A radar echo extrapolation method using three-dimensional environmental field data

Abstract

This invention discloses a radar echo extrapolation method that incorporates three-dimensional environmental field data. Radar baseline data is collected; each set of 21 consecutive radar reflectivity maps is converted into a sample, with the first 11 serving as input and the last 10 serving as labels. Based on the sample information, matching ERA5 global analysis data is collected. The radar reflectivity sequence and the corresponding environmental field are used as input to train a 3D convolutional neural network model. Experimental results show that the inclusion of the environmental field improves the extrapolation performance compared to a model without the environmental field, and is more consistent with meteorological interpretation of the extrapolation principle. By incorporating simultaneous three-dimensional environmental field data, the invention achieves accurate constraints and corrections on the extrapolation results.

Description

Technical Field

[0001] The present invention relates to the fields of meteorology and machine learning, and in particular to a method for improving radar combination reflectivity extrapolation effect by utilizing three-dimensional environmental field data. Background Art

[0002] Nowcasting refers to high-temporal and spatial resolution weather forecasts for the 0-6 hour period (with a focus on the 0-2 hour period). It is a service designed to prevent emergency localized catastrophic weather. Because catastrophic weather can develop rapidly, change in complex ways, and be highly destructive, accurate nowcasting can effectively protect economic property and human safety in areas affected by disasters, and is therefore of great significance in the field of meteorological forecasting.

[0003] Radar echo extrapolation is a common method in nowcasting. It uses information from past radar echo sequences to predict radar echo changes for a fixed time period in the future. Compared to traditional NWP (Numerical Weather Prediction) models, radar data offers greater timeliness and resolution, making radar-based extrapolation methods particularly advantageous for nowcasting within 0-2 hours. [1]

[0004] In recent years, with the rapid development of artificial intelligence and big data technology, the application of deep learning in the field of meteorology has received more and more attention and has made great progress. Compared with traditional extrapolation methods, deep learning can learn complex nonlinear meteorological processes from huge historical data. Therefore, a series of machine learning prediction models have emerged, such as Shi et al. [2] proposed a ConvLSTM model that captures spatial information through CNN and then uses LSTM to learn the correlation of radar echoes in the time dimension, which significantly improves the performance compared with the optical flow method. Du et al. [3] The creative 3D convolution model overcomes the training difficulties and other problems when using the RNN model to predict spatiotemporal sequences.

[0005] In the process of realizing the present invention, the inventors found that the prior art has at least the following shortcomings and deficiencies:

[0006] 1. Due to the short lifespan of radar echoes, the prediction performance of the deep learning-based radar echo prediction model decreases significantly as the prediction time increases, and the prediction effect is poor when the radar echo lifespan exceeds the time limit.

[0007] 2. Most existing radar echo extrapolation models learn the changing patterns of radar echoes through historical sample information, ignoring the influence of the actual environmental factor field and the complex nonlinear meteorological process. The factor environmental field is required to provide the necessary extrapolation clues.

[0008] References

[0009] [1] R.Prudden, S.Adams, D.Kangin, N.Robinson, S.Ravuri, S.Mohamed, A.Aeeibas, "A Review of Radar-Based Nowcasting of Precipitation and ApplicableMachine Learning Techniques", 2020, arXiv:2005.04988.

[0010] [2]

[0011] [3] D.Tran, L.Bourdev, R.Fergus, L.Torresani, M.Paluri, "LearningSpatiotemporal Features with 3D Convolutional Networks" Int.Conf.Comput.Vis, 2015, 4489-4497. Summary of the Invention

[0012] In order to solve the problems in the prior art, the present invention provides a radar echo extrapolation method that introduces three-dimensional environmental field data to solve the problems in the prior art such as insufficient extrapolation information sources and weak meteorological interpretation.

[0013] The technical solution of the present invention is:

[0014] A radar echo extrapolation method using three-dimensional environmental field data includes:

[0015] (1) Collect radar echo sequence base data at a specific moment and multiple three-dimensional environmental field data that match it in time and space;

[0016] (2) Convert the radar-based data from the polar coordinate system to the combined reflectivity in the Cartesian coordinate system, and interpolate the environmental field from the equal latitude and longitude grid field to the equal distance grid field;

[0017] (3) Build a radar echo sequence extrapolation neural network based on 3D convolution;

[0018] (4) Determine the training strategy and training hyperparameters.

