A method for combining liquid ratio and heavy precipitation nowcasting

Abstract

The present application relates to a kind of strong precipitation nowcasting methods of liquid ratio, comprising: collecting radar reflectivity data and next hour precipitation real-time data within one hour;Radar reflectivity data is converted into three-dimensional data under Cartesian rectangular coordinate system, and precipitation real-time data is converted into two-dimensional precipitation grid field;Strong precipitation nowcasting model is built;Model is trained on training set, hyperparameter is optimized on verification set, and model is tested on test set;Obtain precipitation estimation field data.The present application uses the double decoder based on liquid ratio and attention mechanism to build strong precipitation nowcasting model, uses the echo data of liquid water content conversion processing and carries out the feature coding extraction work using encoder module, using two parallel decoder module respectively output prediction result of strong rain area and rainfall, using comprehensive loss function improves the quantitative estimation and qualitative classification of strong precipitation area, after effectiveness test, the present application can significantly improve nowcasting accuracy.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of weather radars, and relates to technologies related to radar signal processing, radar signal analysis and innovative algorithms for nowcasting, in particular to a strong precipitation nowcasting method combined with liquid ratio. BACKGROUND

[0002] As a typical representative of severe convective weather, short-time heavy precipitation has the characteristics of strong burst and high disaster-causing ability, and frequently brings great losses to social economy and people's life and property. Therefore, accurate nowcasting of short-time heavy precipitation to provide reliable basis for disaster prevention and control has become a key problem to be solved in the meteorological field.

[0003] At present, radar observation data has become an important data source for nowcasting of heavy precipitation due to its high temporal and spatial resolution, rich information and wide coverage. However, most current research models are only based on the simplified correlation between single-level reflectivity intensity and precipitation. Such single-layer radar reflectivity data can only reflect the approximate distribution of water vapor particles in the horizontal direction, and cannot reflect the distribution characteristics and dynamic evolution of water vapor particles in the vertical direction, while the formation of severe convective precipitation is often related to strong vertical movement of air. In addition, relying only on single reflectivity data ignores the influence of many other environmental constraints on precipitation prediction. Recent studies have begun to incorporate multi-source data, but the liquid ratio considering the relationship between water particle phase state and liquid water has not been effectively integrated into the nowcasting model of heavy precipitation. The liquid ratio can effectively reduce the interference of hail particles and solid-liquid mixed particles on precipitation prediction by optimizing the empirical formula of reflectivity value and liquid water conversion, and thus improve the prediction accuracy.

[0004] Therefore, in order to further improve the accuracy of the nowcasting method of heavy precipitation in predicting the area of heavy precipitation, it is necessary to comprehensively analyze the weather system as a three-dimensional whole, make the model deeply learn the spatio-temporal correlation between different height layers, and incorporate a preprocessing module based on liquid ratio to convert the liquid water content of reflectivity data, so as to comprehensively improve the precipitation prediction accuracy and explainability of the nowcasting model of heavy precipitation. SUMMARY

[0005] The purpose of the present application is to overcome the shortcomings of the prior art, and to provide a nowcasting method of heavy precipitation combined with liquid ratio, so as to fully consider the three-dimensional structural characteristics and meteorological characteristics of the weather system, and thus comprehensively improve the precipitation prediction accuracy and explainability of the nowcasting model of heavy precipitation.

