Thunderstorm prediction method based on lightweight three-dimensional time domain convolutional network

Abstract

The invention discloses a thunderstorm prediction method based on a lightweight three-dimensional time domain convolutional network. The method comprises the following steps: acquiring a radar echo data set; preprocessing the radar echo data set, and dividing the preprocessed data set into a training set, a verification set and a test set; constructing a 3D-TCN thunderstorm prediction network model which comprises a plurality of residual blocks containing 3D causal expansion convolution layers and is used for capturing time and space features of radar data; training the thunderstorm prediction network model by using a reverse scheduling sampling strategy, and measuring the difference between a predicted image and a real image by using a mean square error and perception loss to obtain a trained thunderstorm prediction network model; and inputting actual radar echo data, and performing prediction through the trained 3D-TCN model to obtain a thunderstorm evolution prediction result. The objective of the invention is to improve the accurate thunderstorm prediction capability of the meteorological radar. According to the method, the time sequence information and the spatial features of the radar image are combined, so that more accurate real-time thunderstorm evolution prediction is realized.

Description

Technical Field

[0001] The present invention relates to the fields of weather radar prediction and deep learning, and in particular to a thunderstorm prediction method based on a lightweight three-dimensional time-domain convolutional network. Background Art

[0002] Thunderstorms are extreme weather phenomena accompanied by lightning and heavy precipitation, which pose a serious threat to aviation, agriculture, and urban infrastructure. Therefore, accurate short-term prediction of thunderstorms is of great significance. Traditional thunderstorm prediction methods are mainly based on physical models and statistical models, which usually rely on complex modeling of atmospheric dynamics and thermodynamics. However, such methods have great limitations in capturing the nonlinear characteristics of thunderstorm generation and evolution, resulting in insufficient prediction accuracy and real-time performance. In recent years, with the development of radar observation technology and deep learning methods, thunderstorm prediction methods based on neural networks have received widespread attention. Recurrent Neural Networks (RNNs) and their variants (such as long short-term memory networks (LSTMs)) are widely used in time series data processing. However, due to the serial computing characteristics of the RNN structure, the training speed is slow, and there are problems such as gradient vanishing and gradient exploding, which limit its application in thunderstorm prediction. To address these problems, the Temporal Convolutional Network (TCN) was proposed for modeling time series data. It has parallel computing capabilities and can effectively capture long-term dependencies. In order to further improve the accuracy and efficiency of thunderstorm prediction, this paper proposes a thunderstorm prediction method based on three-dimensional time domain convolutional network (3D-TCN). This method combines the temporal and spatial features of radar echo images by introducing three-dimensional convolution operations, enhances the deep training capability of the model by using the residual structure, and improves the generalization performance of the model by using the reverse scheduling sampling strategy, ultimately achieving high-precision prediction of thunderstorm evolution. Summary of the invention

[0003] Purpose of the invention: In view of the problems of low thunderstorm prediction accuracy and poor training efficiency in the prior art, the present invention proposes a thunderstorm prediction method based on a lightweight three-dimensional time-domain convolutional network, which improves the real-time and accuracy of thunderstorm prediction.

[0004] Invention content: The thunderstorm prediction method based on lightweight three-dimensional time-domain convolutional network described in the present invention comprises the following steps:

[0005] Acquire a radar echo data set; preprocess the radar echo data set, and divide the preprocessed data set into a training set, a validation set, and a test set;

[0006] Construct a 3D-TCN thunderstorm prediction network model, including multiple residual blocks with 3D causal dilation convolutional layers, to capture the temporal and spatial characteristics of radar data;

[0007] The thunderstorm prediction network model is trained by using a reverse scheduling sampling strategy, and the difference between the predicted image and the real image is measured by using mean square error and perceptual loss to obtain a trained thunderstorm prediction network model;

[0008] The actual radar echo data is input and predicted using the trained 3D-TCN model to obtain the thunderstorm evolution prediction result.

[0009] Furthermore, the preprocessing process of the radar echo data includes quality control, image cropping, normalization, invalid image elimination, and time series synchronous extraction.

[0010] Furthermore, the 3D-TCN thunderstorm prediction network model includes an initial convolution layer, a first residual block, a second residual block, and an output convolution layer; the residual block includes a first 3D causal expansion convolution layer, a first normalization layer, a first ReLU layer, a first Dropout layer, a second 3D causal expansion convolution layer, a second normalization layer, a second ReLU layer, and a second Dropout layer residual connection.

