Efficient approaching rainfall forecasting method

Abstract

The invention discloses an efficient approaching rainfall forecasting method, and belongs to the technical field of meteorology. The method specifically comprises the following steps: firstly, acquiring radar echo data of a public meteorological platform, constructing an echo sequence data set, and dividing the echo sequence data set into a training set and a test set; then, on the basis of the state space model and the convolution model, an efficient double-branch fusion approach rainfall forecasting model is constructed; the model is composed of an encoder, a middle layer and a decoder. The middle layer adopts a state space module and a convolution module to form a double-branch architecture to realize multi-frequency feature extraction, and dynamic weight distribution of multi-scale features is realized through a feature fusion module. Thirdly, training an approaching rainfall forecasting model by using the training set; and finally, inputting radar data observed in real time to obtain a forecast result at a future moment. According to the invention, characteristics of different scales and frequencies in echo data can be effectively captured; while the parameter quantity and the calculation quantity of the model are reduced, the performance of forecasting the short temporary rainfall can be greatly improved.

Description

Technical Field

[0001] The present invention belongs to the field of meteorological technology, and particularly relates to an efficient nowcasting method for precipitation. Background Art

[0002] Nowcasting of precipitation refers to the refined prediction of rainfall intensity and spatial distribution in a specific area within the next 0 - 6 hours using existing observational data. As a core part of the meteorological disaster warning system, high-precision nowcasting of precipitation not only provides scientific decision-making support for urban governance, but also plays a crucial role in the emergency response to extreme weather events; especially in the rapid identification of severe convective precipitation processes, early warning of urban waterlogging risks, and chain prevention and control of basin flood disasters, its prediction performance is directly related to the timeliness and reliability of disaster prevention and mitigation.

[0003] Therefore, achieving effective nowcasting of precipitation is a frontier research direction in the interdisciplinary field of meteorology and data science, attracting the attention of many researchers in the academic community.

[0004] Traditional weather forecasting highly relies on numerical weather prediction. However, when facing the task of nowcasting precipitation, the prediction accuracy of this method is easily affected by the uncertainty of the initial condition field, and the model initialization and calculation inference processes are time-consuming, making it difficult to meet the requirements of refined meteorological services in terms of spatial and temporal resolutions.

[0005] Radar observation data, due to its advantages of high spatio-temporal resolution and high-precision inversion ability for the phase state and particle size distribution of precipitation particles, is widely used in nowcasting of precipitation. Traditional radar echo extrapolation methods, such as the centroid method of single entity, cross-correlation method, and optical flow method, achieve basic tracking and extrapolation of radar echoes by establishing simplified physical models, but show significant limitations in dealing with the complex non-linear characteristics in the spatio-temporal evolution process of cloud clusters.

[0006] With the booming development of deep learning, many effective radar echo extrapolation models have emerged, and the extrapolation accuracy of these models significantly exceeds that of traditional extrapolation methods. Such data-driven methods autonomously mine the non-linear mapping relationships in radar echo data through deep neural networks, thus breaking through the dependence on preset equations of traditional physical models.

[0007] However, radar echo sequences have complex non-linear dynamic characteristics such as high non-stationarity and multi-scale coupling. When facing such data, existing extrapolation models usually fall into a paradox between the accuracy and efficiency of feature representation. Lightweight models are difficult to construct a complete non-linear phase space mapping due to limited capacity, while strategies to enhance feature capture ability by stacking network depth or expanding channel dimensions not only lead to a non-linear increase in computational complexity, but also cause the problem of gradient disappearance caused by over-parameterization. Summary of the Invention

[0008] To solve the problem that it is difficult to balance the computational efficiency and prediction accuracy of existing nowcasting methods for precipitation, the present invention provides an efficient nowcasting method for precipitation, aiming to significantly improve the accuracy of nowcasting for precipitation while enhancing the computational efficiency through an advanced state space model framework.

[0009] The efficient nowcasting method for precipitation includes the following steps:

[0010] Step 1: Obtain radar echo data from the public meteorological data platform, construct an echo sequence dataset, and divide it into a training set and a test set;

[0011] Step 2: Based on the state space model and the convolutional model, construct an efficient dual-branch fusion nowcasting model for precipitation;

[0012] The model mainly consists of three parts: an encoder, an intermediate layer, and a decoder; the intermediate layer includes a state space module, a convolutional module, and a feature fusion module;

[0013] The encoder compresses the input radar echo data into the latent space through downsampling operations to obtain the spatial feature information of the radar echo image; then, the spatial feature information is split in terms of the number of channels and parallelly input into the state space module and the convolutional module of the intermediate layer to extract low-frequency and high-frequency features respectively. The feature fusion module first uses global attention and local attention to extract features from the low-frequency and high-frequency features respectively, and then fuses the frequency feature information of different scales. The data after feature fusion is input into the decoder, and through upsampling operations, the radar echo data at future times is generated according to the spatio-temporal state distribution in the latent space of the feature data, and finally, the nowcasting result for precipitation is represented through inversion transformation.

