A la-unet-lstm-based sub-seasonal precipitation prediction method

Abstract

The application discloses a LA-UNET-LSTM-based subseasonal precipitation prediction method, which comprises the following steps: collecting precipitation and multi-meteorological element prediction data output by a numerical model and actual precipitation observation data, and dividing the data into a training set, a verification set and a test set; extracting a feature sequence of a precipitation predictable mode based on the training set data, diagnosing and analyzing the predictability source of the feature sequence in the numerical model, and extracting a feature mask field; building a LA-UNET-LSTM neural network, and constructing a loss function based on structural similarity and weighted mean square deviation; performing standardization processing on the data, cooperating with the feature mask field to construct a prediction factor feature map, training the model based on the training set data, and adjusting model parameters according to the performance in the model verification set; and bringing the prediction factor feature map in the test set into the trained model, and obtaining precipitation prediction data for 1-4 weeks in the future. The application effectively improves the calculation efficiency and the prediction skill of extreme precipitation.

Description

Technical Field

[0001] This invention relates to a method for forecasting subseasonal precipitation, specifically a method based on LA-UNET-LSTM. Background Technology

[0002] With a deeper understanding of atmospheric physical processes and advancements in observation technology and computer resources, short- and medium-term weather forecasts (within 2 weeks) and short-term climate predictions (more than 3 months) have made significant progress. However, sub-seasonal forecasts, which fall between the two, are still in the early stages of development, and current forecasting techniques are very limited, requiring further exploration and improvement.

[0003] Currently, the precipitation forecast skill of numerical models is mostly limited to within two weeks, making it difficult to directly guide disaster prevention and mitigation. In recent years, some studies have used statistical post-processing methods to correct biases in precipitation forecasts output by numerical models, such as Bayesian averaging, modal projection, and quantile mapping. These methods have improved the subseasonal precipitation skill to some extent, but due to the complexity of precipitation processes, traditional statistical post-processing methods struggle to capture the nonlinear characteristics of the atmosphere, thus limiting their improvement in forecast skill. Against this backdrop, some studies have used deep learning methods to correct subseasonal precipitation forecasts, showing that deep learning methods have significant advantages over traditional statistical methods. However, because the signal-to-noise ratio of numerical models decreases rapidly with increasing forecast lead time, even deep learning methods struggle to significantly improve the 3-4 week precipitation forecast skill. This implies a need to explore additional sources of predictability to further enhance forecast skill. Furthermore, current deep learning solutions mostly train models per forecast lead time, resulting in high computational costs and hindering operationalization. On the other hand, deep learning methods are generally good at capturing average states, but their forecasting skills for extreme values ​​are relatively poor, which also leads to their limited performance in forecasting extreme precipitation. Summary of the Invention

[0004] Purpose of the invention: The purpose of this invention is to provide a subseasonal precipitation forecasting method based on LA-UNET-LSTM, so as to effectively improve computational efficiency and forecasting skills for extreme precipitation.

[0005] Technical solution: The present invention provides a subseasonal precipitation forecasting method based on LA-UNET-LSTM, comprising the following steps: (1) Collect precipitation and multi-meteorological element forecast data output by numerical model, as well as actual precipitation observation data, and divide the precipitation and multi-meteorological element forecast data output by numerical model, as well as actual precipitation observation data, into training set, validation set and test set respectively. (2) Based on the training set data, extract the feature sequences of the predictable precipitation modes, diagnose and analyze the sources of their predictability in the numerical model, and extract the feature mask field; (3) Construct an LA-UNET-LSTM neural network model and build a loss function based on structural similarity and weighted mean square error; the LA-UNET-LSTM neural network includes multiple UNET modules and one LSTM module. Multiple UNET modules are used to extract features from the feature maps of each forecast factor, and then the LSTM module is used to fuse the features. (4) Standardize the training set, validation set and test set data, and construct the forecast factor feature maps of the training set, validation set and test set respectively with the feature mask field. Use the forecast factor feature maps in the training set and the actual precipitation observation data to train the LA-UNET-LSTM neural network model, and use the forecast factor feature maps in the validation set and the actual precipitation observation data to evaluate the model performance and adjust the model parameters, so as to obtain the trained LA-UNET-LSTM neural network model. (5) Input the forecast factor feature map from the test set into the trained model and obtain the precipitation forecast data for the next 1-4 weeks.

