Nearby precipitation prediction method and system based on contribution score of heavy precipitation

Abstract

The application discloses a near precipitation prediction method and system based on strong precipitation contribution score, which carries out near precipitation prediction according to real-time multi-modal meteorological live data through an improved near precipitation prediction model; the improved near precipitation prediction model is established based on a Vision Transformer network, contribution ranking scores of the multi-modal meteorological live data are calculated through a factor importance replacement experiment, optimal dynamic learning weight coefficients of the multi-modal meteorological live data are calculated according to the contribution ranking scores, and feature vectors of the multi-modal meteorological live data are spliced and fused according to the optimal dynamic learning weight coefficients. The application improves strong precipitation physical knowledge guidance of cross-modal dynamic learning, improves the fusion effect of multi-modal data features, and improves the strong precipitation prediction accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of precipitation prediction, in particular to a near-precipitation prediction method and system based on strong precipitation contribution score. BACKGROUND

[0002] Strong precipitation in a short time is easy to induce disasters such as waterlogging and mountain torrents in complex terrain mountainous areas. Near-precipitation prediction is of great significance to industrial and agricultural production, major social activities, and disaster prevention and mitigation. Deep learning extrapolation prediction method is the main means to quickly obtain high-quality near-precipitation prediction. The defects caused by linear vector calculation of traditional optical flow method cannot predict the nonlinear complex changes of precipitation weather process. Therefore, deep learning has become the main technical means for near-precipitation prediction in recent years. Most of the existing deep learning near-precipitation prediction methods pursue model structure improvement, but ignore the meteorological physical connection of massive multi-source meteorological data, and the accuracy of strong precipitation is not high. Some researchers try to input meteorological multi-source modal data, which improves the prediction accuracy of strong precipitation to a certain extent, but due to the "black box" effect of deep learning, it is impossible to know the influence degree of each modal meteorological data on strong precipitation, the multi-modal model data fusion mechanism is not good, it is difficult to capture the strong precipitation physical characteristics of the dominant factor, and the prediction accuracy of strong precipitation is not high. SUMMARY

[0003] The purpose of the present application is to provide a near-precipitation prediction method and system based on strong precipitation contribution score, so as to solve the problem that the influence degree of meteorological data on strong precipitation is not clear, the multi-modal model data fusion effect is not good, the model is difficult to capture the strong precipitation physical characteristics, and the prediction accuracy is not high.

[0004] Technical scheme: the near-precipitation prediction method based on strong precipitation contribution score provided by the present application comprises:

[0005] An improved near-precipitation prediction model is established, real-time multi-modal meteorological real-time data is input into the improved near-precipitation prediction model, and near-precipitation prediction is carried out;

[0006] The improved near-precipitation prediction model is established based on a Vision Transformer network, wherein the feature vectors of the multi-modal meteorological real-time data are spliced and fused according to optimal dynamic learning weight coefficients;

[0007] The method for calculating the optimal dynamic learning weight coefficient comprises the following steps: collecting multi-modal meteorological real-time data and constructing a test data set; establishing a deep learning nowcasting precipitation prediction model for the multi-modal meteorological real-time data based on a Vision Transformer network, and generating a test nowcasting precipitation prediction within a period of time by using the test data set; calculating a contribution ranking score of the multi-modal meteorological real-time data to the heavy precipitation prediction by a factor importance permutation experiment according to the test nowcasting precipitation prediction and the test data set; and calculating the optimal dynamic learning weight coefficient of the multi-modal meteorological real-time data according to the contribution ranking score.

[0008] Further, the method for calculating the contribution ranking score of the multi-modal meteorological real-time data in the nowcasting precipitation prediction model according to the test nowcasting precipitation prediction and the test data set by the factor importance permutation experiment comprises the following steps:

[0009] According to the precipitation data at the same time in the test nowcasting precipitation prediction and the test data set, an undisturbed precipitation prediction ETS score is calculated;

[0010] According to the factor importance permutation experiment, the samples in the test data set are permuted, and the permuted test data set is used for re-prediction to obtain a disturbed precipitation prediction and calculate a disturbed precipitation prediction ETS score;

[0011] The contribution ranking score of the heavy precipitation prediction is obtained by subtracting the disturbed precipitation prediction ETS score from the undisturbed precipitation prediction ETS score.