[0014] The encoder captures the spatial feature information of the radar echo image through downsampling operations, and the downsampling operations include the functions Conv2d and GroupNorm, which are expressed by the formula:

[0015] Z i =σ(GroupNorm(Conv2d(Z i-1 ))), 1≤i≤N s

[0016] where σ is the non-linear activation function SiLU, Z i-1 and Z i are the input and output of the i-th module in the encoder respectively. N s is the total number of all modules in the encoder.

[0017] The intermediate layer adopts a dual-branch structure composed of a state space module and a convolutional module to achieve multi-frequency feature extraction, and realizes the dynamic allocation of weights for multi-scale features through the feature fusion module.

[0018] The state space module utilizes a 2D scanning module constructed based on state space equations to convert image data into sequence data through a four-way scanning mechanism, and captures global low-frequency feature information by modeling long sequences.

[0019] The processing flow of the 2D scanning module includes three stages: cross scanning, selective state space module, and cross merging.

[0020] Specifically, in the cross scanning stage, a four-way path is used to scan the input spatial feature data to form four feature sequences with spatial neighborhood correlations; the selective state space module is respectively used to capture the long-distance feature dependencies of the four feature sequences, thereby obtaining four global low-frequency features; in the cross merging stage, a weighted strategy is used to fuse and reconstruct the four global low-frequency feature sequences into image data, so as to effectively retain the spatial context information in the multi-directional scanning process.

[0021] The formula of the selective state space module is expressed as:

[0022]

[0023] y t =Ch t ,

[0024] where represents the discretized state matrix, represents the discretized mapping parameter, C represents the state parameter, x t is the input feature, y t represents the output feature, h t-1 and h t represent adjacent state variables.

[0025] The convolution module utilizes a gated spatio-temporal attention mechanism to capture local high-frequency mutation features in the spatial features extracted by the encoder; the gated spatio-temporal attention is cascaded with Depth-Wise Conv, Depth-Wise Dilation Conv, and Point-WiseConv, and finally a gated mechanism is added for construction.

[0026] The formula of the gated spatio-temporal attention mechanism is expressed as:

[0027]

[0028] where Y is the input feature, is the output of the gated spatio-temporal attention unit, split refers to splitting in terms of the number of channels, g is the spatio-temporal attention coefficient, ⊙ refers to element-wise multiplication, is an intermediate variable.

[0029] The feature fusion module first uses global attention and local attention to extract features from the low-frequency features X captured by the state space branch and the high-frequency features Y captured by the convolutional branch respectively, obtaining global features and local features; then, it learns frequency feature information at different scales through the cross-fusion module, thereby improving the prediction accuracy of the model.

[0030] Among them, the local attention is expressed as:

[0031]

[0032] Among them, Z represents the input feature, δ refers to the Rectified Linear Unit (ReLU), refers to BatchNormalization, PWConv refers to Point-Wise convolution, and L represents the local feature information extracted through local attention.

[0033] The global attention compresses the H×W image to 1×1 through global average pooling, which is expressed as:

[0034]

[0035] Among them G represents the global feature information extracted through global attention.

[0036] The fusion formula for the global feature and the local feature is expressed as follows:

[0037]

[0038] Among them refers to element-wise addition, refers to channel merging.

[0039] That is, first, the global features and local features at different scales are merged in terms of the number of channels respectively.

[0040] Then, the global features and local features are added element-wise and passed through Sigmoid to obtain the weight coefficients.

[0041] Next, the weight coefficients are multiplied by the input data after channel merging to adjust the spatial distribution of the cross-frequency features.

[0042] The data after feature fusion is input into the decoder for upsampling, and the radar echo data at future times is generated according to the spatio-temporal state distribution in the latent space of the feature data. After inversion transformation, the short-term precipitation forecast results can be obtained.

