A fusion low-level water vapor convergence flux radar echo extrapolation prediction method based on improved Swin UNet

Abstract

The application discloses a kind of fusion low-level water vapor convergence flux radar echo extrapolation forecasting methods based on improved Swin UNet, comprising the following steps: (1) collect radar echo reflectivity data, reflect water vapor convergence flux data, GNSS satellite navigation system wind field data and carry out pretreatment;Constitute fusion low-level water vapor convergence flux extrapolation forecasting data set (2) generate sample data of model input, and carry out strong radar echo sample data screening, and construct strong radar echo convection system prediction training set;(3) build the deep learning model based on improved Swin UNet model, and construct weighted loss function, different weights are respectively given to strong and weak echo area;(4) training based on the training set and model;(5) based on radar echo reflectivity data, reflect water vapor convergence flux data, satellite navigation system wind field data generate prediction product;The application improves extrapolation forecasting skill.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of radar echo extrapolation prediction, and particularly relates to a radar echo extrapolation prediction method based on improved Swin UNet and fusing low-level water vapor convergence flux. BACKGROUND

[0002] At present, the numerical prediction field is still a focus problem. Short-term weather processes are closely related to traffic safety and affect people's lives at all times. Therefore, solving the short-term prediction problem is beneficial to improving the decision-making ability of government departments and making early warning information reach the "last mile". Radar echo data has the advantages of high spatial and temporal resolution, stable real-time availability, mature denoising technology, and the ability to monitor the internal structure of high-resolution precipitation and storms in a larger range compared to other meteorological observation data. Radar echo data is an important basis for short-term prediction in daily business. Accurate radar echo extrapolation prediction is the basis for the prediction of other strong convective systems, such as heavy rain, squall lines, and thunderstorms. Strong convective systems develop very quickly, and general models are difficult to directly capture strong convective systems and accurately predict them on such a rapid time scale. Before the appearance of strong reflectivity signals on the radar picture, the convective cloud cluster has already begun to develop, which is a precursor signal of the generation of strong convective systems, providing a new solution for short-term prediction. In addition, the current short-term prediction problem is faced with the problem of insufficient input physical background information. Simply using radar data as input lacks the constraint of physical background field, and pure data-driven lacks interpretability and is difficult to obtain better prediction results, and has limited application value. SUMMARY

[0003] The purpose of the present application is to provide a radar echo extrapolation prediction method based on improved Swin UNet and fusing low-level water vapor convergence flux, to solve the problem that the current short-term prediction system is difficult to accurately predict the rapidly developing strong convective system in advance.

[0004] Technical scheme: the radar echo extrapolation prediction method based on improved Swin UNet and fusing low-level water vapor convergence flux, comprising the following steps:

[0005] (1) collect radar echo reflectivity data, water vapor convergence flux data, GNSS satellite navigation system wind field data and perform preprocessing; construct a fusion low-level water vapor convergence flux data extrapolation prediction data set;

[0006] (2) generate sample data for model input, and perform strong radar echo sample data screening to construct a strong radar echo convective system prediction training set;

[0007] (3) Build a deep learning model based on the improved Swin UNet model, and construct a weighted loss function to give different weights to strong and weak echo regions respectively;

[0008] (4) Training based on the training set and model;

[0009] (5) Based on the radar echo reflectivity data, water vapor convergence flux data, and satellite navigation system wind field data, generate prediction products.

[0010] Further, in step (1), the water vapor convergence flux data includes: ERA5 ground specific humidity q, ground relative humidity RH, low layer including 925, 850, 700 hPa specific humidity q, relative humidity RH;

[0011] Further, in step (1), the data preprocessing of time difference and spatial interpolation is performed on the multiple types of data.

[0012] Further, step (2) includes the following steps:

[0013] (21) First, generate the time series of the input model, and the time resolution of the radar echo data is every 5 minutes. A single sample first generates 100 minutes, i.e. a total of 20 radar mosaics. The first 8 frames are used for model input, i.e. the previous 40 minutes at time t, and the last 12 frames, i.e. the future 1h at time t, are used for comparison with the model output to calculate the loss function. The specific calculation formula is as follows:

[0014]

[0015] Where X t-7 ,X t-6 ,X t-5 ,X t-4 ,X t-3 ,X t-2 ,X t-1 ,X t0 are the first 8 frames of input data, f is the nonlinear mapping relationship obtained by the model, is the model prediction result for the next 5 minutes and 1h. Process other meteorological factor data into time series data corresponding to the radar data time to generate time series samples fused with low layer water vapor convergence flux data.