[0019] The step (1) means that since the time resolution of the environmental field is 1 hour and the time is near the hour, the central time of the collected radar echo sequence should also be the hour, and then the corresponding environmental field is extracted based on the central time and the latitude and longitude coordinates of the radar site.

[0020] The step (2) refers to using bilinear interpolation to interpolate the data stored in polar coordinate format at 9 elevation angles of the radar into a three-dimensional grid equidistant field of 512×512×31 with a resolution of 1km×1km×0.5km; then taking the maximum basic reflectivity of different altitude layers at the same horizontal position as the combined reflectivity on the horizontal coordinate; and the environmental field data is a grid field of equal longitude and latitude, and the specific method for converting it into an equidistant field is: according to the longitude and latitude information of the radar station at the center of the grid field, calculate the longitude and latitude coordinates of each grid point of the equidistant field, and then use bilinear interpolation to interpolate the data value on each grid point.

[0021] The step (3) refers to the construction of a 3D convolutional neural network that integrates environmental field data; the network is mainly composed of four modules: a radar echo encoder, an environmental field encoder, a reasoner, and a decoder; the radar echo encoder and the environmental field encoder extract important information in the time and space dimensions of the radar echo sequence and the environmental field, and convert high-dimensional data into a low-dimensional space; the reasoner integrates the information provided by the radar and environmental field encoders and performs extrapolation prediction; the decoder restores the information provided by the reasoner into an extrapolation result with the same time and space resolution as the input radar echo sequence;

[0022] Each module consists of several convolutional layers with the same structure. After each convolution, batch normalization and activation function processing are performed. The activation function is the LeakeyRelu function. The radar echo encoder performs downsampling after each two convolutional layers, while the decoder performs upsampling. The inferencer and environment field encoder do not include upsampling or downsampling. The output of the decoder undergoes a 1x1 convolution to merge the channels to obtain the final output.

[0023] The step (4) refers to the training strategy and details:

[0024] (1) Normalize the radar data and environmental field data and then package them into a three-dimensional tensor;

[0025] (2) Load the constructed neural network and determine the training hyperparameters; select the adam optimizer to update the parameters, set the learning rate to 0.001, and the decay coefficient to (0.9, 0.999); set the number of training rounds to 70 rounds and the batch_size to 8;

[0026] (3) Use dataloader to load the pre-processed radar echo sequence and environmental field, and split the radar echo sequence into two parts: input sequence and label sequence;

[0027] (4) Feedforward operation is performed on the radar echo input sequence of the training set and the corresponding environmental field input network to obtain the extrapolated result sequence predicted by the network, and then the WSME loss between this sequence and the label sequence is calculated;

[0028] (5) Based on the calculated loss, use the BP algorithm to update the parameters in the network;

[0029] (6) Repeat steps (3), (4), and (5) until the number of training rounds reaches the preset value; then select the model with the highest average CSI in the last 30 minutes under the 40dBZ threshold for testing.

[0030] The beneficial effects of the present invention are: compared with the extrapolation model that simply uses the radar echo sequence as input, the method of the present invention provides additional constraint information, improves the extrapolation accuracy, delays the trend of the extrapolation effect decreasing sharply with the increase of extrapolation time, and enhances the interpretability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 This is a schematic diagram of the locations of 15 radar weather stations;

[0032] Figure 2 Examples of pseudo-color images of combined reflectivity obtained after converting radar-based data from Yantai Station at 13:00-14:00 on August 1, 2016 (Figures a-j: combined reflectivity images obtained every 6 minutes from 13:00 to 14:00);

[0033] Figure 3 Grayscale images of the temperature fields at seven altitude layers extracted from Yantai Station at 2:00 PM on August 1, 2016 (Figure a: temperature field at 100 hPa, Figure b: temperature field at 200 hPa, Figure c: temperature field at 300 hPa, Figure d: temperature field at 500 hPa, Figure e: temperature field at 700 hPa, Figure f: temperature field at 850 hPa, Figure g: temperature field at 925 hPa);

[0034] Figure 4 Schematic diagram of the model network structure;

[0035] Figure 5 Flowchart for model training;

[0036] Figure 6 Comparison chart of test results under different extrapolation methods (Figure a: extrapolated CSI values ​​of different methods at different times with a threshold of 20dBz; Figure b: extrapolated CSI values ​​of different methods at different times with a threshold of 30dBz; Figure c: extrapolated CSI values ​​of different methods at different times with a threshold of 40dBz; Figure d: extrapolated CSI values ​​of different methods at different times with a threshold of 50dBz). DETAILED DESCRIPTION

[0037] 1. Radar echo dataset preparation

[0038] (1) Download radar base data of 15 radar weather stations in North China and East China from 2015 to 2016. The radar station locations are as follows: Figure 1 The recording range of each radar station is 512 km × 512 km, containing data at 9 elevation angles, and is scanned once every 6 minutes on average.