[0012] Further, the method for splicing and fusing the feature vectors of the multi-modal meteorological real-time data according to the optimal dynamic learning weight coefficient comprises the following steps:

[0013] The optimal dynamic learning weight coefficient is used as an initial weight coefficient of a multi-head attention mechanism, and a multi-head attention mechanism feature vector and a dynamic weight dependent coefficient of cross-modal dynamic weight learning are calculated;

[0014] The dynamic weight dependent coefficient is normalized to obtain a dynamic weight coefficient of cross-modal learning, and the optimal weight coefficient is dynamically updated;

[0015] The feature vectors of each mode are weighted and spliced according to the dynamic weight coefficient to obtain a spliced feature vector.

[0016] Further, the multi-modal meteorological real-time data comprises ground meteorological automatic station multi-element minute-level observation data and radar picture product data of a prediction area;

[0017] The ground meteorological automatic station multi-element minute-level observation data comprises temperature, pressure, relative humidity, wind direction, wind speed and precipitation observation data;

[0018] Radar imagery product data includes: radar composite reflectivity factor network mosaic and radar vertical liquid integral water content image.

[0019] Furthermore, it also includes preprocessing and normalizing multimodal meteorological data, including: using the bilinear interpolation method to process multimodal meteorological data into gridded field data with uniform spatiotemporal resolution, and normalizing the gridded field data.

[0020] Furthermore, both the now-near precipitation forecast model and the improved now-near precipitation forecast model include a location coding layer. The location coding layer divides and positions the input multimodal meteorological data to obtain the feature vector of the multimodal meteorological data.

[0021] Furthermore, the near-term precipitation forecast is a precipitation forecast every 6 minutes from 0 to 2 hours.

[0022] Furthermore, the meteorological data with the highest contribution ranking score was the dominant factor in heavy precipitation.

[0023] The now-nearest precipitation forecasting system based on heavy precipitation contribution score as described in this invention includes:

[0024] A multimodal data contribution scoring unit is used to collect multimodal meteorological data and construct a test dataset; a deep learning nowcasting model for multimodal meteorological data is established based on the Vision Transformer network, and a test nowcasting forecast for a period of time is generated using the test dataset; based on the test nowcasting forecast and the test dataset, the contribution ranking score of multimodal meteorological data to heavy precipitation forecast is calculated through factor importance permutation experiments.

[0025] The modeling unit for the now-near precipitation forecast is used to establish an improved now-near precipitation forecast model. The improved now-near precipitation forecast model is based on the Vision Transformer network. The optimal dynamic learning weight coefficients of the multimodal meteorological data are calculated according to the contribution ranking score. The feature vectors of the multimodal meteorological data are spliced ​​and fused according to the optimal dynamic learning weight coefficients.

[0026] The now-near precipitation forecasting unit is used to acquire real-time multimodal meteorological data to input into the improved now-near precipitation forecasting model for now-near precipitation forecasting.

[0027] The electronic device of 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 onto the processor, it implements the now-nearest precipitation forecast based on the contribution score of heavy precipitation.

[0028] Beneficial effects: Compared with the prior art, the advantages of this invention are as follows: Based on factor importance permutation experiments and cross-modal dynamic weight learning, this invention establishes an improved near-term precipitation forecasting model based on the Vision Transformer network. Through factor importance permutation experiments, the heavy precipitation contribution ranking and dominant factors of the deep learning model are calculated, and the optimal initial values ​​of the weight coefficients are determined accordingly. This improves the guidance of heavy precipitation physics knowledge in cross-modal dynamic learning, enhances the fusion effect of multimodal data features, guides the model to better capture the physical characteristics of heavy precipitation, and improves the accuracy of heavy precipitation forecasting. Attached Figure Description

[0029] Figure 1 This is a flowchart of the near-term precipitation forecasting method according to an embodiment of the present invention.

[0030] Figure 2 This is a schematic diagram of the improved nowcasting structure for precipitation according to an embodiment of the present invention. Detailed Implementation

[0031] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0032] like Figure 1 As shown, the near-term precipitation forecasting method includes the following steps.

[0033] Step 1: Collect multimodal real-time meteorological data, perform preprocessing and normalization, and construct a training dataset. This includes the following steps.