[0043] The upsampling operation includes upConv2d and GroupNorm, where upConv2d contains Conv2d and Pixelshuffle, and can be specifically expressed by the formula:

[0044] Z i = σ(GroupNorm(upConv2d(Z i-1 ))),N s +N t ≤i≤2N s +N t

[0045] where N t represents the number of intermediate layers. N s is the total number of all modules in the decoder, which is the same as that of the encoder.

[0046] Step 3: Use the training set of radar echo data to train the nowcasting precipitation prediction model;

[0047] The specific training process is as follows:

[0048] After inputting the feature data set of the training data set into the nowcasting precipitation model, the mean squared error loss function continuously calculates the difference between the prediction result of the model and the label true value data, and backpropagates it to the deep learning model.

[0049] Step 4: Input the radar precipitation observation data observed at historical time T into the trained prediction model to obtain the precipitation prediction result at future time S.

[0050] The advantages and beneficial effects of the present invention are as follows:

[0051] (1) The present invention proposes to use the state space model and convolutional neural network to process radar echo data in parallel, which can effectively capture the features of different scales and frequencies in the echo data.

[0052] (2) The present invention introduces the global-local attention cross-fusion module, and through the weight allocation strategy, it can learn the spatio-temporal distribution law of different frequency features to achieve adaptive fusion.

[0053] (3) Compared with the existing methods, the present invention can significantly improve the performance of short-term and nowcasting precipitation prediction while reducing the number of model parameters and the amount of calculation. Description of the Drawings

[0054] Figure 1 is the model training flow chart for realizing nowcasting precipitation prediction using precipitation observation data according to the present invention;

[0055] Figure 2 is the overall structure diagram of the nowcasting precipitation prediction model based on the state space model and convolutional neural network according to the present invention. Detailed implementation manners

[0056] For the convenience of those of ordinary skill in the art to understand and implement the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0057] An efficient nowcasting method for precipitation of the present invention is a key challenge in the process of jointly optimizing the efficiency and accuracy in the nonlinear feature modeling of radar echo sequences. The idea is to decouple the echo sequence into high-frequency and low-frequency features, and respectively use a state space model and a convolutional neural network to fully capture the high-frequency mutations and low-frequency evolution information of the radar echo field. In addition, a feature fusion module is introduced to dynamically fuse the spatial distributions of the captured high-frequency and low-frequency feature information to improve the accuracy of the forecasting results.

[0058] The specific principle is as follows: First, construct a radar echo sequence data set required for the experiment; then, based on the basic frameworks of the state space model and the convolutional neural network, construct a dual-branch module for extracting features of different frequencies in the echo data respectively. The cross-frequency features extracted are input into the feature fusion module to dynamically model the spatio-temporal distribution laws of different-frequency features and achieve adaptive fusion. Finally, the historical observed radar echo data is input into the trained model to obtain the high-resolution nowcasting results of precipitation at future times.

[0059] The efficient nowcasting method for precipitation, as Figure 1 shown, includes the following steps:

[0060] Step 1: Obtain the radar echo data of the public meteorological data platform, construct an echo sequence data set, and divide it into a training set and a test set;

[0061] Step 2: Based on the state space model and the convolutional model, construct an efficient dual-branch fusion nowcasting model for precipitation;

[0062] The model mainly consists of three parts: an encoder, an intermediate layer, and a decoder; the intermediate layer includes a state space module, a convolutional module, and a feature fusion module;

[0063] The encoder compresses the input radar echo data into the latent space through downsampling operations to obtain the spatial feature information of the radar echo image. Then, the spatial feature information is split in terms of the number of channels and fed into the state space module and the convolutional module in the middle layer in parallel to extract low-frequency and high-frequency features respectively. The feature fusion module first uses global attention and local attention to extract features from the low-frequency features and the convolutional high-frequency features respectively, and then fuses the frequency feature information of different scales. The data after feature fusion is input into the decoder, and through upsampling operations, the radar echo data at future times is generated according to the spatio-temporal state distribution in the latent space of the feature data, and finally represents the adjacent precipitation forecast result through inverse transformation.

[0064] The encoder compresses the input radar echo data into the latent space through successive convolution and downsampling operations to extract the spatial distribution feature information of the radar echo image sequence. The downsampling operations include the functions Conv2d and GroupNorm, which are expressed by the formula:

[0065] Z i = σ(GroupNorm(Conv2d(Z i-1 ))), 1 ≤ i ≤ N s

[0066] where σ is the non-linear activation function SiLU, Z i-1 and Z i are the input and output of the i-th module in the encoder respectively. N s is the total number of all modules in the encoder.