[0039] (2) Using bilinear interpolation, the radar base data in the polar coordinate system is interpolated into 512×512×31 three-dimensional grid data in the Cartesian coordinate system with a resolution of 1km×1km×0.5km. Then, the maximum basic reflectivity at different height layers at the same horizontal position is taken as the combined reflectivity of the point, such as Figure 2 shown.

[0040] (3) All hourly time data are recorded. If the time between a certain hourly time data and the two adjacent hourly time data is complete, that is, it contains 21 time data in total, then the recording cycle is complete. These 21 data are packaged as a sample. The middle hourly time and the station name are combined as the sample name. For example, the sample record of Qingdao station at 20:00 on October 1, 2015 is 2015_10_01_20_qingdao.

[0041] (4) Screening samples with higher combined reflectivity: If there are more than 600 points with a reflectivity greater than 40 dBz or more than 40,000 points with a reflectivity greater than 25 dBz at more than 10 moments in a sample, the sample is retained; otherwise, it is discarded. A total of 8,979 samples were obtained after screening.

[0042] (5) The data were randomly divided into training set, validation set and test set according to the ratio of 4:1:1, and finally 5987 training set data, 1496 validation set data and 1496 test set data were obtained.

[0043] 2. Preparation of environmental field dataset.

[0044] (1) Download the full-time ERA5 global analysis data for China from 2015 to 2016. The data has a temporal resolution of 1 hour and a spatial resolution of 0.25° × 0.25°, and contains seven altitude layers. The environmental factors mainly include ten categories: divergence, gravitational potential energy, potential vorticity, relative humidity, specific humidity, temperature, wind U component, wind V component, vertical velocity, and relative vorticity.

[0045] (2) Extract the corresponding environmental field according to the radar echo sample name, and the extraction range is a 1024km×1024km square area centered on the radar site. Since ERA5 provides a grid field of equal longitude and latitude, it cannot be aligned with the equidistant grid field provided by the radar. Therefore, it is necessary to first convert the environmental field into an equidistant grid field. The specific method is: according to the longitude and latitude of the radar site at the center of the equidistant grid field, calculate the longitude and latitude values ​​of each grid point in the corresponding equidistant field, and then use bilinear interpolation to interpolate the specific values ​​of each grid point in the equidistant field. Finally, the 0.25°×0.25° equidistant grid field is converted to a 32km×32km equidistant grid field with a size of 32×32. The grayscale images of the environmental fields of the seven altitude layers after extraction are shown as follows: Figure 3 shown.

[0046] 3. Model network construction

[0047] Taking into account factors such as training difficulty and computational resources, this method selects 3D convolution as the extrapolation model. Traditional CNN models mostly use 2D convolution, where the depth of the convolution kernel is equal to the number of channels in the input image. Therefore, regardless of the number of input images, 2D convolution ultimately outputs only one feature map. The model can only learn spatial feature information, while information in the temporal dimension is overwritten. Compared with traditional 2D convolution, the convolution kernel of 3D convolution adds the temporal dimension, allowing the convolution kernel to perform convolution not only in the spatial dimension but also along the temporal dimension, thereby learning the temporal correlation between radar echoes at different times.

[0048] The entire network can be divided into four sub-networks: radar echo field encoder, environmental field encoder, inference unit, and decoder. Each sub-network is composed of stacked convolution blocks with the same structure. Each convolution block includes a three-dimensional convolution, followed by a batch normalization, and finally output through an activation function. The main function of batch normalization is to pull a group of eigenvalues ​​with large distribution differences back to the standard normal distribution. In this way, the eigenvalues ​​will fall into the range where the activation function is more sensitive to the input, avoiding gradient vanishing and accelerating the convergence speed. The activation function uses the LeakeyRelu function. Compared with the sigmoid function and tanh function, this function has no saturation range, which solves the problem of gradient vanishing. Compared with the Relu function, it still has a smaller gradient when the input is less than 0, avoiding inactivation when the input is less than 0.