[0034] Step 1.1: Collect multimodal meteorological data for the forecast area (depending on specific circumstances), including minute-level multi-element observation data from automatic surface weather stations and radar imagery products. The minute-level multi-element observation data from automatic surface weather stations includes: temperature, air pressure, relative humidity, wind direction, wind speed, and precipitation observation data; the radar imagery products include: radar composite reflectivity factor mosaic and radar vertical liquid integral water content images.

[0035] Step 1.2 involves preprocessing and normalizing the data to construct training and testing datasets. Outliers in the minute-level observation data of multi-element automatic weather stations are removed using the average climate standard value (average over the past 30 years), and missing data are supplemented. Then, bilinear interpolation is used to process the minute-level observation data of multi-element automatic weather stations and radar image products into gridded field data with a uniform spatiotemporal resolution of 6 minutes and 0.125°×0.125° (equidistant latitude and longitude grids). Next, the gridded field data is normalized to reduce the data to the range [0,1] to improve model training and forecasting performance. Finally, the training and testing datasets are constructed using the sample method of "inputting gridded field data for the first 2 hours at 6-minute intervals → outputting precipitation gridded field data for the next 2 hours".

[0036] The bilinear interpolation method is calculated as follows: Assume the grid points to be interpolated are... Given the values ​​of the four surrounding points: the bottom left is... The bottom right is The top left is The top right is in , First calculate shaft and Weight in the axial direction , :

[0037] , (1)

[0038] , Given four points in Values ​​in the axial direction, , for Values ​​along the axis. Then, along... Direction interpolation between the upper and lower rows:

[0039] , (2)

[0040] , They represent along Points obtained by interpolation of the upper and lower rows.

[0041] Then along Directional interpolation is used to obtain the required grid point values:

[0042] (3)

[0043] Finally, data normalization is performed.

[0044] (4)

[0045] , These represent the maximum and minimum values ​​in the entire sample data, respectively. For the normalized dataset samples, For grid point values ​​that need to be normalized.

[0046] Step 2: Based on the Vision Transformer network and the training dataset, construct a deep learning-based nowcasting precipitation model. This includes the following steps.

[0047] Step 2.1: Based on the Vision Transformer network, construct a deep learning multimodal data-based nowcasting precipitation forecasting model network. The model network consists of: a location encoding layer, a spatiotemporal feature mixing layer, and an output layer. The location encoding layer is constructed based on location encoding and block-based computation; a multi-head attention mechanism is used to construct the spatiotemporal feature mixing layer; and a feedforward neural network combined with a regularization layer is used to construct the output layer. Details are as follows:

[0048] (a) The location encoding layer performs block segmentation and location encoding on the input multimodal meteorological data, replacing the convolution operation in traditional deep learning. This improves the computational efficiency of data feature extraction, yields the location-encoded feature vectors of the multimodal meteorological data, and prepares for subsequent multi-head attention mechanism fusion of spatiotemporal features. First, non-overlapping block segmentation is performed on the multimodal data, dividing the data grid field into N grids of size [missing information]. (like (blocks)

[0049] (5)

[0050] N= (6)

[0051] This represents the input real-time meteorological data. This represents the training dataset obtained in step 1. To account for the number of channels in the model, the multimodal meteorological data in this embodiment requires a total of 12 channels, including: 6 channels of observation data of temperature, air pressure, relative humidity, wind direction, wind speed and precipitation from ground automatic weather stations; 6 channels of radar combined reflectivity factor network mosaic and radar vertical liquid integral water content image. Each type of image data includes 3 color channels: red, green and blue. This indicates the number of latitudinal grid points in the input meteorological data. This represents the number of meridional grid points in the input meteorological data, where S is the time step.

[0052] Then position encoding is performed:

[0053] , (6)

[0054] The positional encoding feature vector is added after encoding. This represents the data block that needs to be encoded. This indicates the position code.

[0055] (b) Spatiotemporal feature mixing layer: This layer uses a multi-head attention mechanism to capture the spatiotemporal dependencies between position-encoded feature vectors. First, spatiotemporal feature projection is performed, which decomposes high-dimensional, complex spatiotemporal features into multiple low-dimensional components, improving computational efficiency and model expressive power. The calculation formula is as follows:

[0056] , , (7)

[0057] Encode the feature vector for the location. , and For learnable projection matrices, For querying the matrix, The key matrix, This is the value matrix. Then, a multi-head attention mechanism is used to extract the global and local dependencies of the feature vectors, resulting in the multi-head attention mechanism feature vectors:

[0058] (8)

[0059] in, This is the feature vector for the multi-head attention mechanism. Represents a nonlinear function. For querying the matrix, Key matrix The transpose of the matrix, Projected dimensions of queries and keys in each header.