[0011] Furthermore, the number of output channels of the residual block gradually increases.

[0012] Furthermore, the 3D causal dilated convolution is:

[0013]

[0014] Among them, Y(n,h,w) is the output tensor, which represents the feature map of the radar echo sequence after 3D convolution; n, h, w represent the index of the convolution operation in the time dimension, image height and width respectively; X(ni,hj,wl) is the input of the convolution, which is the feature map of the radar echo sequence or the last convolution operation; d is the expansion factor; i, j, l represent the number of sliding steps of the convolution kernel in time, spatial height and spatial width respectively; W(i,j,l) is the weight of the convolution kernel.

[0015] Furthermore, the implementation process of using mean square error and perceptual loss to measure the difference between the predicted image and the real image is as follows:

[0016] The mean square error MSE is as follows:

[0017]

[0018] Among them, x i Represents the pixel value of the real radar echo image, represents the pixel value of the radar echo image predicted by the model, and N represents the total number of pixels;

[0019] The perceptual loss LPIPS is as follows:

[0020]

[0021] Among them, x represents the real image, represents the generated image; φ l Represents the features extracted in the lth layer of the convolutional neural network; Indicates the difference between features; w l The weights representing the differences in features at different levels;

[0022] L total =α·L MSE +β·L LPIPS

[0023] Among them, α and β are loss weights used to balance the impact of MSE and LPIPS in the total loss.

[0024] Furthermore, the reverse scheduling sampling strategy achieves a smooth transition from complete reliance on real data to complete reliance on model predictions by gradually adjusting the probability of using real data in the training steps.

[0025] Furthermore, the residual connection includes adding the input directly to the convolution output, and when the number of channels of the input and the output are inconsistent, adjustment is performed through a 1×1 convolution.

[0026] Furthermore, the initial convolution layer uses a 3×3×3 convolution kernel, a step size of 1, and a padding of 1 to ensure that the output size is consistent with the input size.

[0027] Furthermore, the convolution kernel size of the output convolution layer is 1×1×1.

[0028] Beneficial effects: Compared with the prior art, the present invention has the following beneficial effects: 1. The present invention utilizes the advantages of causality and parallel computing of time-domain convolutional neural networks, improves training efficiency and reduces the consumption of computing resources through lightweight design and optimization of network structure, making the model more suitable for resource-constrained environments; 2. The present invention utilizes three-dimensional convolution to combine time and space characteristics, and utilizes the advantages of causality and parallel computing of time-domain convolutional neural networks to make the model more accurate in capturing the generation, development and dissipation of thunderstorms, significantly improving the prediction effect of thunderstorm evolution. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is the architecture diagram of the 3D-TCN thunderstorm prediction network proposed by the present invention;

[0030] Figure 2 It is a set of samples after data preprocessing;

[0031] Figure 3 It is the data month proportion diagram of the present invention;

[0032] Figure 4 It is the data year statistical chart of the present invention;

[0033] Figure 5 It is a diagram of the actual observation image of thunderstorm echo and the model prediction result of Example 1 of the present invention;

[0034] Figure 6 It is a CSI curve diagram of the real observation image of thunderstorm echo and the model prediction result of Example 1 of the present invention;

[0035] Figure 7 It is a diagram of the actual observation image of thunderstorm echo and the model prediction result of Example 2 of the present invention;

[0036] Figure 8 It is a CSI curve diagram of the real observation image of thunderstorm echo and the model prediction result of Example 2 of the present invention. DETAILED DESCRIPTION

[0037] The present invention will be further described below in conjunction with the accompanying drawings.

[0038] The present invention proposes a thunderstorm prediction method for meteorological radar based on a lightweight three-dimensional time-domain convolutional network, which specifically includes the following steps:

[0039] Step 1: Obtain a weather radar echo image sample dataset.

[0040] In the embodiment, the data set adopts a thunderstorm data set measured by a weather radar, and the original samples are weather radar echo images in PNG format.

[0041] The meteorological radar echo image samples in the dataset are preprocessed, including quality control, size cropping, normalization, invalid image removal, and time series synchronization extraction to obtain image sequence samples. Finally, the processed data are divided into training set, validation set and test set.

[0042] The weather radar echo images are smoothed and filtered, and a high-pass filter is constructed to eliminate the reflectivity factor values ​​that are not related to precipitation. In order to reduce the computational complexity and improve the training efficiency, the original images are cropped from 480×480 pixels each to 9 128×128 pixels.