[0006] Furthermore, the multi-meteorological element forecast data in step (1) includes: the sea surface temperature field, multi-layer geopotential height field, and multi-layer wind field forecasted by the model. The forecast data all include multiple ensemble members.

[0007] Furthermore, in step (2), feature sequences of predictable precipitation modes are extracted based on the training set data, specifically as follows: First, the residual value of the precipitation forecast is calculated based on the precipitation ensemble forecast data in the training set. The calculation formula is as follows: ; in, These are the forecast values ​​for the precipitation ensemble members. It is the forecast average of the precipitation ensemble members. This indicates the residual in the precipitation forecast; for The noise modes are obtained by empirical orthogonal decomposition of the covariance matrix: ; in, express The covariance matrix, For noise mode, It is by A diagonal matrix composed of eigenvalues. for The transpose of the matrix; then... Projected to That is, to achieve white noise conversion: ; in, After white noise conversion , For the number of members in the precipitation ensemble; then for Singular value decomposition yields the characteristic sequence of predictable modes: ; in, For left singular values, for A diagonal matrix composed of eigenvalues. These are the characteristic sequences of predictable modes.

[0008] Furthermore, in step (2), the predictability sources in the numerical model are diagnosed and analyzed, and the feature mask field is extracted; specifically: Correlation analysis was conducted based on the characteristic sequences of predictable precipitation modes and forecast data of multiple meteorological elements, and the correlation coefficient was determined. The calculation formula is: ; in, The characteristic sequence representing the predictable mode is the first... One element, This represents the average value of the characteristic sequence of the predictable mode. For meteorological elements at a certain grid point One value, This represents the average value of meteorological elements at a certain grid point. The length of the characteristic sequence of the predictable mode; A significance test is performed, and a feature mask field is obtained based on the significance test results. Regions that pass the test are marked as 1, and regions that fail the test are marked as 0.

[0009] Furthermore, in step (3), a loss function based on structural similarity and weighted mean square error is constructed; specifically: The formula for calculating structural similarity (SSIM) is: , , ; in, and These represent forecast data and observational data, respectively. and These represent the mean values ​​of the forecast data and the observed data, respectively. and These represent the standard deviations of the forecast data and the observed data, respectively. This represents the covariance between forecast data and observed data; and They are respectively and The parameters, The maximum value of the observed data and the forecast data; The formula for calculating the weighted mean square error (WMSE) is: ; in, and Representing the first One forecast and one observation sample, Indicates the total number of samples; The loss function is: ; in, and These are two hyperparameters that are adjusted based on the actual training conditions to obtain the optimal model prediction performance.

[0010] Furthermore, in step (4), the training set, validation set, and test set data are standardized, and feature maps of the prediction factors for the training set, validation set, and test set are constructed in conjunction with the feature mask field; including: The standardization process is as follows: First, outliers in the observed and forecast data are removed. Then, the data is processed into weekly average data, and the weekly average anomaly is extracted based on the average value of historical data. The calculation formula is as follows: ; in, It is the average value of the observed / forecasted variable over a certain week. It is the average value of the variable over the same historical period; This is the weekly average anomaly of the variable; after extracting the anomaly, the variable needs to be normalized to its minimum and maximum values. ; in, It is the weekly average anomaly of a certain variable. express The minimum value, express The maximum value, This is the result of normalizing the maximum and minimum weekly average anomalies of the variable; Multiply the forecast data in the standardized training set, validation set, and test set by the corresponding feature mask field obtained in step (2) to obtain the forecast factor feature maps of the training set, validation set, and test set, respectively.

[0011] Furthermore, in step (4), the model performance is evaluated and the model parameters are adjusted using the forecast factor feature map and actual precipitation observation data from the validation set; specifically: The structural similarity (SSIM) and weighted mean square error (WMSE) in the model were evaluated using the forecast factor feature map and actual precipitation observation data in the validation set. If the WMSE generated by the model in the validation set is greater than the WMSE of the original model, it indicates that the model has poor forecasting ability for extreme precipitation and the hyperparameters need to be increased. Value; if the SSIM of the model-generated forecast in the validation set is less than the SSIM of the original model forecast, it indicates that the model has a poor grasp of the rainband structure and the hyperparameters need to be increased. value.