[0016] (22) Select the sample and construct a strong radar echo sample set. The specific method is: for each time series sample, in the last 12 frames of data, the number of pixel points with radar reflectivity exceeding 35dBz in any one frame image accounts for more than 20% of the total number of pixel points in the whole image. If yes, keep the sample, otherwise, remove the sample.

[0017] Further, step (3) includes the following steps:

[0018] (31) The neural network extrapolation prediction model based on Swin UNet is built, and the specific process is as follows: first, the input variables of each channel are divided into non-overlapping blocks by a patch segmentation block; then two consecutive SwinTransformer blocks are used to build the network bottleneck part, each feature variable is extracted and then fused in the bottleneck part, and the fused feature variables are decoded together in the decoder, which is composed of a Swin Transformer module, a patch expansion layer and a skip connection; the patch expansion layer restores the feature maps of adjacent dimensions to the original resolution, and the extracted context features are fused with the multi-scale features in the encoder through the skip connection; the skip connection is used to fuse the multi-scale features from the encoder with the up-sampled features and restore the spatial dimension of the picture through a linear layer;

[0019] (32) A weighted loss function combining mean square error (MSE) and absolute mean error (MAE) is constructed.

[0020] Further, the input variables of each channel are divided into non-overlapping blocks by a patch segmentation block, and the specific process is as follows: the 8 consecutive time series are converted into sequence embeddings, the feature dimension is changed, and the sequence blocks are projected to any dimension through a linear embedding layer, and two consecutive Swin Transformer blocks are used for self-attention calculation and transformation; the specific calculation formula is as follows:

[0021]

[0022] Wherein: Q is the query vector, which represents a vector that needs to be matched with the input sequence; K is the key vector, which represents the feature value of each element in the input sequence, V is the value vector, which represents the actual information and content of each element in the sequence, d k is the dimension of the key vector, which is used to scale the score value to avoid too large or too small value, and the SoftMax function converts the input score into a probability distribution.

[0023] Further, the Swin Transformer is composed of a normalization layer, a multi-head self-attention module, a residual connection and a two-layer multi-layer perception (MLP) with a non-linear activation function GELU; the window-based multi-head self-attention (W-MSA) module and the displacement window-based multi-head self-attention (SW-MSA) module are applied to two consecutive Transformer blocks; the specific formula is as follows:

[0024]

[0025] Wherein: z l-1 represents the feature value of the input of the previous layer, z represents the feature value after passing through the window-based multi-head self-attention module l z represents the feature value after passing through the multi-layer perception, z represents the feature value after passing through the position window multi-head attention mechanism, l+1 z represents the feature value after passing through the multi-layer perception, W-MSA represents the window-based multi-head self-attention module, MLP represents the multi-layer perception, SW-MSA represents the displacement window-based multi-head self-attention module, and LN represents the layer normalization operation.

[0026] Further, the step (32) is specifically formulaed as:

[0027]

[0028] wherein, Weighted Loss is the weighted loss, N represents the total number of pixel points of each picture, H and W respectively represent the length and width size of each picture, I t+n,i,j represents the radar reflectivity at the (i, j) position point at the time t+n moment in the forecast time n, represents the corresponding model predicted radar reflectivity value, weighted t+n,i,j represents the weight coefficient of the point.

[0029] Further, the step (5) comprises the following steps:

[0030] (51) Obtain radar echo reflectivity data and corresponding reanalysis ERA5 ground specific humidity q, ground relative humidity RH, low layer including 925, 850, 700 hPa specific humidity q, relative humidity RH, GNSS satellite navigation system wind field data, after data preprocessing, jointly constitute the input variable.

[0031] The electronic device provided by the application comprises a memory, a processor and a computer program stored in the memory and capable of running on the processor, and the computer program realizes the fusion low-level water vapor convergence flux radar echo extrapolation prediction method based on the improved Swin UNet when loaded into the processor.