[0043] Normalize the picture, the normalization formula is:

[0044]

[0045] Where I is the pixel value of the original grayscale image, ranging from [0,255], I norm is the normalized pixel value, ranging from [0,1]. Then, invalid images are removed from the normalized dataset, and only samples with more than 10% of valid pixels are retained.

[0046] Radar echo images are extracted from multiple folders. The images are stored in the folders as follows: under the total folder, there are three layers of folders divided according to year, month, and day. Under the day folder are 9 bottom-level folders, because the original image is divided into 9 images. Under the bottom-level folder are images named and arranged according to precise time. Use label files to distinguish between training sets and test sets, extract 20 images at a time from each bottom-level folder to form a sequence, and determine whether this sequence is a continuous sequence with an interval of 6 minutes through the image naming. If the sequence is discontinuous due to the missing of an image at a certain moment due to previous image screening, then discard this sample. The batch management mechanism is used to achieve efficient reading and use of data, thereby providing input data with time series continuity for thunderstorm prediction. Figure 2 Shown is a set of samples after preprocessing.

[0047] Step 2: Establish a 3D-TCN thunderstorm prediction network model. The model is a lightweight three-dimensional time-domain convolutional network. The network extracts the temporal and spatial features of the input sequence and outputs the predicted radar echo image sequence through the trained model.

[0048] like Figure 1 As shown, in this embodiment, the neural network includes an initial convolution layer, a first residual block, a second residual block, and an output convolution layer connected in sequence. Each residual block includes a first 3D causal expansion convolution layer, a first normalization layer, a first ReLU layer, a first Dropout layer, a second 3D causal expansion convolution layer, a second normalization layer, a second ReLU layer, and a second Dropout layer. The residual connection is that the input of the residual block does not pass through a series of internal layers, but is directly added to the results of two convolutions after dimension matching.

[0049] The initial convolution layer uses a 3×3×3 convolution kernel with a stride of 1 and padding of 1 to ensure that the output size is consistent with the input size. The number of input channels is equal to the number of channels of the image (1 grayscale channel), and the number of output channels is set to 16 to increase the feature dimension. The number of residual blocks is 2 to retain sufficient feature extraction capabilities and reduce the risk of overfitting. Each residual block consists of two 3D causal convolution layers. The convolution kernel size of the first 3D causal dilated convolution layer is 3×3×3 and the padding is 1 to maintain the spatial and temporal dimensions of the feature map. The convolution layer is followed by batch normalization to stabilize the numerical distribution during training. The subsequent activation function layer is a ReLU layer to introduce nonlinearity to enhance the feature expression ability of the model. The second 3D causal dilated convolution layer and batch normalization layer have the same processing configuration as the first convolution layer and its subsequent processing. Dimension adaptation is added between residual blocks. When the number of input and output channels is inconsistent, it is adjusted through 1×1 convolution to ensure the consistency of the dimension when the residual is added.

[0050] Residual path: Add the input directly to the features after two convolutions, and then activate it through ReLU to ensure that the dynamic range of the network output is reasonable. Multi-layer residual block stacking: The model is composed of multiple residual blocks stacked together. Each residual block extracts different levels of spatiotemporal features, gradually improving the modeling ability of the thunderstorm evolution process. The number of output channels of each residual block can be gradually increased to enhance the model's ability to express features. The convolution kernel size of the output convolution layer is 1×1×1, and its function is to compress the feature channels to the required number of output channels (for example, the number of channels for predicting future radar echo images).

[0051] The formula for the 3D causal dilated convolution operation is:

[0052]

[0053] Among them, Y(n,h,w) is the output tensor, which represents the feature map of the radar echo sequence after 3D convolution; n, h, w represent the index of the convolution operation in the time dimension, image height and width respectively; X(ni,hj,wl) is the input of the convolution, which may be the radar echo sequence or the feature map generated by the previous convolution operation; d is the dilation factor; i, j, l represent the number of sliding steps of the convolution kernel in time, spatial height and spatial width respectively; W(i,j,l) is the weight of the convolution kernel.

[0054] Step 3: Input the training set and the validation set into the thunderstorm prediction network model for training to obtain a trained thunderstorm prediction network model.

[0055] The learning rate of the hyperparameters in the lightweight 3D time-domain convolutional network training process is set to 0.001, the iteration round is set to 40000, the batch size is set to 8, the optimizer selects the Adam algorithm, and the exponential decay rate of the first-order moment estimate in the Adam algorithm is set to 0.9, and the exponential decay rate of the second-order moment estimate is set to 0.999. The loss function is the mean square error (MSE) and the perceptual loss (LPIPS), which is used to measure the difference between the predicted image and the real image, which is suitable for the prediction task of continuous values.