[0012] In one embodiment of the present invention, a subseasonal precipitation forecasting system based on LA-UNET-LSTM includes: The dataset creation unit is used to collect precipitation and multi-meteorological element forecast data output by numerical models, as well as actual precipitation observation data. The precipitation and multi-meteorological element forecast data output by numerical models, as well as actual precipitation observation data, are divided into training set, validation set and test set respectively. The feature map construction unit is used to extract feature sequences of predictable precipitation modes based on training set data, diagnose and analyze the sources of their predictability in numerical models, and extract feature mask fields. The network model building unit is used to build the LA-UNET-LSTM neural network model and construct a loss function based on structural similarity and weighted mean square error. The LA-UNET-LSTM neural network includes multiple UNET modules and one LSTM module. Multiple UNET modules are used to extract features from the feature maps of each forecast factor, and then the LSTM module is used to fuse the features. The model training and validation unit is used to standardize the training, validation, and test data, and construct the forecast factor feature maps of the training, validation, and test sets respectively with the feature mask field. The LA-UNET-LSTM neural network model is trained using the forecast factor feature maps in the training set and the actual precipitation observation data. The model performance is evaluated and the model parameters are adjusted using the forecast factor feature maps in the validation set and the actual precipitation observation data, thus obtaining the trained LA-UNET-LSTM neural network model. The test unit is used to input the feature maps of forecast factors from the test set into the trained model, and at the same time obtain precipitation forecast data for the next 1-4 weeks.

[0013] In one embodiment of the present invention, an electronic device includes: Memory containing executable program code; A processor coupled to the memory; The processor calls the executable program code stored in the memory to execute the steps of the subseasonal precipitation forecasting method based on LA-UNET-LSTM as described above.

[0014] In one embodiment of the present invention, a computer-readable storage medium stores computer instructions that, when invoked, are used to perform the steps of the LA-UNET-LSTM-based subseasonal precipitation forecasting method described above.

[0015] Beneficial effects: Compared with the prior art, the significant technical effects of the present invention are as follows: (1) Based on predictable modal analysis, predictability sources are extracted from geopotential height, sea surface temperature and wind field forecast products of numerical models, providing physical constraints and additional opportunity windows for the next season precipitation forecast; (2) A neural network of LA-UNET-LSTM is proposed, which extracts feature modes of each forecast factor and performs feature fusion, and generates precipitation forecast products for the next 1-4 weeks, effectively improving computational efficiency and forecasting skills; (3) A loss function based on SSIM and MSE is introduced, which effectively improves the model's grasp of the rain belt location and improves the forecasting skills of extreme precipitation. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a schematic diagram of the network model structure. Detailed Implementation

[0017] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0018] like Figure 1 As shown, the subseasonal precipitation forecasting method based on LA-UNET-LSTM provided by this invention includes the following steps: (1) Collect precipitation and multi-meteorological element forecast data and actual precipitation observation data output by the numerical model, and divide the precipitation and multi-meteorological element forecast data and actual precipitation observation data into training set, validation set and test set respectively. Among them, the multi-meteorological element forecast data includes: the sea surface temperature field, multi-layer geopotential height field and multi-layer wind field forecast by the model. All forecast data include multiple ensemble members, that is, the forecast data of each meteorological element includes multiple ensemble members. For example, for a certain day's forecast, there are 10 forecast data for the sea surface temperature field, multi-layer geopotential height field and multi-layer wind field, that is, there are 10 ensemble members. The precipitation and multi-meteorological element forecast data and actual precipitation observation data output by the numerical model are divided into training set, validation set and test set according to the proportions of 80%, 10% and 10% respectively.

[0019] (2) Based on the training set data (including the training set of precipitation and multi-meteorological element forecast data output by numerical models and the training set of actual precipitation observation data), the feature sequences of the predictable precipitation modes are extracted, and their predictability sources in the numerical models are diagnosed and analyzed, thereby constructing a forecast factor feature map. The specific process is as follows: (21) Based on precipitation forecast data in the training set, the predictable modes of precipitation are extracted using the maximum signal-to-noise ratio empirical orthogonal decomposition method. First, the internal variability of the precipitation forecast data needs to be converted into white noise; First, the residual value of the precipitation forecast is calculated based on the precipitation ensemble forecast data in the training set. The calculation formula is as follows: ; in, These are the forecast values ​​for the precipitation ensemble members. It is the forecast average of the precipitation ensemble members. This indicates the residual in the precipitation forecast. The noise modes can be obtained by empirical orthogonal decomposition of the covariance matrix: ; in, express The covariance matrix, For noise mode, It is by A diagonal matrix composed of eigenvalues. To represent the transpose of a matrix, for The transpose of the matrix. Then... Projected to This allows for white noise conversion: ; in, After white noise conversion , This represents the number of members in the precipitation collection. Next, regarding... Singular value decomposition can be used to obtain the feature sequence of predictable modes: ; in, For left singular values, for A diagonal matrix composed of eigenvalues. These are the characteristic sequences of predictable modes.