[0060] (c) Output layer: First, the feature vectors calculated by the multi-head attention mechanism are spliced ​​and fused. Then, a certain regularization layer is added to prevent overfitting, gradient vanishing or exploding, which would lead to a decrease in generalization performance. Then, a feedforward neural network is used and a nonlinear transformation is performed to map it to the 0-2h normalized precipitation grid field to obtain the normalized near-term precipitation forecast. In this embodiment, the 0-2h 6-minute precipitation grid field is used as an example.

[0061] (7)

[0062] That is, the feature vector after concatenation and fusion. This is the weight matrix. For the first to the The feature vectors of the multi-head attention mechanism for each model's data channels This indicates a linear splicing and merging operation.

[0063] (8)

[0064] (9)

[0065] , These represent the feature vectors output by the regularization layer and the feedforward neural network, respectively. , Here is the weight matrix of the feedforward neural network. , For bias terms, Indicates a regularization layer. This represents a feedforward neural network. This is a nonlinear activation function. Finally, by inversely normalizing the normalized nowcast precipitation forecast, we can obtain the 0-2h 6-minute precipitation forecast grid field:

[0066] (10)

[0067] , These represent the maximum and minimum values ​​in the dataset samples, respectively. This is the normalized value of precipitation. This is for deep learning of the 0-2h 6-minute nowcasting grid field for precipitation.

[0068] Step 2.2: Train the Vision Transformer nowcasting model network based on the training dataset, converge to obtain a stable model, and use the test dataset to obtain nowcasting forecasts for at least one year. By gradually adjusting model parameters such as learning rate, batch size, and training epochs, train the Vision Transformer nowcasting model network using the training dataset, gradually converge to obtain a stable deep learning nowcasting model, and then use the gridded field data from the test dataset to input the stable model to obtain nowcasting forecasts for at least one year, from 0-2 hours every 6 minutes.

[0069] Step 3: Based on deep learning-based nowcasting of precipitation and factor importance permutation experiments, calculate the ranking of heavy precipitation contribution scores and dominant physical factors of the deep learning model. This determines the optimal dynamic learning weight coefficients for multimodal data features, guiding model improvement based on heavy precipitation physics knowledge and enhancing the accuracy of heavy precipitation forecasts. Specifically, this includes the following steps.

[0070] First, based on at least one year of 0-2h 6-minute nowcast precipitation forecasts and precipitation gridded field data at the same time in the test dataset, the unperturbed precipitation forecast ETS score is calculated:

[0071] (11)

[0072] (12)

[0073] NA represents the number of events that occurred in both forecasts and observations; NB represents the number of events that occurred in forecasts but not in observations; NC represents the number of events that did not occur in forecasts but were observed; ND represents the number of events that occurred in both forecasts and observations; and RA represents the number of events that did not occur in both forecasts and observations.

[0074] Then, based on the principle of factor importance permutation experiment, the temporal order of the uniform spatiotemporal resolution grid field data in the test dataset samples is shuffled one by one using methods such as reverse order and random order, and mixed with other samples. The permuted sample data is used to make a new prediction to obtain the perturbation prediction, and the ETS score of the perturbation prediction is calculated.

[0075] Finally, the difference between the ETS score of the perturbed precipitation forecast and the ETS score of the undisturbed precipitation forecast is calculated to obtain the ETS score difference, which is used as the ranking score for heavy precipitation contribution. The larger the difference, the more important the corresponding feature. To obtain a stable and reliable heavy precipitation contribution ranking score, the above steps are repeated 2-3 times to obtain the final heavy precipitation contribution ranking score. This score is then used to calculate and determine the optimal dynamic learning weight coefficients for the multimodal forecast, as shown in the following formula:

[0076] (13)

[0077] The optimal dynamic learning weight coefficients, Contribute scores to heavy precipitation data in multimodal data. Indicates the number of channels.

[0078] Furthermore, the dominant factors of heavy precipitation can be screened out, namely the meteorological data with the highest contribution ranking score.