[0067] The middle layer adopts a two-branch architecture composed of a state space module and a convolutional module to achieve multi-frequency feature extraction, and realizes the dynamic allocation of weights for multi-scale features through the feature fusion module. The dynamically adjusted weight coefficients can assist the model in learning the spatial distribution of frequency features and fuse the frequency feature information of different scales in the best way.

[0068] The state space module utilizes a 2D scanning module constructed based on the state space equation to convert the image data into sequence data through a four-way scanning mechanism, and captures the global low-frequency feature information through the modeling of long sequences; modeling low-frequency features helps to capture the macroscopic motion trajectories and structural evolution laws of convective systems, and ensures the prediction accuracy of the motion trends of large-scale weather systems.

[0069] The 2D scanning module mainly relies on the four-way scanning strategy to achieve the dynamic modeling of two-dimensional spatial features, and the processing process includes three stages: cross scanning, selective state space module, and cross merging.

[0070] Specifically, in the cross-scanning stage, a four-way path (from top left to bottom right, from bottom right to top left, from top right to bottom left, from bottom left to top right) is used to scan the input spatial feature image data, forming four feature sequences with spatial neighborhood correlations. The selective state space module is used to capture the long-distance feature dependencies of the four feature sequences respectively, thereby obtaining four global low-frequency features; in the cross-merging stage, a weighted strategy is used to fuse and reconstruct the four global low-frequency feature sequences into two-dimensional feature image data, effectively retaining the spatial context information in the multi-directional scanning process.

[0071] The formula of the selective state space module is expressed as:

[0072]

[0073] y t =Ch t ,

[0074] where represents the discretized state matrix, represents the discretized mapping parameter, C represents the state parameter, x t is the input feature, y t represents the output feature, h t-1 and h t represent adjacent state variables.

[0075] The convolution module uses the gated spatio-temporal attention mechanism to capture local high-frequency mutation features. The extraction of high-frequency features focuses on the refined evolution of local strong convection regions, which can effectively improve the model's representation ability for short-term heavy precipitation regions and enhance the forecasting accuracy of small-scale meteorological times. The gated spatio-temporal attention cascades Depth-Wise Conv, Depth-Wise Dilation Conv, and Point-Wise Conv, and finally adds a gated mechanism to construct. Through the convolution module, high-frequency features are captured from the features extracted from the encoder to the latent space.

[0076] The formula of the gated spatio-temporal attention mechanism is expressed as:

[0077]

[0078] where Y is the input feature, is the output of the gated spatio-temporal attention unit, split refers to splitting in the number of channels, g is the spatio-temporal attention coefficient, ⊙ refers to element-wise multiplication, and σ represents the Sigmoid function. is an intermediate variable.

[0079] The feature fusion module learns the spatio-temporal distribution law and fuses different frequency characteristics to achieve dynamic calibration and adaptive fusion of cross-scale spatio-temporal features. First, global attention and local attention are used to extract features from the low-frequency feature X captured by the state space branch and the high-frequency feature Y captured by the convolutional branch respectively, obtaining global features and local features. Then, the cross-fusion module learns the frequency feature information of different scales, thereby improving the prediction accuracy of the model.

[0080] Local attention focuses on the small-scale detailed image signals in the latent space, and PWConv is selected as the local channel context aggregator in this attention. Local attention is expressed as:

[0081]

[0082] where Z represents the input feature, δ refers to the Rectified Linear Unit (ReLU), refers to BatchNormalization, PWConv refers to Point-Wise convolution, and L represents the local feature information extracted by local attention.

[0083] Based on local attention, global attention adds global average pooling, focusing on the large-scale image signals in the global range of the latent space; the H×W image is compressed to 1×1 through global average pooling, effectively integrating the feature space information and comprehensively considering the feature distribution of the entire image region; it is expressed as:

[0084]

[0085] where G represents the global feature information extracted by global attention.

[0086] After global and local feature extraction of the input data respectively, global information and local information are fused through channel merging and element-wise addition, and then the weight coefficient is obtained through the Sigmoid function. This weight coefficient can learn the multi-frequency spatial distribution law of the input data. Multiply this coefficient by the data after channel merging X and Y to adjust the spatial distribution of global low-frequency and local high-frequency information. The fusion module is finally expressed as the following formula:

[0087]

[0088] where refers to element-wise addition, refers to channel merging.