[0020] (22) Based on the characteristic sequences of the precipitation predictable mode and the forecast data of multiple meteorological elements, correlation analysis was performed to extract the sources of predictability in the numerical model, and then a forecast factor feature map was constructed. The correlation coefficient calculation formula is as follows: ; in, The characteristic sequence representing the predictable mode is the first... One element, This represents the average value of the characteristic sequence of the predictable mode. For meteorological elements (i.e., sea surface temperature / geopotential height / wind field) at a certain grid point One value, This represents the average value of meteorological elements at a certain grid point. The length of the feature sequence of the predictable mode is given. After calculating the correlation coefficient, a significance test is performed, and the feature mask field is obtained based on the significance test results. The over-detected region is marked as 1, and the under-detected region is marked as 0.

[0021] (3) Construct the LA-UNET-LSTM neural network model and implement a loss function based on structural similarity and weighted mean square error. The LA-UNET-LSTM neural network structure is as follows: Figure 2 As shown, the model includes multiple UNET modules and one LSTM module. The number of UNET modules is the same as the number of forecast factor feature maps. Using the forecast factor feature maps as model input, the UNET modules extract features from the feature maps of each variable and each forecast lead time separately, then fuse them, and perform time-series prediction based on LSTM, thereby simultaneously obtaining precipitation forecast data for the next 1-4 weeks. This invention chooses to extract features from each feature map separately before fusing them, rather than extracting features from all feature maps simultaneously. This strategy is called late fusion (LA), which effectively avoids mutual interference between different forecast factors, allowing the model to maximize the use of information from all forecast factors.

[0022] It is worth noting that the loss function can significantly influence the learning ability of a deep learning model, and a suitable loss function plays a crucial role in improving the performance of a deep learning model. Commonly used loss functions mainly include mean absolute error and mean square error, but these loss functions can lead to a severe "smoothing" effect, causing the model to focus excessively on the average predicted state, which is detrimental to the prediction of extreme cases and system structures. Against this backdrop, this invention introduces a loss function combining Structure Similarity Index Measure (SSIM) and Weighted Mean Square Error (WMSE) to improve the model's prediction skills for rainband structure and extreme precipitation. The SSIM calculation formula is as follows: ; ; ; in, and These represent forecast data and observational data, respectively. and These represent the mean values ​​of the forecast data and the observed data, respectively. and These represent the standard deviations of the forecast and the observation, respectively. This represents the covariance between the two. Parameter and The default values ​​are 0.01 and 0.03. This represents the maximum value of the observed and forecast data. The introduction of SSIM can effectively improve the model's ability to predict rainband features; a larger SSIM indicates a greater similarity between the forecast and observation results.

[0023] On the other hand, to further improve the prediction ability of deep learning models for extreme cases, this invention also introduces weighted mean square error to enhance the model's sensitivity to extreme values. The calculation formula is as follows: ; in, and Representing the first One forecast and one observation sample, This represents the total number of samples. Combining SSIM and WMSE can simultaneously improve the model's predictive skill for both system structure and extreme cases. Therefore, the loss function defined in this invention is: ; in, and These are two hyperparameters, namely the weight coefficients of SSIM and WMSE, which can be adjusted according to the actual training situation to obtain the optimal model prediction performance.