[0006] The technical problem of the present application is solved by the following technical scheme:

[0007] A liquid ratio combined heavy rainfall nowcasting method, characterized in that the steps of the method are:

[0008] S1, collect radar reflectivity data within one hour and next hour rainfall real data in hour units;

[0009] S2, convert radar reflectivity data at each radar scanning time from polar coordinate system to three-dimensional data in Cartesian rectangular coordinate system, convert next hour rainfall real data into two-dimensional rainfall grid field by interpolation method, use radar reflectivity three-dimensional data within one hour as input, and next hour two-dimensional rainfall grid field as label to make sample set;

[0010] S3, build a heavy rainfall nowcasting model taking radar reflectivity three-dimensional data within one hour as input and next hour rainfall estimation field data as output, the heavy rainfall nowcasting model is a dual-decoder heavy rainfall nowcasting model based on liquid ratio and attention mechanism;

[0011] S4, divide the sample set into training set, validation set and test set, train the heavy rainfall nowcasting model on the training set, optimize hyperparameters on the validation set, and test the model on the test set;

[0012] S5, use radar reflectivity data within any one hour, perform S2 to convert radar reflectivity data at each radar scanning time from polar coordinate system to three-dimensional data in Cartesian rectangular coordinate system, and use data within this hour as input of the model, to obtain next hour rainfall estimation field data.

[0013] Moreover, S2 converting radar reflectivity data from polar coordinate system to three-dimensional data in Cartesian rectangular coordinate system means first converting radar base data from polar coordinate system to Cartesian coordinate system using bilinear interpolation method in horizontal dimension, and then interpolating in vertical dimension to obtain three-dimensional equidistant grid field data; converting next hour rainfall real data into two-dimensional rainfall grid field by interpolation method means using inverse distance weighted interpolation method to interpolate rainfall record points to a grid of size 128x128 to form a uniform two-dimensional rainfall grid field.

[0014] Moreover, the heavy rainfall nowcasting model of S3 includes an encoder module and two parallel decoder modules, the encoder module is provided with a liquid ratio based preprocessing module at the front end, the liquid ratio based preprocessing module aims to reduce the interference of hail particles and solid-liquid mixed particles on rainfall estimation; the encoder module receives echo data after liquid water content conversion processing and performs feature coding and extraction work, two parallel decoder modules respectively output prediction results of heavy rain area and rainfall, and finally output next hour two-dimensional rainfall estimation field;

[0015] The liquid ratio-based preprocessing module is used to process the special case when the convection system crosses the zero-degree layer. When the convection system crosses the zero-degree layer, the precipitation particles will gradually freeze from liquid to solid, but still appear as a high-value area on the reflectivity map. The module uses the liquid ratio to attenuate the liquid water content M converted from the reflectivity. Assuming that the bottom height of the convection cell strong echo body is H L and the top height is H U , the weighted vertical cumulative liquid water content density HF is obtained according to the following formula:

[0016]

[0017] Wherein: s(h) is the profile area of the cell at height h;

[0018] k(h) is the attenuation coefficient, which undertakes the role of “attenuation” M;

[0019] With the increase of height, ice crystals increase, and actual liquid water content decreases. The attenuation coefficient k is defined as follows:

[0020]

[0021] The liquid ratio-based preprocessing module is based on the construction idea of the liquid ratio feature. Before encoding, the data is converted by height layer using the attenuation coefficient to avoid the misleading of the high reflectivity value generated by the ice crystal particles above the zero-degree layer to the model learning.

[0022] The attention mechanism of the strong precipitation nowcasting model is a convolution attention module. The convolution attention module enhances the ability to capture key features of precipitation by analyzing the channel and spatial dimensions. The module is composed of a channel attention module and a spatial attention module in series. The channel attention module effectively aggregates the global information of the input feature map to highlight the contribution of important channels and suppress irrelevant or redundant information. The spatial attention module captures the mutual dependence between different spatial positions to enhance the attention of the model to key areas.

[0023] Moreover, the S4 trains the strong precipitation nowcasting model using a comprehensive loss function L ALL that combines the rain loss L WMSE and the rain area loss L CLA to guide the model to pay more attention to the prediction accuracy of strong precipitation areas.

[0024] The comprehensive loss function L ALL aims to improve both the quantitative estimation and the qualitative classification of strong precipitation areas. The loss function is defined as follows:

[0025] L ALL =λ WMSE L WMSE +λCLA L CLA

[0026]

[0027] Where: λ WMSE and λ CLA These represent scaling factors for rainfall loss and rain area loss, respectively.

