast and satellite precipitation data, is a hyperparameter, This is the normalized data that is finally input into the model training.

6. The Att-TS-UKAN precipitation deviation correction method based on conditional dynamic regional sampling according to claim 1 is characterized in that: Step (5) is as follows: three scoring indices, namely, hit rate POD, false alarm rate FAR and Heidke SkillScore HSS, are introduced to conduct a deterministic evaluation of the precipitation bias correction capability. The three scoring indices are calculated based on the binary confusion matrix. The calculation formula is as follows: ； ； ； Among them, TP means that the observation is positive and the prediction is positive; FN means that the observation is positive and the prediction is negative; FP means that the observation is negative and the prediction is positive; TN means that the observation is negative and the prediction is negative.

7. An Att-TS-UKAN precipitation deviation correction system based on conditional dynamic regional sampling, characterized in that: include: Preprocessing module: used to obtain global precipitation forecasts, satellite data and related meteorological factor data and perform preprocessing; Att-UKAN module: used to build the deep learning model of Att-UKAN; including: convolutional blocks, attention modules and self-attention KAN modules; Design module: used to design weighted mean square error and differentiable meteorological TS score based on sample data distribution; Enhancement module: used to train the model using global precipitation forecast data and related factors and precipitation satellite data through data enhancement technology of conditional dynamic area sampling; Prediction module: used to use the optimized model to correct the deviation of precipitation forecast data in the target area and generate more accurate precipitation forecast results.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the computer program is loaded into a processor, it implements the Att-TS-UKAN precipitation bias correction method based on conditional dynamic regional sampling according to any one of claims 1 to 6.

9. A storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, it implements the Att-TS-UKAN precipitation deviation correction method based on conditional dynamic area sampling according to any one of claims 1 to 6.

Images