[0024] (4) Standardize the training, validation, and test sets of data divided in step (1), and construct the forecast factor feature maps of the training, validation, and test sets respectively with the feature mask field. Use the forecast factor feature maps in the training set and the actual precipitation observation data to train the LA-UNET-LSTM neural network model, and use the forecast factor feature maps in the validation set and the actual precipitation observation data to evaluate the model performance and adjust the model parameters, thereby obtaining the trained LA-UNET-LSTM neural network model. The standardization process is as follows: First, remove outliers from the observation and forecast data, then process the data into weekly average data, and extract the weekly average anomaly based on the average value of the historical 20-year data. The calculation formula is as follows: ; in, It is the average value of the observed / forecasted variable over a certain week. It is the average value of this variable over the same period in the past 20 years. This represents the weekly average anomaly of the variable. After extracting the anomaly, the variable needs to be normalized to its minimum and maximum values. ; in, It is the weekly average anomaly of a certain variable. express The minimum value, express The maximum value, This is the result of normalizing the maximum and minimum weekly average anomalies of the variable.

[0025] Then, the forecast data in the standardized training set, validation set, and test set are multiplied by the corresponding feature mask fields obtained in step (2) to obtain the forecast factor feature maps of the training set, validation set, and test set, respectively. The LA-UNET-LSTM neural network model is trained using the forecast factor feature maps in the training set and the standardized actual precipitation observation data, and the model performance is evaluated (evaluating SSIM and WMSE in the validation set) and the model parameters are adjusted (adjusting hyperparameters) using the forecast factor feature maps in the validation set and the standardized actual precipitation observation data. and (The default values ​​are all 0.5), thus obtaining the trained LA-UNET-LSTM neural network model; if the WMSE of the model's generated forecast in the validation set is greater than the WMSE of the original model forecast, it indicates that the model has poor ability to predict extreme precipitation and the hyperparameters need to be increased. If the SSIM of the model-generated forecast in the validation set is lower than that of the original model forecast, it indicates that the model has a poor grasp of the rainband structure and the hyperparameters need to be increased. value.

[0026] (5) Input the forecast factor feature map from the test set into the trained model and obtain the precipitation forecast data for the next 1-4 weeks.

[0027] In another embodiment of the present invention, a subseasonal precipitation forecasting system based on LA-UNET-LSTM includes: The dataset creation unit is used to collect precipitation and multi-meteorological element forecast data output by numerical models, as well as actual precipitation observation data. The precipitation and multi-meteorological element forecast data output by numerical models, as well as actual precipitation observation data, are divided into training set, validation set and test set respectively. The feature map construction unit is used to extract feature sequences of predictable precipitation modes based on training set data, diagnose and analyze the sources of their predictability in numerical models, and extract feature mask fields. The network model building unit is used to build the LA-UNET-LSTM neural network model and construct a loss function based on structural similarity and weighted mean square error. The LA-UNET-LSTM neural network includes multiple UNET modules and one LSTM module. Multiple UNET modules are used to extract features from the feature maps of each forecast factor, and then the LSTM module is used to fuse the features. The model training and validation unit is used to standardize the training, validation, and test data, and construct the forecast factor feature maps of the training, validation, and test sets respectively with the feature mask field. The LA-UNET-LSTM neural network model is trained using the forecast factor feature maps in the training set and the actual precipitation observation data. The model performance is evaluated and the model parameters are adjusted using the forecast factor feature maps in the validation set and the actual precipitation observation data, thus obtaining the trained LA-UNET-LSTM neural network model. The test unit is used to input the feature maps of forecast factors from the test set into the trained model, and at the same time obtain precipitation forecast data for the next 1-4 weeks.

[0028] In another embodiment of the present invention, an electronic device includes: Memory containing executable program code; A processor coupled to the memory; The processor calls the executable program code stored in the memory to execute the steps of the LA-UNET-LSTM-based subseasonal precipitation forecasting method described above, and can achieve the technical effects of the above method.

[0029] In another embodiment of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing computer instructions, which, when invoked, are used to execute the steps of the above-described LA-UNET-LSTM-based subseasonal precipitation forecasting method, and can achieve the technical effects of the above-described method.