[0028] Q x,y This represents the precipitation amount for the label located at coordinates (x, y).

[0029] This represents the predicted precipitation at coordinates (x, y).

[0030] l i ∈{0,1} represents the label of the real heavy rainfall area;

[0031] p i The corresponding predicted probability.

[0032] The positive effects that this invention can produce are:

[0033] This invention comprehensively analyzes the weather system as a three-dimensional whole, incorporating a preprocessing module based on liquid ratio to convert reflectance data for liquid water content. A dual decoder based on liquid ratio and attention mechanism is used to build a heavy precipitation nowcasting model. An encoder module receives the echo data after liquid water content conversion and performs feature encoding and extraction. Two parallel decoder modules output the predicted heavy rain area and rainfall amount, respectively. A design incorporating rainfall loss (L...) is also included. WMSE ) and rain area losses (L CLA The comprehensive loss function (L) ALL This invention improves the quantitative estimation and qualitative classification of areas with heavy precipitation. Through effectiveness testing of the dual-decoder structure and preprocessing module, it was found that the application of the dual-decoder structure and the integration of a liquid ratio-based preprocessing module into the nowcasting model significantly improves the accuracy of nowcasting. Attached Figure Description

[0034] Figure 1 This is a flowchart of the present invention;

[0035] Figure 2 This is a schematic diagram of the heavy precipitation nowcasting model of the present invention;

[0036] Figure 3 This is a schematic diagram illustrating how the attenuation coefficient of the present invention varies with height.

[0037] Figure 4 This is a schematic diagram of the convolutional attention module structure of the present invention. Detailed Implementation

[0038] The application will be further described in detail by specific examples, the following examples are only descriptive, not limiting, and cannot limit the protection scope of the application.

[0039] The application discloses a liquid ratio-based heavy rain nowcasting method, which has the following steps:

[0040] (1) Collect S-band Doppler weather radar data from May 2015 to October 2016 in units of hours, involving 12 radar stations at 15 different locations in North China and East China, and the radar obtains reflectivity data in VCP21 body scanning mode; the next hour of rainfall real-time data is derived from the historical rainfall information recorded by each rainfall automatic station.

[0041] (2) Convert the radar reflectivity data of each radar scanning moment from the polar coordinate system to the three-dimensional data in the Cartesian rectangular coordinate system, convert the next hour of rainfall real-time data into a two-dimensional rainfall grid field by using an interpolation method, take the radar reflectivity three-dimensional data within one hour as input, and take the two-dimensional rainfall grid field of the next hour as a label to make a sample set.

[0042] The two-dimensional rainfall grid field is obtained by using an inverse distance weighted interpolation method (IDW) to interpolate the automatic station rainfall data within the scanning range of the 15 radars into a uniform two-dimensional rainfall grid field. g The influence is more significant.

[0043] Therefore, the contribution degree of the automatic weather station i (i = 1, 2,..., n) to the specific grid point g is inversely proportional to the distance di to the center of the grid point. Wherein, the inverse distance weighted interpolation calculation method is as follows:

[0044]

[0045] In the formula, n is 6, representing the selection of the six automatic weather stations closest to the grid point g; R i represents the rainfall provided by the i-th automatic station; k is an integer, and in the embodiment, is set to 2.

[0046] (3) Build a heavy rain nowcasting model with three-dimensional radar reflectivity data within one hour as input and precipitation estimation field in the next hour as output. In this embodiment, a dual-decoder heavy rain nowcasting model based on liquid ratio and attention mechanism is used to process radar reflectivity data within one hour. The model mainly consists of three core modules: echo encoder, rainfall decoder and heavy rain area decoder. At the front end of the echo encoder, a preprocessing module based on liquid ratio is set to reduce the interference of hail particles, solid-liquid mixed particles and other factors on precipitation estimation. The echo encoder module is responsible for receiving echo data after liquid water content conversion and performing feature encoding and extraction. The feature map obtained after encoding operation is input into two parallel decoders. Under the guidance of the comprehensive loss function, the rainfall decoder reconstructs the precipitation grid field, and the heavy rain area decoder predicts the position of the area with precipitation greater than 20 mm. The model structure is shown in Figure 2