[0089] The data after feature fusion is input into the decoder, and through continuous convolution and upsampling operations, the spatio-temporal features in the latent space are decoded, and the radar echo data at future moments is generated according to the spatio-temporal state distribution of the feature data. After inversion transformation, it can represent the nowcasting precipitation forecast result.

[0090] The upsampling operation includes upConv2d and GroupNorm, where upConv2d contains Conv2d and Pixelshuffle, and can be specifically expressed by the formula:

[0091] Z i =σ(GroupNorm(upConv2d(Z i-1 ))),N s +N t ≤i≤2N s +N t

[0092] where σ is the non-linear activation function SiLU, and N t represents the number of intermediate layers. N s is the total number of all modules of the decoder, which is the same as that of the encoder.

[0093] Step 3: Use the training set of radar echo data to train the nowcasting precipitation forecast model;

[0094] The specific training process is as follows:

[0095] After the feature data set of the training data set is input into the nowcasting precipitation model, the loss function continuously calculates the difference between the prediction result of the model and the label true value data, and backpropagates it to the deep learning model. Evaluate the trained model, and select the forecast model with the optimal parameters according to the evaluation results.

[0096] Step 4: Input the radar precipitation observation data observed at historical time T into the trained forecast model to obtain the precipitation forecast result at future time S.

[0097] Example:

[0098] As Figure 1 shown, the present invention is mainly divided into the following three steps:

[0099] Step (1): Collect the existing publicly available radar echo data and construct the data set required for the experiment.

[0100] The present invention collects effective radar echo data from the existing publicly available radar observation data, divides it at intervals of 2 hours, uses the data of the first 60 minutes (12 frames) as input, and predicts the changes in the next 60 minutes (12 frames); divides the obtained data into a training set, a validation set, and a test set according to a ratio, and normalizes the data, thereby constructing a data set required for model training.

[0101] Step (2): Based on the state space model and the convolutional model, construct an efficient dual-branch fusion short-term precipitation forecasting model.

[0102] As Figure 2 shown, the model uses an encoder-middle layer-decoder structure. After the input data passes through the encoder, the spatial feature information of the radar echo image is obtained, which is sent to the middle layer to capture spatio-temporal coupling features, and finally sent to the decoder for decoding to obtain the forecasting result.

[0103] The encoder and the decoder are realized by downsampling and upsampling. The middle layer is the core module, and the state space module and the convolutional module are used as a dual-branch structure to realize the dynamic coupling of spatio-temporal features. By parallel processing the data input to the middle layer, low-frequency and high-frequency features are respectively extracted; and the spatio-temporal distribution law is learned through the feature fusion module and different frequency characteristics are fused to realize the dynamic calibration and adaptive fusion of cross-scale spatio-temporal features.

[0104] Step (3): Use the radar echo data to train the short-term precipitation forecasting model. Input the radar data observed at the existing time T into the trained forecasting model, and the forecasting result at the future time S can be obtained.

[0105] The above are only the preferred embodiments of the present disclosure, and are not intended to limit the present disclosure. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present disclosure shall be included within the scope of protection of the present disclosure.

Claims

1. An efficient nowcasting method for precipitation, characterized in that: The following steps are involved: Step 1: Obtain radar echo data from the public meteorological data platform, construct an echo sequence data set and divide it into a training set and a test set; Step 2: Based on the state space model and convolution model, an efficient dual-branch fusion nowcasting precipitation forecast model is constructed; The model is mainly composed of three parts: encoder, middle layer and decoder; the middle layer includes state space module, convolution module and feature fusion module; The encoder compresses the input radar echo data into the latent space through downsampling operation to obtain the spatial feature information of the radar echo image; Then, the spatial feature information is split based on the number of channels and input in parallel to the state space module and convolution module of the middle layer to extract low-frequency and high-frequency features respectively; feature The fusion module first uses global attention and local attention to extract low-frequency features and high-frequency features respectively, and then fuses the frequency feature information of different scales; The data after feature fusion is input into the decoder. Through upsampling operation, the radar echo data of the future time is generated according to the spatiotemporal state distribution in the latent space of the feature data. After inversion conversion, it finally represents the near-term precipitation forecast result. Step 3: Use the training set of radar echo data to train the nowcasting precipitation forecast model; Step 4: Input the radar precipitation observation data observed at the historical time T into the trained forecast model to obtain the precipitation forecast result at the future time S.