[0056] The formula for MSE is as follows:

[0057]

[0058] Among them, x i Represents the pixel value of the real radar echo image, It represents the pixel value of the radar echo image predicted by the model, and N represents the total number of pixels.

[0059] The formula for LPIPS is as follows:

[0060]

[0061] Among them, x represents the real image, represents the generated image; φ l Represents the features extracted in the lth layer of the convolutional neural network; Indicates the difference between features; w l The weights representing the feature differences at different levels.

[0062] The ReLU activation function commonly used in this embodiment is a common nonlinear activation function, which is widely used in deep learning. Compared with the linear activation function, ReLU has the advantages of sparsity and accelerated convergence. Sparsity makes only a few neurons respond to the input data, thereby helping the neural network to mine and learn relevant features and improve the expression and generalization capabilities of the model. At the same time, the simple calculation form of the ReLU function and the characteristics of the constant gradient in the positive interval speed up the convergence speed of the model and accelerate the training process.

[0063] The loss function of the thunderstorm prediction network model during training combines the mean square error and perceptual loss, which can be expressed by the following formula:

[0064] L total =α·L MSE +β·L LPIPS

[0065] Among them, L MSE represents the mean square error loss, which is used to measure the difference between the predicted image and the real image at the pixel level; L LPIPSrepresents the perceptual loss, which measures the perceptual similarity between the generated image and the real image through a deep neural network to improve the visual quality of the image; α and β are loss weights, which are used to balance the impact of MSE and LPIPS in the total loss. The specific weights need to be determined through experiments to strike a balance between numerical accuracy and visual quality.

[0066] Step 4: Input the test sequence samples into the trained thunderstorm prediction network model for prediction to obtain the thunderstorm prediction results of the weather radar.

[0067] This embodiment uses a thunderstorm dataset measured by a weather radar, and performs preprocessing operations to process it into a dataset suitable for training. It covers radar images collected every 6 minutes from 2009 to 2015, and these images cover the entire Hong Kong area. Given that the precipitation distribution in Hong Kong varies significantly with the seasons, especially during the rainy season (May to September), when precipitation is more concentrated, more data samples are selected in these months to ensure that the model can fully capture the precipitation characteristics of these periods. On the contrary, data sampling is reduced in months with less precipitation (October to April of the following year). Such a data selection strategy helps to improve the model's prediction accuracy for precipitation patterns in different seasons. Therefore, a total of 371 days of data from 2009 to 2015 were selected to form the dataset used in this article. The specific data distribution is as follows Figure 3 and Figure 4 shown.

[0068] In order to ensure the quality of training data, the original radar echo data set needs to be quality controlled. First, smoothing filtering is performed, and then a high-pass filter is constructed to eliminate non-precipitation related reflectivity factor values. At the same time, in order to improve the efficiency of network training, the echo images in the original data set need to be cropped into small-size images with a resolution of 128×128.

[0069] Subsequently, the cropped images need to be further screened to remove images with echo areas that are too small, because these images contribute very little to the training of the thunderstorm prediction model, so only images with valid pixel values ​​exceeding 10% are retained. Finally, a refined time series synchronization extraction technology is used to calculate the difference in the identification of the end frame index and its corresponding start frame to ensure data continuity and integrity, and obtain image sequence samples. Each sequence sample contains 20 frames of images, of which the first 10 frames are input frames, which are input into the model during training and testing, and then output prediction frames of equal length; the last 10 frames are target frames, which are used to evaluate the degree of match between the predicted frames and the actual situation.

[0070] The optimization algorithm is the Adam algorithm, which is an improvement of the stochastic gradient descent algorithm and a combination of the Momentum algorithm and the RMSprop algorithm. In the process of gradient descent, the Adam algorithm not only considers the first-order moment estimation of the gradient (i.e. the average value of the gradient), but also considers the second-order moment estimation of the gradient (i.e. the variance of the gradient).