[0047] The echo encoder maps the preprocessed three-dimensional radar echo sequence to the feature space. Among them, the convolutional layer is used to extract and fuse the local features in the image to generate high-level semantic representations; the convolutional attention module focuses on identifying the most indicative areas for precipitation prediction in different time periods; the 3d-ConvLSTM unit further enhances the time series processing capability of the model by capturing the long-term evolution trend of the global echo pattern. The parameter settings of the encoder are shown in Table 1.

[0048] Table 1 Parameter setting information of echo encoder

[0049]

[0050] The two parallel decoders undertake different tasks. Among them, the rainfall decoder is similar to the same component in other existing models, which is responsible for reconstructing the pixel-level precipitation grid field; while the heavy rain area decoder is specifically used to predict the position of the area with precipitation greater than 20 mm, and its output is a binary image. In view of the output characteristics of the two decoders, this embodiment designs a comprehensive loss function to optimize the prediction performance of the two parts and guide the model to focus on the prediction accuracy of heavy rain areas during training. This method ensures that the model pays attention to the key areas during the learning process, thereby improving the precipitation prediction accuracy and reliability of the strong convective area. The parameter settings of the decoder structure are shown in Tables 2 and 3.

[0051] Table 2 Parameter setting information of rainfall decoder

[0052]

[0053]

[0054] Table 3 Parameter setting information of heavy rain area decoder ​

[0055]

[0056] To facilitate the decoder to make full use of the multi-level feature information of the encoder, the three-layer output feature information of the encoder is connected to two parallel decoders through the skip connection structure of the U-Net network. The decoding process combines three-dimensional convolution and up-sampling operation to keep consistency with the dimension of the output feature of the encoder. Finally, the decoder obtains the expected output size through two-dimensional convolution.

[0057] (4) dividing the sample set into a training set, a validation set and a test set, training the heavy precipitation nowcasting model on the training set, performing hyperparameter tuning on the validation set, and performing model testing on the test set.

[0058] Each radar completes a volume scan every 6 minutes on average, and 10 three-dimensional radar reflectivity images within one hour are used as the input of the heavy precipitation nowcasting model, and the two-dimensional precipitation grid field of the next hour is used as the label, a total of 7167 samples, and the samples are randomly allocated to the training set, the validation set and the test set according to the ratio of 4:1:1. Finally, 4777 samples are obtained in the training set, and 1195 samples are obtained in the validation set and the test set.

[0059] The heavy precipitation nowcasting model is trained using a comprehensive loss function (L WMSE ) combining rain loss (L CLA ) and rain area loss (L ALL ) to guide the model to pay more attention to the prediction accuracy of the heavy precipitation area.

[0060] The rain loss LWMSE is defined as a weighted MSE (Weighted Mean Square Error, WMSE) loss function, which adjusts the weight according to the intensity of precipitation, and higher intensity of precipitation is given a higher weight coefficient, to ensure that the heavy precipitation area receives more attention during training.

[0061] For strong rain area positioning and identification, the precipitation image is converted to a binary image based on a 20mm / h heavy precipitation threshold, and the cross-entropy loss (Cross-Entropy Loss) is used to calculate the difference between the output of the strong rain area decoder and the true label, to define the rain area classification loss L CLA , to give the model classification ability.