[0032] The storage medium provided by the application stores a computer program, and the computer program realizes the fusion low-level water vapor convergence flux radar echo extrapolation prediction method based on the improved Swin UNet when executed by the processor.

[0033] ​Beneficial effects: compared with the prior art, the present application has the following remarkable advantages: using the fusion low-level water vapor convergence flux data, in order to promote the early monitoring of strong convective system by adding the information of physical field reflecting the generation of convective cloud, in order to further improve the extrapolation prediction skill; using the improved Swin UNet model based on the Transformer framework, the model has attention mechanism, compared with the prior art, no additional attention module is needed to play the role of feature enhancement, and the key area is adaptively focused, thereby improving the prediction ability of strong convective system. The weighted loss function is written, different weights are given in different reflectivity areas, the "smoothing" effect caused by simply using the root mean square error as the loss function is avoided, and the prediction skill of strong convective system is effectively improved. BRIEF DESCRIPTION OF DRAWINGS

[0034] Figure 1 is the general flowchart of the present application

[0035] Figure 2 is the model structure schematic diagram of the present application. DETAILED DESCRIPTION

[0036] The technical scheme of the present application will be further described below in combination with the drawings.

[0037] As Figure 1 shown, the embodiment of the present application provides a fusion low-level water vapor convergence flux radar echo extrapolation prediction method based on improved Swin UNet, which comprises the following steps:

[0038] Step 1, radar echo reflectivity data, water vapor convergence flux data including (ERA5 ground specific humidity q, ground relative humidity RH, low layer including 925, 850, 700 hPa specific humidity q, relative humidity RH), GNSS satellite navigation system wind field data), because the space-time resolution of several data is inconsistent, the space-time resolution of radar data is higher, the data preprocessing of time difference and space interpolation is needed for other data, the space-time resolution of two kinds of data is unified, normalized processing is carried out, and the data set is constructed.

[0039] The space-time interpolation method adopts the commonly used bilinear interpolation method in meteorology. The main process of bilinear interpolation is to perform linear interpolation in two directions respectively, that is, first linear interpolation in X axis and then linear interpolation in Y axis, so that one bilinear interpolation is completed. The specific formula is as follows:

[0040]

[0041] Wherein f(X a ,Y a ), f(X b ,Y a ), f(Xa ,Y b ), f(X b ,Y b ) represents the 4 directions of the wind field variable 4 points, X a , X b , Y a , Y b is the position of the variable on each coordinate axis, X represents the horizontal axis, Y represents the vertical axis, a, b represent different position coordinates, f(X r ,Y a ), f(X r ,Y b ) is the point obtained by linear interpolation in the X direction, and then linear interpolation in the Y direction to obtain the interpolated unknown point, and the specific formula is as follows:

[0042]

[0043] Where f(X r ,Y r ) is the interpolation point obtained by bilinear interpolation.

[0044] After the data space resolution is corresponded, the maximum and minimum normalized data preprocessing is carried out, and the extrapolation prediction data set of radar reflectivity data fusion of multiple meteorological elements is constructed. The maximum and minimum normalization formula is as follows:

[0045]

[0046] Where Z' is the normalized meteorological element value, Z is the original meteorological element value, Z max , Z min is the maximum and minimum value of the training set, to ensure the consistency of data normalization.

[0047] Step 2, generate sample data for model input, in order to avoid data imbalance caused by a large amount of weak echo data, strong radar echo sample data filtering is carried out, and strong radar echo convection system prediction training set is constructed:

[0048] Step 2.1: First, generate the time series of the input model, the radar echo data time resolution is every 5 min, and a single sample is generated 100 min, that is, a total of 20 radar mosaics, the first 8 frames are used for model input (i.e. 40 minutes before t time), and the last 12 frames (t time 1h in the future) are used for comparison with model output, and the loss function is calculated, and the specific calculation formula is as follows:

[0049]

[0050] Where X t-7 , X t-6 , X t-5 , X t-4X t-3 X t-2 X t-1 X t0 are the input data of the previous 8 frames, respectively, f is the nonlinear mapping relationship obtained by the model, are the model prediction results for the next 5 minutes and 1 hour, respectively. The other meteorological factor data is processed into time series data corresponding to the radar data time to generate time series samples fused with low-level water vapor convergence flux data.