Claims

1. A method for predicting subseasonal precipitation based on LA-UNET-LSTM, characterized in that, Includes the following steps: (1) Collect precipitation and multi-meteorological element forecast data output by numerical model, as well as actual precipitation observation data, and divide the precipitation and multi-meteorological element forecast data output by numerical model, as well as actual precipitation observation data, into training set, validation set and test set respectively. (2) Based on the training set data, extract the feature sequences of the predictable precipitation modes, diagnose and analyze the sources of their predictability in the numerical model, and extract the feature mask field; (3) Construct an LA-UNET-LSTM neural network model and build a loss function based on structural similarity and weighted mean square error; the LA-UNET-LSTM neural network includes multiple UNET modules and one LSTM module. Multiple UNET modules are used to extract features from the feature maps of each forecast factor, and then the LSTM module is used for feature fusion; specifically: The formula for calculating structural similarity (SSIM) is: ； ； ； in, and These represent forecast data and observational data, respectively. and These represent the mean values ​​of the forecast data and the observed data, respectively. and These represent the standard deviations of the forecast data and the observed data, respectively. This represents the covariance between forecast data and observed data; and They are respectively and The parameters, The maximum value of the observed data and the forecast data; The formula for calculating the weighted mean square error (WMSE) is: ； in, and Representing the first One forecast and one observation sample, Indicates the total number of samples; The loss function is: ； in, and These are two hyperparameters that are adjusted based on the actual training situation to obtain the optimal model prediction performance; (4) Standardize the training, validation, and test sets, and construct the forecast factor feature maps for the training, validation, and test sets respectively using the feature mask field. Train the LA-UNET-LSTM neural network model using the forecast factor feature maps in the training set and actual precipitation observation data. Evaluate the model performance and adjust the model parameters using the forecast factor feature maps in the validation set and actual precipitation observation data to obtain the trained LA-UNET-LSTM neural network model; including: The standardization process is as follows: First, outliers in the observed and forecast data are removed. Then, the data is processed into weekly average data, and the weekly average anomaly is extracted based on the average value of historical data. The calculation formula is as follows: ； in, It is the average value of the observed / forecasted variable over a certain week. It is the average value of the variable over the same historical period; This is the weekly average anomaly of the variable; after extracting the anomaly, the variable needs to be normalized to its minimum and maximum values. ； in, It is the weekly average anomaly of a certain variable. express The minimum value, express The maximum value, This is the result of normalizing the maximum and minimum weekly average anomalies of the variable; Multiply the forecast data in the standardized training set, validation set, and test set by the corresponding feature mask field obtained in step (2) to obtain the forecast factor feature maps of the training set, validation set, and test set respectively. The structural similarity (SSIM) and weighted mean square error (WMSE) in the model were evaluated using the forecast factor feature map and actual precipitation observation data in the validation set. If the WMSE generated by the model in the validation set is greater than the WMSE of the original model, it indicates that the model has poor forecasting ability for extreme precipitation and the hyperparameters need to be increased. If the SSIM of the model-generated forecast in the validation set is lower than that of the original model forecast, it indicates that the model has a poor grasp of the rainband structure and the hyperparameters need to be increased. value; (5) Input the forecast factor feature map from the test set into the trained model and obtain the precipitation forecast data for the next 1-4 weeks.

2. The subseasonal precipitation forecasting method based on LA-UNET-LSTM according to claim 1, characterized in that, The multi-meteorological element forecast data in step (1) includes: the sea surface temperature field, multi-layer geopotential height field, and multi-layer wind field forecasted by the model. The forecast data all include multiple ensemble members.

3. The subseasonal precipitation forecasting method based on LA-UNET-LSTM according to claim 1, characterized in that, In step (2), feature sequences of predictable precipitation modes are extracted based on the training set data, specifically as follows: First, the residual value of the precipitation forecast is calculated based on the precipitation ensemble forecast data in the training set. The calculation formula is as follows: ； in, These are the forecast values ​​for the precipitation ensemble members. It is the forecast average of the precipitation ensemble members. This indicates the residual in the precipitation forecast; for The noise modes are obtained by empirical orthogonal decomposition of the covariance matrix: ； in, express The covariance matrix, For noise mode, It is by A diagonal matrix composed of eigenvalues. for The transpose of the matrix; then... Projected to That is, to achieve white noise conversion: ； in, After white noise conversion , For the number of members in the precipitation ensemble; then for Singular value decomposition yields the characteristic sequence of predictable modes: ； in, For left singular values, for A diagonal matrix composed of eigenvalues. These are the characteristic sequences of predictable modes.