[0062] The comprehensive loss function L ALL aims to improve both quantitative estimation and qualitative classification of the heavy precipitation area, and the definition of the loss function is as follows:

[0063] L ALL =λ WMSE L WMSE +λCLA L CLA

[0064]

[0065] where λ WMSE and λ CLA are the scaling factors for rain loss and rain area loss, respectively, Q x,y is the label precipitation at coordinate (x, y), is the predicted precipitation at coordinate (x, y), l i ∈ {0, 1} is the true heavy rain area label, p i is its corresponding prediction probability. Through experimental tests, λ WMSE is set to 1, λ CLA is set to 1, and weight x,y is set as follows:

[0066]

[0067] (5) Using the radar reflectivity data in any hour, the radar reflectivity data at each radar scanning moment is converted from the polar coordinate system to the three-dimensional data in the Cartesian rectangular coordinate system, and the data in this hour is taken as the input of the model, so that the next hour of precipitation estimation field can be obtained.

[0068] To comprehensively evaluate the performance of the model, the embodiment performs pixel-level evaluation from the angles of precipitation events and precipitation values. For the evaluation of precipitation events under different precipitation intensity thresholds, the embodiment uses multiple evaluation indexes, including the Critical Success Index (CSI), the Equitable Threat Score (ETS), and the Heidke Skill Score (HSS) as the event-level evaluation indexes. The greater the values of CSI, ETS, and HSS, the better the model performs. The definitions of CSI, HSS, and ETS are as follows:

[0069]

[0070] The calculation formula of REF is as follows:

[0071]

[0072] Where TP, FN, FP, and TN are defined by the confusion matrix in Table 4:

[0073] Table 4 Confusion Matrix

[0074]

[0075] For the evaluation of precipitation values, the Root Mean Square Error (RMSE) is used as the evaluation index. RMSE provides a quantitative measure of the accuracy of model predictions. Let y i be the observed precipitation value, and be the predicted precipitation value, the RMSE is defined as follows:

[0076]

[0077] The experimental results and analysis of this embodiment include three parts, including the effectiveness of the dual-decoder structure, the effectiveness of the liquid ratio-based preprocessing module, and the comparison test with other typical precipitation prediction methods.

[0078] 1) Effectiveness verification of the dual-decoder structure

[0079] The end-to-end heavy precipitation nowcasting model proposed in this embodiment consists of an encoder and two parallel decoders (as shown in Figure 2 ), denoted as 3dNowcast2. In order to verify its modeling reasoning ability, this section compares and analyzes the 3dNowcast2 model with a single-decoder model. The single-decoder model (denoted as 3dNowcast1) only contains one rainfall decoder, which outputs a pixel-level precipitation grid field, and uses the L WMSE loss function for training. Table 5 and Table 6 show the performance of the two models in various evaluation indicators, where the values in bold represent the optimal scores, and the values underlined represent the suboptimal results.

[0080] Table 5: Two model score results obtained from the perspective of precipitation events

[0081]

[0082] Table 6: Two model score results obtained from the perspective of precipitation values

[0083]

[0084] The dual-decoder model 3dNowcast2 performs well in all types of evaluation indicators, especially under high precipitation threshold conditions. Taking the critical success index CSI as an example, at a 20mm threshold, the performance of 3dNowcast2 is improved by 2.26% compared to 3dNowcast1; at a 30mm threshold, this improvement increases to 7.31%, with a relative improvement rate of 60.8%. In addition, the Root Mean Square Error RMSE, which is used to quantify the evaluation of precipitation values, also performs better.

[0085] These results show that the design of the parallel decoder is reasonable and effective, which significantly enhances the inference ability of the model and the accuracy of precipitation inversion, and has obvious advantages in dealing with heavy precipitation events.