[0051] Step 2.2: Screening the samples to construct a strong radar echo sample set. The specific method is as follows: for each time series sample, the proportion of the number of pixel points with radar reflectivity exceeding 35 dBz in any frame of the last 12 frames of data to the total number of pixel points in the entire image is more than 20%, then the sample is retained, otherwise the sample is excluded. To prevent data imbalance caused by a large number of weak echo sample data, which may have a small impact on the prediction results.

[0052] Step 3, build a deep learning model based on the improved Swin UNet model, and construct a weighted loss function, respectively, to give different weights to strong and weak echo regions. The specific performance is as follows:

[0053] Step 3.1, build a neural network extrapolation prediction model based on Swin UNet, as shown in Figure 2 . First, the input variables of each channel are divided into non-overlapping blocks by the patch segmentation block. This process converts 8 continuous time series into sequence embeddings, changes the feature dimension, and projects the sequence blocks to any dimension through a linear embedding layer. Two consecutive Swin Transformer blocks perform self-attention calculation and transformation. The specific calculation formula is as follows:

[0054]

[0055] Where: Q is the query vector, representing a vector that needs to be matched with the input sequence; K is the key vector, representing the feature value of each element in the input sequence, V is the value vector, representing the actual information and content of each element in the sequence, d k is the dimension of the key vector, which is used to scale the score value to avoid too large or too small values, and the SoftMax function converts the input score into a probability distribution.

[0056] The feature values and feature dimensions remain unchanged after the calculation of two consecutive Swin Transformers, and then the patch merging layer is used for down-sampling process, and the spatial resolution is reduced. In order to prevent the problem of Transformer being too deep to converge, two consecutive Swin Transformer blocks are used to construct the network bottleneck part, and each feature variable is extracted and fused in the bottleneck part. The crosstalk effect between different channels can be prevented, and the fused feature variables are then input into the decoder together. The decoding part is composed of Swin Transformer module and patch expansion layer. The patch expansion layer restores the feature maps of adjacent dimensions to the original resolution, and the extracted context features are fused with the multi-scale features in the encoder through the jump connection to make up for the spatial information loss caused by the down-sampling process. The jump connection is used to fuse the multi-scale features from the encoder with the up-sampled features to reduce the spatial loss, and the spatial dimension of the picture is restored through the linear layer.

[0057] The internal structure of Swin Transformer is as follows: Figure 2 Each block is composed of a normalization layer, a multi-head self-attention module, a residual connection and a two-layer multi-layer perception (MLP) with a nonlinear activation function GELU. The window-based multi-head self-attention (W-MSA) module and the displacement window-based multi-head self-attention (SW-MSA) module are applied to two consecutive Transformer blocks. The specific formula is as follows:

[0058]

[0059] wherein: z l-1 represents the feature value of the input of the previous layer, represents the feature value after the window-based multi-head self-attention module, z l represents the feature value after the multi-layer perception, represents the feature value after the position window-based multi-head attention mechanism, z l+1 represents the feature value after the multi-layer perception, W-MSA represents the window-based multi-head self-attention module, MLP represents the multi-layer perception, SW-MSA represents the displacement window-based multi-head self-attention module, and LN represents the layer normalization operation.

[0060] The Swin Transformer used in the feature extraction part of this model has an attention mechanism inside, which can focus on the strong echo area without adding additional attention mechanism, thereby improving the prediction ability of the strong echo system.

[0061] Step 3.2: Construct a weighted loss function combining mean square error (MSE) and mean absolute error (MAE). For radar reflectivity images, the number of low-intensity pixels is much larger than that of high-intensity pixels, and the original MAE or MSE error will pay more attention to weak echo areas or no echo areas, making it difficult to focus on the development of some small range of strong echo areas. A weighted loss function is constructed to give different weights in different reflectivity area ranges of each picture. The specific formula is:

[0062]

[0063] where Weighted Loss is the weighted loss, N represents the total number of pixels in each picture, H and W represent the length and width of each picture, respectively, I t+n,i,j represents the radar reflectivity at position (i,j) at time t+n in the forecast period n, I represents the corresponding model predicted radar reflectivity value, and weighetd t+n,i,j represents the weight coefficient of the point. The selection of the weight is shown in formula 12.