4. The subseasonal precipitation forecasting method based on LA-UNET-LSTM according to claim 1, characterized in that, Step (2) involves diagnosing and analyzing the sources of predictability in the numerical model and extracting the feature mask field; specifically: Correlation analysis was conducted based on the characteristic sequences of predictable precipitation modes and forecast data of multiple meteorological elements, and the correlation coefficient was determined. The calculation formula is: ； in, The characteristic sequence representing the predictable mode is the first... One element, This represents the average value of the characteristic sequence of the predictable mode. For meteorological elements at a certain grid point One value, This represents the average value of meteorological elements at a certain grid point. The length of the characteristic sequence of the predictable mode; A significance test is performed, and a feature mask field is obtained based on the significance test results. Regions that pass the test are marked as 1, and regions that fail the test are marked as 0.

5. A subseasonal precipitation forecasting system based on LA-UNET-LSTM, characterized in that, include: The dataset creation unit is used to collect precipitation and multi-meteorological element forecast data output by numerical models, as well as actual precipitation observation data. The precipitation and multi-meteorological element forecast data output by numerical models, as well as actual precipitation observation data, are divided into training set, validation set and test set respectively. The feature map construction unit is used to extract feature sequences of predictable precipitation modes based on training set data, diagnose and analyze the sources of their predictability in numerical models, and extract feature mask fields. The network model building unit is used to build the LA-UNET-LSTM neural network model and construct a loss function based on structural similarity and weighted mean square error. The LA-UNET-LSTM neural network includes multiple UNET modules and one LSTM module. Multiple UNET modules are used to extract features from the feature maps of each forecast factor, and then the LSTM module is used for feature fusion. Specifically: The formula for calculating structural similarity (SSIM) is: ； ； ； in, and These represent forecast data and observational data, respectively. and These represent the mean values ​​of the forecast data and the observed data, respectively. and These represent the standard deviations of the forecast data and the observed data, respectively. This represents the covariance between forecast data and observed data; and They are respectively and The parameters, The maximum value of the observed data and the forecast data; The formula for calculating the weighted mean square error (WMSE) is: ； in, and Representing the first One forecast and one observation sample, Indicates the total number of samples; The loss function is: ； in, and These are two hyperparameters that are adjusted based on the actual training situation to obtain the optimal model prediction performance; The model training and validation unit is used to standardize the training, validation, and test sets, and construct forecast factor feature maps for each set using a feature mask field. The LA-UNET-LSTM neural network model is trained using the forecast factor feature maps from the training set and actual precipitation observation data. The model performance is evaluated and parameters are adjusted using the forecast factor feature maps from the validation set and actual precipitation observation data, resulting in a well-trained LA-UNET-LSTM neural network model. This includes: The standardization process is as follows: First, outliers in the observed and forecast data are removed. Then, the data is processed into weekly average data, and the weekly average anomaly is extracted based on the average value of historical data. The calculation formula is as follows: ； in, It is the average value of the observed / forecasted variable over a certain week. It is the average value of the variable over the same historical period; This is the weekly average anomaly of the variable; after extracting the anomaly, the variable needs to be normalized to its minimum and maximum values. ； in, It is the weekly average anomaly of a certain variable. express The minimum value, express The maximum value, This is the result of normalizing the maximum and minimum weekly average anomalies of the variable; Multiply the forecast data in the standardized training set, validation set, and test set by the corresponding feature mask field obtained in step (2) to obtain the forecast factor feature maps of the training set, validation set, and test set respectively. The structural similarity (SSIM) and weighted mean square error (WMSE) in the model were evaluated using the forecast factor feature map and actual precipitation observation data in the validation set. If the WMSE generated by the model in the validation set is greater than the WMSE of the original model, it indicates that the model has poor forecasting ability for extreme precipitation and the hyperparameters need to be increased. If the SSIM of the model-generated forecast in the validation set is lower than that of the original model forecast, it indicates that the model has a poor grasp of the rainband structure and the hyperparameters need to be increased. value; The test unit is used to input the feature maps of forecast factors from the test set into the trained model, and at the same time obtain precipitation forecast data for the next 1-4 weeks.

6. An electronic device, characterized in that, The device includes: Memory containing executable program code; A processor coupled to the memory; The processor calls the executable program code stored in the memory to perform the steps of the subseasonal precipitation forecasting method based on LA-UNET-LSTM as described in any one of claims 1-4.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions, which, when invoked, are used to perform the steps of the subseasonal precipitation forecasting method based on LA-UNET-LSTM as described in any one of claims 1-4.

Images