[0086] 2) Effectiveness verification of the liquid ratio-based preprocessing module

[0087] To verify the effect of the liquid ratio-based preprocessing module, this embodiment compares the 3dNowcast2-LRN model with the preprocessing module introduced and the 3dNowcast1 and 3dNowcast2 models without the preprocessing module. The evaluation results of the three models for the same test sample are shown in Tables 7 and 8. It can be seen that:

[0088] Table 7 Model score results obtained from the perspective of precipitation events

[0089]

[0090] Table 8 Model score results obtained from the perspective of precipitation values

[0091]

[0092] ① The 3dNowcast2-LRN model exhibits the optimal performance in the CSI, HSS and ETS score indicators. In particular, under the condition of a 30mm threshold, the CSI score of the model reaches 21.21%, and the relative improvement rates thereof compared with the 3dNowcast1 and 3dNowcast2 models are as high as 76.5% and 9.7%, respectively. From the evaluation results of the RMSE, the 3dNowcast2-LRN model is slightly lower than the 3dNowcast2 model by 0.1557.

[0093] ② Considering the classification evaluation of precipitation events and the results of precipitation value prediction, the overall performance of the 3dNowcast2-LRN model is the best. This result confirms that the liquid ratio-based preprocessing module can effectively reduce the interference of the high reflectivity value of ice crystal particles on the model learning process, thereby weakening or even eliminating the overestimation problem of precipitation prediction caused thereby. It can be said that the prediction ability of the 3dNowcast2-LRN model in dealing with complex precipitation conditions has been significantly enhanced.

[0094] 3) Comparison experiment with other typical methods

[0095] The 3dNowcast2-LRN model proposed in this embodiment is compared with three typical nowcasting methods. The comparison method and model are the traditional Z-R relationship method, the ConvLSTM model, and the Attention model with the introduction of a self-attention mechanism. Among them, the Z-R relationship method uses the 2km height layer reflectivity data in the echo sequence obtained by the radar extrapolation model as the input. Tables 9 and 10 show the scores obtained by each model on the test set.

[0096] Table 9: Model score results obtained from the perspective of precipitation events

[0097]

[0098] Table 10: Model score results obtained from the perspective of precipitation values

[0099]

[0100] It can be seen that the model accuracy ranking from high to low is: 3dNowcast2-LRN, Attention, ConvLSTM, Z-R relationship. Among the three comparison algorithms, the ConvLSTM model and the Attention model perform significantly better than the Z-R relationship, and the Attention model is better than the ConvLSTM model. The 3dNowcast2-LRN model proposed in this embodiment has obvious improvement compared with the Attention model. Taking CSI as an example, the improvement of the 3dNowcast2-LRN model is 6.88% and 5.91% at the thresholds of 20mm and 30mm, respectively, and the relative improvement rate is 26.3% and 38.7%, respectively. In terms of RMSE, the 3dNowcast2-LRN model has more significant improvement, and its RMSE value in the heavy precipitation area is reduced by 4.9242, with a relative reduction rate of 44.1%, further confirming the advantage of the model in dealing with complex precipitation conditions.

[0101] Although the embodiments of the present application and the drawings are disclosed for illustrative purposes, those skilled in the art can understand that various alternatives, changes and modifications are possible without departing from the spirit and scope of the present application and the appended claims, therefore, the scope of the present application is not limited to the disclosed content of the embodiments and drawings.

Claims

1. A method for nowcasting heavy precipitation based on liquid ratio, characterized in that: The steps of the method are as follows: S1. Collect radar reflectivity data for one hour and precipitation data for the next hour, on an hourly basis. S2. Convert the radar reflectivity data at each radar scan time from polar coordinates to three-dimensional data in Cartesian coordinates. Convert the actual precipitation data for the next hour into a two-dimensional precipitation grid field using interpolation. Use the three-dimensional radar reflectivity data for one hour as input and the two-dimensional precipitation grid field for the next hour as labels to create a sample set. S3. Construct a heavy precipitation nowcasting model that takes three-dimensional radar reflectivity data within one hour as input and precipitation estimation field data within the next hour as output. The heavy precipitation nowcasting model is a dual-decoder heavy precipitation nowcasting model based on liquid ratio and attention mechanism. The heavy precipitation nowcasting model includes an encoder module and two parallel decoder modules. The encoder module is equipped with a preprocessing module based on liquid ratio at its front end. The preprocessing module based on liquid ratio aims to reduce the interference of hail particles and solid-liquid mixed particles on precipitation estimation. The encoder module receives echo data after liquid water content conversion and processing and performs feature encoding and extraction. The two parallel decoder modules output the prediction results of heavy rain area and precipitation amount, respectively. Finally, the two-dimensional precipitation estimation field for the next hour is output. S4. Divide the sample set into a training set, a validation set, and a test set. Train the heavy precipitation nowcasting model on the training set, perform hyperparameter tuning on the validation set, and test the model on the test set. S5. Using radar reflectivity data for any given hour, execute step S2 to convert the radar reflectivity data for each radar scan moment from polar coordinates to three-dimensional data in Cartesian coordinates. Use the data for that hour as input to the model to obtain the estimated precipitation field data for the next hour.