[0049] The radar echo field encoder primarily captures spatial and temporal information about the input echo sequence, encoding high-dimensional spatiotemporal data into a low-dimensional space. During the encoding process, the number of feature maps increases after each two convolutional blocks, followed by a spatial average pooling operation, halving the spatial resolution of the feature map. Two convolutional blocks and one average pooling layer constitute an encoding block. The entire encoder network is composed of several stacked encoding blocks, with the pooling in the final encoding block also reducing temporal resolution.

[0050] The ambient field encoder performs a similar function to the radar echo encoder, differing in that the ambient field has a greater temporal resolution. Therefore, the ambient field encoder must first perform linear interpolation in the temporal dimension to align the ambient field with the radar echo. Secondly, because the spatial range represented by the ambient field is wider than that of the radar, pooling methods cannot effectively integrate environmental information outside the radar detection range into the extrapolated spatial range. Therefore, large convolution kernels are used instead of pooling layers to reduce the spatial dimensionality.

[0051] After the radar echo sequence and the ambient field sequence pass through their respective encoders, they generate a series of feature maps of equal spatial dimensions. These maps are stacked in the channel dimension and input into the inference network. The inference network primarily learns the correlation between the radar echo and the ambient field and fuses their features. The inference network consists of a series of convolutional blocks, does not include pooling operations, and maintains the same number of feature layers.

[0052] The decoder's task is to restore the feature maps received from the inference engine into an extrapolated sequence with the same spatial resolution as the input sequence. The decoder structure is essentially the same as the encoder, differing in that the decoder replaces the encoder's average pooling operation with upsampling, and the number of feature layers decreases with each decoding block. The entire decoder network consists of a stack of decoding blocks, with the pooling in the first decoding block also restoring temporal resolution.

[0053] The model network diagram is as follows Figure 4 As shown in Figure 1, the entire network consists of four parts: a radar echo encoder, an ambient field encoder, an inference unit, and a decoder. The radar echo sequence and ambient field are input into the radar echo encoder and the ambient field encoder, respectively, which then output several layers of feature maps. The radar echo feature maps and the ambient field feature maps are stacked along the channel dimension and input into the inference unit, which fuses the information extracted from the radar echo sequence and the ambient field. The feature maps transformed by the inference unit are then fed into the decoder, which ultimately restores the feature maps to an extrapolated result with the same resolution as the input radar echo.

[0054] The radar echo encoder consists of four encoding blocks, each of which includes two convolutions and one average pooling operation. After each convolution operation, batch normalization and an activation function are applied. The activation function uniformly uses LeakeyReLU with a slope of 0.1. The number of feature layers increases with each encoding block. The ambient field encoder first performs upsampling in the temporal dimension, followed by three convolution operations, the second of which uses a large convolution kernel to incorporate peripheral information.

[0055] Compared to the encoder, the inferencer discards the pooling operation while maintaining the same number of feature layers. This is primarily to fuse radar echo information with environmental field information. The inferencer output is fed into the decoder, which has a similar structure to the radar echo encoder, differing in that it replaces average pooling with upsampling. The specific network parameters are shown in Table 1:

[0056] Table 1 Parameters of each layer of the neural network

[0057] Module Convolution kernel size filling step length Output size Radar_encoder1 4x3x3 1 1 32x10x128x128 Radar_encoder2 3x3x3 1 1 64x10x64x64 Radar_encoder3 3x3x3 1 1 128x10x32x32 Radar_encoder4 3x3x3 1 1 256x5x16x16 Env_encoder1 3x3x3 1 1 32x5x32x32 Env_encoder2 1x17x17 0 1 64x5x16x16 Env_encoder3 3x3x3 1 1 128x5x16x16 5*Inference 3x3x3 1 1 256x5x16x16 Decoder1 3x3x3 1 1 128x10x32x32 Decoder2 3x3x3 1 1 64x10x64x64 Decoder3 3x3x3 1 1 32x10x128x128 Decoder4 3x3x3 1 1 32x10x256x256 To_img 3x3x3 1 1 1x10x256x256

[0058] 4. Model Training

[0059] Model training mainly includes data preprocessing, selection of loss function and optimizer, determination of learning rate and number of training rounds, etc.