[0079] Step 4: Based on the optimal dynamic learning weight coefficients, cross-modal dynamic weight learning is introduced to improve the VisionTransformer nowcasting model network, establishing a VisionTransformer nowcasting model with improved heavy precipitation dominant factors. A late-stage fusion layer is established based on cross-modal dynamic weight learning and the optimal dynamic learning weight coefficients to achieve cross-modal dynamic weight learning of multimodal data, improve the fusion effect of multimodal data features, guide the model to better capture the physical characteristics of heavy precipitation, improve the model's physical knowledge guidance, and increase the accuracy of heavy precipitation forecasts. This specifically includes the following steps.

[0080] Step 4.1: Based on the optimal dynamic learning weight coefficients and cross-modal dynamic weight learning, construct the late fusion layer to improve the spatiotemporal feature mixing layer of the Vison Transformer nowcasting model network. Figure 2 The diagram shows the improved nowcasting structure. In the improved Vison Transformer nowcasting model, the location encoding layer and output layer are the same as in step 2. A total of 12 channels are needed for the multimodal meteorological data. Figure 2 Some input channels have been omitted for demonstration purposes.

[0081] In the spatiotemporal feature mixing layer, the optimal dynamic learning weight coefficients are used as the initial weight coefficients of the multi-head attention mechanism feature vector. The multi-head attention mechanism feature vector of cross-modal dynamic weight learning is calculated to realize the initial splicing and fusion of multi-modal data feature vectors, thereby guiding the model to capture the physical characteristics of heavy precipitation and optimizing the multi-modal data feature fusion effect.

[0082] In the spatiotemporal feature mixing layer, the feature vector calculation formula using the multi-head attention mechanism learned through cross-modal dynamic weights is as follows:

[0083] (14)

[0084] For cross-modal multi-head attention mechanism feature vectors, Represents a nonlinear function. For querying the matrix, Key matrix The transpose of the matrix, Projected dimensions of queries and keys in each header. The optimal dynamic learning weight coefficients will be continuously improved and changed during cross-modal dynamic weight learning.

[0085] Simultaneously, the dynamic weight dependency coefficients of each data channel are calculated in parallel. The calculation formula is as follows;

[0086] (15)

[0087] Let be the dynamic weight dependency coefficient of the i-th data channel. Let i be the key matrix of the i-th data channel. Let represent the transpose of the query matrix for the i-th data channel, and exp represent the nonlinear empirical function; in this embodiment, the trigonometric cosine function is used as an example. The number of data channels in the model. This represents the projection dimension of the query and key in each header.

[0088] The dynamic weight coefficients are then obtained through Softmax normalization, which are used to update the optimal weight coefficients of the spatiotemporal feature mixing layer. These coefficients also serve as weight coefficients for later feature vector fusion and are continuously optimized and improved during dynamic learning. The calculation formula is as follows:

[0089] (16)

[0090] Represents a nonlinear function. This represents the dynamic weight coefficients obtained from cross-modal learning of the i-th channel. Let be the dynamic weight dependency coefficient of the i-th data channel. This represents the number of data channels in the model.

[0091] Then, a post-fusion layer is added after the spatiotemporal feature mixing layer to weight and concatenate the features of each modality according to the obtained dynamic weight coefficients:

[0092] (17)

[0093] This refers to the feature vector obtained after concatenation and fusion, specifically the feature vector obtained through cross-modal multi-head attention mechanism. For dynamic weighting coefficients, For the first to the Feature vectors of multi-head attention mechanism for each data channel This indicates a splicing and merging operation.

[0094] Then, using the same calculation method as the output layer in step 2.1, the feature vectors obtained from the later fusion layer are mapped to the 0-2h normalized precipitation grid field using a feedforward neural network and a regularization layer, and then inverse normalization is performed to obtain the improved 0-2h 6-minute nowcast precipitation forecast.

[0095] Step 4.2: Train the Vision Transformer nowcasting model network with improved heavy precipitation dominant factors based on the training dataset to obtain a stable nowcasting model. Input real-time multimodal meteorological data to obtain the improved 0-2h precipitation forecast every 6 minutes.