2. The method for nowcasting heavy precipitation based on liquid ratio according to claim 1, characterized in that: The process of converting radar reflectivity data from polar coordinates to Cartesian coordinates in S2 refers to first converting the radar base data from polar coordinates to Cartesian coordinates using bilinear interpolation in the horizontal dimension, and then interpolating in the vertical dimension to obtain three-dimensional equidistant grid field data. The process of converting the next hour's precipitation data into a two-dimensional precipitation grid field using interpolation refers to using inverse distance weighted interpolation to interpolate precipitation recording points onto a 128×128 grid to form a uniform two-dimensional precipitation grid field.

3. The method for nowcasting heavy precipitation based on liquid ratio according to claim 1, characterized in that: The liquid ratio-based preprocessing module is used to handle special cases that occur when a convective system crosses the zero-degree layer. When the convective system crosses the zero-degree layer, precipitation particles gradually freeze from a liquid state to a solid state, but still appear as high-value regions on the reflectivity map. This module uses the liquid ratio to attenuate the liquid water content M obtained from reflectivity conversion. Let H be the bottom height of the strong echo body of the convective cell. L The top height is H U The weighted vertical cumulative liquid water content density HF is obtained according to the following formula: Where: s(h) is the cross-sectional area of ​​the single unit located at height h; k(h) is the attenuation coefficient, which is responsible for the "attenuation" of M; As altitude increases, ice crystals increase, and the actual liquid water content decreases. The attenuation coefficient k is defined as follows: The preprocessing module based on liquid ratio is based on the construction idea of ​​liquid ratio characteristics. Before encoding, it uses the attenuation coefficient to convert the data according to the height layer to avoid the high reflectivity value generated by ice crystal particles above the zero degree layer from misleading the model learning. The attention mechanism of the heavy precipitation nowcasting model is a convolutional attention module. This module enhances the ability to capture key precipitation features by analyzing channel and spatial dimensions. The module consists of a channel attention module and a spatial attention module connected in series. The channel attention module effectively aggregates global information from the input feature map to highlight the contribution of important channels and suppress irrelevant or redundant information. The spatial attention module captures the interdependence between different spatial locations to enhance the model's attention to key areas.

4. The method for nowcasting heavy precipitation based on liquid ratio as described in claim 1, characterized in that: The S4 training of the heavy precipitation nowcasting model uses a combination of rainfall loss L WMSE and rain area losses L CLA The comprehensive loss function L ALL This is to guide the model to pay more attention to the accuracy of predictions for areas with heavy rainfall; Comprehensive loss function L ALL The loss function is defined as follows, aiming to simultaneously improve the quantitative estimation and qualitative classification of areas with heavy precipitation: L ALL =λ WMSE L WMSE +λ CLA L CLA Where: λ WMSE and λ CLA These represent scaling factors for rainfall loss and rain area loss, respectively. Q x,y This represents the precipitation amount for the label located at coordinates (x, y). This represents the predicted precipitation at coordinates (x, y). l i ∈{0,1} represents the label of a real heavy rainfall area; p i The corresponding predicted probability.

Images