2. The efficient nowcasting method for precipitation according to claim 1, characterized in that: The encoder captures the spatial feature information of the radar echo image through a downsampling operation. The downsampling operation includes functions Conv2d and GroupNorm, which are expressed as follows: WITH i =σ(GroupNorm(Conv2d(Z i-1 ))),1≤i≤N s Where σ is the nonlinear activation function SiLU, Z i-1 and Z i are the input and output of the i-th module in the encoder, N s is the total number of all encoder modules.

3. The efficient nowcasting method for precipitation according to claim 1, characterized in that: The middle layer adopts a dual-branch architecture consisting of a state space module and a convolution module to realize multi-frequency feature extraction, and realizes dynamic weight allocation of multi-scale features through a feature fusion module.

4. An efficient nowcasting method for precipitation according to claim 1 or 3, characterized in that: The state space module uses a 2D scanning module constructed based on the state space equation to convert image data into sequence data through a four-way scanning mechanism, and captures global low-frequency feature information by modeling long sequences; The processing flow of the 2D scanning module consists of three stages: cross scanning, selective state space module and cross merging; Specifically, in the cross-scanning stage, a four-way path is used to scan the input spatial feature data to form four feature sequences with spatial neighborhood correlation; the selective state space module is used to capture the long-range feature dependencies of the four feature sequences, thereby obtaining four global low-frequency features; In the cross-merging stage, a weighted strategy is used to fuse the four global low-frequency feature sequences and reconstruct them into image data, so as to effectively retain the spatial context information in the multi-directional scanning process.

5. An efficient nowcasting method for precipitation according to claim 4, characterized in that: The selective state space module formula is expressed as: in Represented as the discretized state matrix, is represented as the discretized mapping parameter, C is represented as the state parameter, and x t is the input feature, y t Denoted as output feature, h t-1 and h t Represented as adjacent state variables.

6. An efficient precipitation forecasting method as claimed in claim 1 or 3, characterized in that: The convolution module uses a gated spatiotemporal attention mechanism to capture local high-frequency mutation features in the spatial features extracted by the encoder; the gated spatiotemporal attention cascades Depth-Wise Conv, Depth-Wise Dilation Conv and Point-Wise Conv, and finally adds a gating mechanism to construct; The gated spatiotemporal attention mechanism formula is expressed as: Where Y is the input feature, is the output of the gated spatiotemporal attention unit, split refers to the split on the number of channels, g is the spatiotemporal attention coefficient, ⊙ refers to element-wise multiplication, is an intermediate variable.

7. An efficient nowcasting method for precipitation according to claim 1 or 3, characterized in that: The feature fusion module first uses global attention and local attention to extract the low-frequency feature X captured by the state space branch and the high-frequency feature Y captured by the convolution branch, respectively, to obtain global features and local features; then the cross-fusion module learns the frequency feature information of different scales, thereby improving the prediction accuracy of the model; Among them, local attention is expressed as: Where Z represents the input feature, δ refers to the Rectified Linear Unit (ReLU), refers to BatchNormalization, PWConv refers to Point-Wise convolution, and L represents the local feature information extracted by local attention; Global attention compresses the H×W image to 1×1 through global average pooling, expressed as: in G represents the global feature information extracted by global attention; The fusion formula of global features and local features is expressed as follows: in refers to element-wise addition, It refers to channel merging; That is, first, the global features and local features of different scales are merged in terms of the number of channels; Then, the global features and local features are added element by element and the weight coefficient is obtained through Sigmoid. Next, the weight coefficient is multiplied by the channel-merged input data to adjust the spatial distribution of cross-frequency features.

8. An efficient nowcasting method for precipitation according to claim 7, characterized in that: The data after the feature fusion is input into the decoder for upsampling, and the radar echo data at the future time is generated according to the spatiotemporal state distribution in the latent space of the feature data, and the near-term precipitation forecast result can be obtained through inversion conversion; The upsampling operation includes upConv2d and GroupNorm, where upConv2d includes Conv2d and Pixelshuffle, which can be expressed by the formula: WITH i =σ(GroupNorm(upConv2d(Z i-1 ))),N s +N t ≤i≤2N s +N t Where N t Indicates the number of intermediate layers, N s is the total number of all modules of the decoder, which is consistent with the encoder.

9. The efficient nowcasting method for precipitation according to claim 1, characterized in that: The specific training process of step 3 is as follows: After the feature dataset of the training dataset is input into the nearby precipitation model, the mean square error loss function continuously calculates the difference between the model's prediction results and the label true value data, and back-propagates it to the deep learning model.

Images