[0096] By gradually adjusting model parameters such as learning rate, batch size, and training epochs, a Vision Transformer nowcasting model network improved by the dominant factor of heavy precipitation was trained using the training dataset, gradually converging to obtain a stable deep learning nowcasting model. Then, real-time multi-element minute-level observation data from automatic surface meteorological stations and radar imagery products of the forecast area were collected. The data were preprocessed and normalized using the same method as in step 1 to form a gridded field with uniform spatiotemporal resolution, which was then input into the converged stable model to obtain a 0-2h 6-minute precipitation forecast improved by the dominant factor of heavy precipitation.

[0097] The now-nearest precipitation forecasting system based on heavy precipitation contribution score as described in this invention includes:

[0098] A multimodal data contribution scoring unit is used to collect multimodal meteorological data and construct a test dataset; a deep learning nowcasting model for multimodal meteorological data is established based on the Vision Transformer network, and a test nowcasting forecast for a period of time is generated using the test dataset; based on the test nowcasting forecast and the test dataset, the contribution ranking score of multimodal meteorological data to heavy precipitation forecast is calculated through factor importance permutation experiments.

[0099] The modeling unit for the now-near precipitation forecast is used to establish an improved now-near precipitation forecast model. The improved now-near precipitation forecast model is based on the Vision Transformer network. The optimal dynamic learning weight coefficients of the multimodal meteorological data are calculated according to the contribution ranking score. The feature vectors of the multimodal meteorological data are spliced ​​and fused according to the optimal dynamic learning weight coefficients.

[0100] The now-near precipitation forecasting unit is used to acquire real-time multimodal meteorological data to input into the improved now-near precipitation forecasting model for now-near precipitation forecasting.

[0101] The electronic device of 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 onto the processor, it implements the now-nearest precipitation forecast based on the contribution score of heavy precipitation.

[0102] The electronic device of 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 onto the processor, it implements the now-nearest precipitation forecast based on a heavy precipitation contribution score. The processor executes the computer program stored in the memory to implement the various steps of the methods described in the above embodiments.

Claims

1. A nowcasting method of precipitation based on a contribution score of heavy precipitation, characterized in that, The application relates to a method for improving a near precipitation prediction model, and comprises the following steps: An improved near precipitation prediction model is established, real-time multi-modal meteorological live data is input into the improved near precipitation prediction model, and near precipitation prediction is carried out; The improved near precipitation prediction model is established based on a Vision Transformer network, wherein feature vectors of the multi-modal meteorological live data are spliced and fused according to optimal dynamic learning weight coefficients; The calculation method of the optimal dynamic learning weight coefficients comprises the following steps: multi-modal meteorological live data is collected and a test data set is constructed; a deep learning near precipitation prediction model of the multi-modal meteorological live data is established based on a Vision Transformer network, and test near precipitation prediction in a period of time is generated by using the test data set; according to the test near precipitation prediction and the test data set, the contribution ranking score of the multi-modal meteorological live data to strong precipitation prediction is calculated through a factor importance permutation experiment; and the optimal dynamic learning weight coefficients of the multi-modal meteorological live data are calculated according to the contribution ranking score. The splicing and fusing of the feature vectors of the multi-modal meteorological live data according to the optimal dynamic learning weight coefficients comprises the following steps: The optimal dynamic learning weight coefficients are used as initial weight coefficients of a multi-head attention mechanism, multi-head attention mechanism feature vectors of cross-modal dynamic weight learning and dynamic weight dependent coefficients are calculated, and the dynamic weight dependent coefficients are normalized to obtain dynamic weight coefficients of cross-modal learning; The feature vectors of various modes are weighted and spliced according to the dynamic weight coefficients to obtain spliced feature vectors. The optimal dynamic learning weight coefficients are as follows: ； optimal dynamic learning weight coefficient for the i-th data channel, denotes the number of channels;​ The dynamic weight dependent coefficients are as follows: ； a dynamic weight-dependent coefficient for the i-th data channel, a key matrix for the i-th data channel, a transpose matrix of a query matrix for the i-th data channel, exp is a non-linear experience function, denotes the projection dimension of the query and the key in each head; The dynamic weight coefficients are as follows: ； denotes a non-linear function, denotes a dynamic weight coefficient obtained by cross-modal learning of the i th channel.