[0060] (1) Data preprocessing: The main preprocessing method adopted is normalization. Since the intensity of most recorded radar echoes is less than 70dBz, the data in the radar echo sequence is divided by 70 to scale the data between 0 and 1. However, the distribution of data in different environmental fields varies greatly, so Z-score normalization is used to normalize the data of different types of environmental fields to a Gaussian distribution of N(0,1).

[0061] (2) Loss function: In radar extrapolation results, high reflectivity areas often indicate a greater likelihood of severe weather in those areas. However, since the probability of radar detecting high reflectivity is very low, if the MSE loss function is used alone for training, the model may ignore the extrapolated results of high reflectivity areas. Therefore, the weighted MSE loss function (WMSE) is used to train the model. Higher reflectivity areas are assigned higher weights, allowing the model to focus more on these areas.

[0062] (3) Optimizer: Since the Adam optimizer can adaptively change the learning rate and is a momentum-based algorithm that is easy to escape from local minima, the Adam optimizer is used to train the model.

[0063] (4) Training process: The pre-processed radar echo sequence and the environmental field are input into the network for forward operation. The network finally outputs the extrapolated result. The WMSE function is used to calculate the loss between the extrapolated result and the true result. The BP algorithm is then used to update the network weights.

[0064] The flow chart of model training is as follows Figure 5 The number of training rounds is set to 70. Since we want to highlight the advantages of the model in predicting severe weather, areas with larger combined reflectivity need to be given more attention. Therefore, the loss function uses the weighted MSE function, and the weight is set to:

[0065]

[0066] Because the Adam optimizer can adaptively change the learning rate and is a momentum-based algorithm that easily escapes local minima, we used it to train the model. The learning rate was set to 0.001, and the decay coefficients were set to (0.9, 0.999). To ensure training efficiency and accelerate convergence, the batch_size was set to 8.

[0067] Training process: A batch of pre-processed radar echo sequences and environmental fields are fed into the network for forward computation. The network then outputs an extrapolated result. The loss between the extrapolated result and the true result is calculated using the WMSE function. The network weights are then updated using the backpropagation algorithm. The next batch of radar echo sequences and environmental fields is then fed into the network, and the above steps are repeated until the preset number of training rounds is reached.

[0068] 5. Model Evaluation

[0069] In order to verify that the extrapolation effect after adding the environmental field is better than that without the environmental field, it is necessary to reasonably evaluate the model prediction effect. The meteorological forecast indicators we introduced mainly include: hit rate (POD), false alarm rate (FAR), critical success index (CSI), fair skill score (ETS) and Heidke skill score (HSS). Before calculating the score, a threshold needs to be set. Points greater than the threshold are recorded as positive samples, and points less than the threshold are recorded as negative samples. The calculation formulas for each score are as follows:

[0070] Table 2 Classification results confusion matrix

[0071] Predicted as positive sample Predicted as negative sample True positive sample TP FN True negative sample FP TN

[0072]

[0073]

[0074]

[0075]

[0076]

[0077]

[0078] We selected the model with the highest average CSI in the last 30 minutes when the threshold was 40dBZ on the validation set as the test model. The test results of the optical flow method, the 3D convolution model without the environment field, and the 3D convolution model with the optimal environment field in each extrapolation period are as follows: Figure 6 The average results for all time periods at different thresholds are shown in the table below. As can be seen from the figure, at thresholds of 20, 30, and 40 dBZ, the inclusion of certain environmental fields improves the model's extrapolation performance to a certain extent. The larger the threshold, the more pronounced the improvement, extending the effective time of extrapolation to a certain extent. However, a threshold of 50 dBZ shows no significant effect. This is presumably because the proportion of grid points above 50 dBZ is too small, and the WMSE weights grid points above 50 dBZ too small, making it difficult for the model to learn the changing patterns of these points.