[0064] Step 4, training based on the training set and the model: adjusting the model structure and parameters (mainly including the number of SwinTransformer blocks, learning rate, optimization algorithm, and weight coefficient in the loss function), and training to obtain the best model.

[0065] Step 5, generating prediction products based on radar echo reflectivity data, water vapor convergence flux data including ERA5 ground specific humidity q, ground relative humidity RH, low layer including 925, 850, 700 hPa specific humidity q, relative humidity RH, and GNSS satellite navigation system wind field data.

[0066] Step 5.1: Obtain radar echo reflectivity data (spatial range, time resolution consistent with the training set) and corresponding reanalysis ERA5 ground specific humidity q, ground relative humidity RH, low layer including 925, 850, 700 hPa specific humidity q, relative humidity RH, and GNSS satellite navigation system wind field data. After data preprocessing, they are jointly composed of input variables.

[0067] Step 5.2: Based on the input variables and the trained model, the radar echo extrapolation prediction result of the future 1h is obtained.

[0068] The embodiment of the application also provides an electronic device, which includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the computer program is loaded into the processor, it implements any one of the fusion low-level water vapor convergence flux radar echo extrapolation prediction methods based on the improved Swin UNet.

[0069] The embodiment of the present application also provides a storage medium, which stores a computer program, and the computer program is executed by a processor to realize the fusion low-level water vapor convergence flux radar echo extrapolation prediction method based on the improved Swin UNet.

Claims

1. A method for extrapolating and predicting low-level water vapor convergence flux radar echoes based on an improved Swin UNet, characterized in that, Includes the following steps: (1) Collect radar echo reflectivity data, water vapor convergence flux data, and GNSS satellite navigation system wind field data and perform preprocessing; Construct an extrapolation forecast dataset that integrates low-level water vapor convergence flux data; (2) Generate sample data for model input and perform strong radar echo sample data screening to construct a training set for strong radar echo convection system forecasting; (3) Construct a deep learning model based on the improved Swin UNet model and build a weighted loss function, assigning different weights to the strong and weak echo regions respectively; including the following steps: (31) A neural network extrapolation prediction model based on Swin UNet is constructed as follows: First, the input variables of each channel are divided into non-overlapping blocks by patch segmentation blocks; then, two consecutive Swin Transformer blocks are used to construct the bottleneck part of the network. Each feature variable is extracted separately and then fused in the bottleneck part. The fused feature variables are then fed into the decoder for decoding. The decoding part consists of a Swin Transformer module, a patch extension layer, and a skip connection. The patch extension layer reshapes the feature maps of adjacent dimensions to their original resolution. The extracted context features are fused with the multi-scale features in the encoder through skip connections. Skip connections are used to fuse the multi-scale features from the encoder with the upsampled features and restore the spatial dimension of the image through a linear layer. (32) Construct a weighted loss function that combines mean square error (MSE) and mean absolute error (MAE); (4) Training is performed based on the training set and the model; (5) Generate forecast products based on radar echo reflectivity data, water vapor convergence flux data, and wind field data from satellite navigation systems.

2. The method for extrapolating and predicting low-level water vapor convergence flux radar echoes based on an improved Swin UNet according to claim 1, characterized in that, In step (1), the data reflecting water vapor convergence flux include: ERA5 surface specific humidity q, surface relative humidity RH, and lower layer specific humidity q and relative humidity RH at 925, 850, and 700 hPa.

3. The method for extrapolating and predicting low-level water vapor convergence flux radar echoes based on an improved Swin UNet according to claim 1, characterized in that, In step (1), data preprocessing is performed on the multi-type data using time difference and spatial interpolation.