2. The nowcasting method of claim 1, wherein the nowcasting method is based on a contribution score of heavy precipitation. The calculation of the contribution ranking score of the multi-modal meteorological live data in the near precipitation prediction model according to the test near precipitation prediction and the test data set through the factor importance permutation experiment comprises the following steps: According to the precipitation data of the same time in the test near precipitation prediction and the test data set, undisturbed precipitation prediction ETS scores are calculated; According to the factor importance permutation experiment, samples in the test data set are permuted, the permuted test data set is used for re-prediction, disturbed precipitation prediction is obtained, and disturbed precipitation prediction ETS scores are calculated; The contribution ranking score of strong precipitation prediction is obtained by subtracting the disturbed precipitation prediction ETS scores from the undisturbed precipitation prediction ETS scores.

3. The nowcasting method based on the contribution score of heavy precipitation according to claim 1, wherein, The multi-modal meteorological live data comprises ground meteorological automatic station multi-element minute-level observation data and radar picture product data of a prediction area. The ground meteorological automatic station multi-element minute-level observation data comprises air temperature, air pressure, relative humidity, wind direction, wind speed and precipitation observation data. The radar picture product data comprises radar combined reflectivity factor networking mosaics and radar vertical liquid water content pictures.

4. The nowcasting method based on the contribution score of heavy precipitation according to claim 1, wherein, The method further comprises pre-processing and normalization operations on the multi-modal meteorological live data, which comprises processing the multi-modal meteorological live data into grid field data of a unified space-time resolution by using a bilinear difference method, and performing normalization processing on the grid field data.

5. The nowcasting method based on the contribution score of heavy precipitation according to claim 1, wherein, The near precipitation prediction model and the improved near precipitation prediction model both comprise a position coding layer, which blocks and positionally encodes the input multi-modal meteorological live data to obtain a position coding feature vector of the multi-modal meteorological live data.

6. The nowcasting method based on the contribution score of heavy precipitation according to claim 1, wherein, The near precipitation prediction is a 0-2 hour 6-minute precipitation prediction.

7. The nowcasting method based on the contribution score of heavy precipitation according to claim 1, wherein, The meteorological live data with the highest contribution ranking score is the dominant factor of heavy precipitation.

8. A nowcasting precipitation forecast system based on a contribution score of heavy precipitation, characterized in that, Comprise: A multi-modal data heavy precipitation contribution scoring unit is configured to collect multi-modal meteorological live data and construct a test data set; a deep learning near precipitation prediction model of the multi-modal meteorological live data is established based on a Vision Transformer network, and a test near precipitation prediction within a period of time is generated by using the test data set; and a factor importance permutation experiment is performed to calculate a contribution ranking score of the multi-modal meteorological live data to heavy precipitation prediction according to the test near precipitation prediction and the test data set; A near precipitation prediction model modeling unit is configured to establish an improved near precipitation prediction model, which is established based on a Vision Transformer network, wherein an optimal dynamic learning weight coefficient of the multi-modal meteorological live data is calculated according to the contribution ranking score, and a feature vector of the multi-modal meteorological live data is spliced and fused according to the optimal dynamic learning weight coefficient; A near precipitation prediction unit is configured to input real-time multi-modal meteorological live data into the improved near precipitation prediction model to perform near precipitation prediction; The splicing and fusing of the feature vector of the multi-modal meteorological live data according to the optimal dynamic learning weight coefficient comprises: The optimal dynamic learning weight coefficient is used as an initial weight coefficient of a multi-head attention mechanism to calculate a multi-head attention mechanism feature vector of cross-modal dynamic weight learning and a dynamic weight dependent coefficient; The dynamic weight dependent coefficient is normalized to obtain a dynamic weight coefficient of cross-modal learning; The feature vectors of each modality are weighted and spliced according to the dynamic weight coefficient to obtain a spliced feature vector; The optimal dynamic learning weight coefficient is: ； optimal dynamic learning weight coefficient for the i-th data channel, a heavy precipitation contribution score for the i-th data channel, denotes the number of channels; The dynamic weight dependent coefficient is: ； a dynamic weight-dependent coefficient for the i-th data channel, a key matrix for the i-th data channel, a transpose matrix of a query matrix for the i-th data channel, exp is a non-linear experience function, denotes the projection dimension of the query and the key in each head; The dynamic weight coefficient is: ； represents a nonlinear function, represents a dynamic weight coefficient of the first i channel obtained by cross-modal learning.

9. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The computer program, when loaded into the processor, realizes the near precipitation prediction method based on the heavy precipitation contribution score according to any one of claims 1-7.

Images