[0079] Table 3 Average CSI (validation set) with different environment fields and different thresholds

[0080]

[0081] (Note: Bold font indicates the optimal value of the indicator, and underlined font indicates the suboptimal value of the indicator)

[0082] Table 4 Average indicators of different methods at all times when the threshold is 20dBZ (test set)

[0083]

[0084] (Note: Bold font indicates the optimal value of the indicator, and underlined font indicates the suboptimal value of the indicator)

[0085] Table 5 Average indicators of different methods at all times when the threshold is 30dBZ (test set)

[0086]

[0087] (Note: Bold font indicates the optimal value of the indicator, and underlined font indicates the suboptimal value of the indicator)

[0088] Table 6 Average indicators of different methods at all times when the threshold is 40dBZ (test set)

[0089]

[0090]

[0091] (Note: Bold font indicates the optimal value of the indicator, and underlined font indicates the suboptimal value of the indicator)

[0092] Those skilled in the art will appreciate that the accompanying drawings are merely schematic diagrams of a preferred embodiment. The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A radar echo extrapolation method using three-dimensional environmental field data, characterized in that: include: (1) Collect radar echo sequence base data at a specific moment and multiple three-dimensional environmental field data that match it in time and space; (2) Convert the radar-based data from the polar coordinate system to the combined reflectivity in the Cartesian coordinate system, and interpolate the environmental field from the equal latitude and longitude grid field to the equal distance grid field; (3) Build a radar echo sequence extrapolation neural network based on 3D convolution; (4) Determine the training strategy and training hyperparameters.

2. The radar echo extrapolation method using three-dimensional environmental field data according to claim 1, characterized in that: The step (1) means that since the time resolution of the environmental field is 1 hour and the time is near the hour, the central time of the collected radar echo sequence should also be the hour, and then the corresponding environmental field is extracted based on the central time and the latitude and longitude coordinates of the radar site.

3. The radar echo extrapolation method using three-dimensional environmental field data according to claim 1, characterized in that: The step (2) refers to using bilinear interpolation to interpolate the data stored in polar coordinate format at 9 elevation angles of the radar into a three-dimensional grid equidistant field of 512×512×31 with a resolution of 1km×1km×0.5km; then taking the maximum basic reflectivity of different altitude layers at the same horizontal position as the combined reflectivity on the horizontal coordinate; and the environmental field data is a grid field of equal longitude and latitude, and the specific method for converting it into an equidistant field is: according to the longitude and latitude information of the radar station at the center of the grid field, calculate the longitude and latitude coordinates of each grid point of the equidistant field, and then use bilinear interpolation to interpolate the data value on each grid point.

4. The radar echo extrapolation method using three-dimensional environmental field data according to claim 1, characterized in that: The step (3) refers to the construction of a 3D convolutional neural network that integrates environmental field data; the network is mainly composed of four modules: a radar echo encoder, an environmental field encoder, an inference unit, and a decoder; the radar echo encoder and the environmental field encoder extract important information in the time and space dimensions of the radar echo sequence and the environmental field, and convert high-dimensional data into a low-dimensional space; the inference unit integrates the information provided by the radar and the environmental field encoder and performs extrapolation prediction; The decoder restores the information provided by the reasoner into an extrapolated result with the same temporal and spatial resolution as the input radar echo sequence; Each module consists of several convolutional layers with the same structure. After each convolution, batch normalization and activation function processing are performed. The activation function is the Leakey ReLU function. The radar echo encoder performs downsampling after each two convolutional layers, while the decoder performs upsampling. The inferencer and environment field encoder do not include upsampling or downsampling. The output of the decoder undergoes a 1x1 convolution to merge the channels to obtain the final output.

5. The radar echo extrapolation method using three-dimensional environmental field data according to claim 1, characterized in that: The step (4) refers to the training strategy and details: (1) Normalize the radar data and environmental field data and then package them into a three-dimensional tensor; (2) Load the constructed neural network and determine the training hyperparameters; select the adam optimizer to update the parameters, set the learning rate to 0.001, and the decay coefficient to (0.9, 0.999); set the number of training rounds to 70 rounds and the batch_size to 8; (3) Use dataloader to load the pre-processed radar echo sequence and environmental field, and split the radar echo sequence into two parts: input sequence and label sequence; (4) Feedforward operation is performed on the radar echo input sequence of the training set and the corresponding environmental field input network to obtain the extrapolated result sequence predicted by the network, and then the WSME loss between this sequence and the label sequence is calculated; (5) Based on the calculated loss, use the BP algorithm to update the parameters in the network; (6) Repeat steps (3), (4), and (5) until the number of training rounds reaches the preset value; Then, the model with the highest average CSI in the last 30 minutes under the 40dBZ threshold is selected for testing.

Images