4. The method for extrapolating and predicting low-level water vapor convergence flux radar echoes based on an improved Swin UNet according to claim 1, characterized in that, Step (2) includes the following steps: (21) First, the time series of the input model is generated. The radar echo data has a time resolution of one frame every 5 minutes. For a single sample, 100 minutes, or a total of 20 frames of radar mosaic, are generated first. The first 8 frames, i.e., the 40 minutes before time t, are used for model input. The last 12 frames, i.e., the next hour before time t, are used for comparison with the model output. The loss function is calculated. The specific calculation formula is as follows: Where X t-7 ,X t-6 ,X t-5 ,X t-4 ,X t-3 ,X t-2 ,X t-1 ,X t0 These represent the first 8 frames of input data for the variables, and f represents the nonlinear mapping relationship obtained by the model. The model forecast results for the next 5 minutes and 1 hour are obtained; other meteorological factor data are processed into time series data corresponding to the radar data time, and time series samples are generated by fusing low-level water vapor convergence flux data. (22) Screen the samples and construct a strong radar echo sample set. The specific method is as follows: for each time series sample, if the number of pixels with radar reflectivity exceeding 35dBz in any frame of the last 12 frames accounts for more than 20% of the total number of pixels in the whole image, then the sample is retained; otherwise, the sample is discarded.

5. The method for extrapolating and predicting low-level water vapor convergence flux radar echoes based on an improved Swin UNet according to claim 1, characterized in that, The input variable of each channel is segmented into non-overlapping blocks using patch segmentation, as follows: Eight consecutive time series are transformed into sequence embeddings, the feature dimension is changed, and then the sequence blocks are projected onto arbitrary dimensions through a linear embedding layer. Self-attention calculation and transformation are performed on two consecutive Swin Transformer blocks. The specific calculation formula is as follows: in; Q is the query vector, representing a vector that needs to be matched against the input sequence; K is the key vector, representing the feature value of each element in the input sequence; V is the value vector, representing the actual information and content of each element in the sequence; d k The dimension of the key vector is used to scale the score values ​​to avoid them being too large or too small. The SoftMax function transforms the input scores into a probability distribution.

6. The method for extrapolating radar echoes of fused low-level water vapor convergence flux based on an improved Swin UNet according to claim 1, characterized in that, The Swin Transformer consists of a normalization layer, a multi-head self-attention module, residual connections, and a two-layer multilayer perceptron (MLP) with a nonlinear activation function GELU. A window-based multi-head self-attention (W-MSA) module and a shift-window-based multi-head self-attention (SW-MSA) module are applied to two consecutive Transformer blocks. The specific formulas are as follows: Where: z l-1 Represents the feature values ​​input from the previous layer. z represents the feature value after passing through a window-based multi-head self-attention module. l This represents the feature values ​​after passing through a multilayer perceptron. z represents the feature value after passing through the multi-head attention mechanism of the position window. l+1 represent The feature values ​​after passing through the multilayer perceptron are: W-MSA represents a window-based multi-head self-attention module, MLP represents a multilayer perceptron, SW-MSA represents a shift-window-based multi-head self-attention module, and LN represents layer normalization operation.

7. The method for extrapolating and predicting low-level water vapor convergence flux radar echoes based on an improved Swin UNet according to claim 1, characterized in that, The specific formula for step (32) is as follows: Among them, Weighted Loss For weighted loss, N represents the total number of pixels in each image, H and W represent the length and width of each image, respectively, and I... t+n,i,j This represents the radar reflectivity at location (i,j) at time t+n, with a forecast lead time n. This represents the corresponding model-predicted radar reflectivity value, weighetd t+n,i,j This represents the weighting coefficient for that point.

8. The method for extrapolating radar echoes of fused low-level water vapor convergence flux based on an improved Swin UNet according to claim 1, characterized in that, Step (5) includes the following steps: (51) Obtain radar echo reflectivity data and corresponding reanalysis ERA5 ground specific humidity q, ground relative humidity RH, lower layer specific humidity q, relative humidity RH including 925, 850, and 700 hPa, and GNSS satellite navigation system wind field data. After data preprocessing, they together form the input variables. (52) Based on the input variables and the trained model, the radar echo extrapolation forecast results for the next 1 hour are obtained.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is loaded into the processor, it implements the radar echo extrapolation prediction method based on the improved Swin UNet according to any one of claims 1-8.

10. A storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the radar echo extrapolation prediction method based on the improved Swin UNet according to any one of claims 1-8.

Images