[0071] In order to evaluate the performance of thunderstorm prediction, the present invention adopts several key evaluation indicators in the field of thunderstorm identification and prediction: Probability of Detection (POD), False Alarm Ratio (FAR) and Critical Success Index (CSI). POD evaluates the ability of the model in correctly capturing events, emphasizing the sensitivity of the forecast; FAR measures the proportion of false alarms in the forecast, focusing on the problem of excessive alarms; CSI, as a comprehensive indicator, takes into account the hit rate, missed alarms and false alarms, and provides a comprehensive evaluation of the overall effectiveness of the model. The better the prediction performance of the model, the higher the POD and CSI, and the lower the FAR. The definitions of various model evaluation indicators are listed in the following table and formula to facilitate systematic analysis and comparison of the effectiveness of different prediction models.

[0072] Table 1 Evaluation indexes of thunderstorm prediction

[0073]

[0074]

[0075] Among them, NA is the number of pixels predicted successfully; NB is the number of pixels predicted falsely; NC is the number of pixels missed; ND is the number of missed reports.

[0076] Given the complexity of extreme meteorological phenomena such as thunderstorms, the reflectivity values ​​associated with thunderstorms are roughly distributed between 30dBz and 80dBz, which is a large range. Therefore, 30dBz, 40dBz, and 50dBz are selected as key reflectivity thresholds for evaluation. In addition, the forecast time range is set to 1 hour, and a forecast image is generated every 6 minutes.

[0077] The present invention calculates an image that linearly reverse maps grayscale pixels to reflectivity factor values, so the reflectivity factor range is 0 to 80.

[0078] In order to demonstrate the superiority of the present invention, the effects of the present invention and several existing thunderstorm prediction algorithms will be compared through specific experimental data and two case studies. These comparisons will demonstrate the advantages of the present invention in thunderstorm prediction from multiple perspectives such as prediction accuracy and computational efficiency.

[0079] The results of the thunderstorm prediction scoring factor indicators of the 3D-TCN and the existing memory decoupled ST-LSTM network models in the radar echo test set are given in Tables 2 and 3 below. The results under different reflectivity thresholds when the prediction time is 6 minutes, 30 minutes and 60 minutes are selected for detailed comparison. From the data in the table, it can be seen that, except for individual data points, the 3D-TCN network has a higher detection success rate and critical success index and a lower false alarm rate under the same conditions. Therefore, the overall performance of the 3D-TCN network is better than that of the ST-LSTM+memory decoupled network. This is especially obvious when the reflectivity threshold is 30dBz. When the prediction time is 60 minutes, the improvement of the critical success index is close to 0.13.

[0080] Table 23D-TCN network test set indicator test results

[0081]

[0082] Table 3. Test results of ST-LSTM+memory decoupling network test set indicators

[0083]

[0084] In addition, it can be seen from Tables 2 and 3 that as the prediction time increases, the critical success index and detection success probability of the 3D-TCN network and the ST-LSTM+ memory decoupling network are decreasing, while the false alarm rate is increasing, indicating that the prediction performance of both is decreasing. This is because as the prediction time increases, the correlation between current and recent information and the prediction target time point gradually weakens. This makes it difficult for the model to effectively extract prediction information about future states from existing data, thus affecting the accuracy of the prediction. As the reflectivity threshold increases, the overall performance of both networks is observed to decline. This is because when the reflectivity threshold increases, the thunderstorm areas targeted by the evaluation indicators are mainly concentrated in the strong echo area. These areas usually have high echo intensity and drastic changes, and their coverage area is relatively small. Therefore, the prediction of these areas becomes more complex and difficult.

[0085] In order to compare the overall effects of more different prediction methods, the prediction evaluation indicators of the cross-correlation method, MIM network, ST-LSTM+memory decoupling network, and 3D-TCN network when the prediction time is 30 minutes are given in Table 4.

[0086] Table 4 Results of 30-minute evaluation indicators for different prediction methods

[0087]

[0088] From the data analysis in Table 4, it can be seen that the prediction performance of the three neural network algorithms exceeds that of the traditional cross-correlation method. In addition, the prediction results of the 3D-TCN network are better than those of the other two neural networks based on improved RNN.

[0089] In order to further demonstrate the performance of the present invention, the prediction results of two different dates are selected from the test set for detailed analysis and compared with the existing ST-LSTM+ memory decoupling network.

[0090] In thunderstorm example 1, the thunderstorm on that day is shown in the image as multiple extensive cluster-shaped strong echo blocks covering a large area. This analysis involves five key moments from 6 minutes to 60 minutes after the start of the forecast. Figure 5 The actual radar echo images at these five moments are displayed and compared with the predicted images generated by the memory decoupled ST-LSTM and 3D-TCN network models. From the comparison results, the distribution of strong echo areas predicted by the 3D-TCN network is consistent with the actual situation in most aspects, especially in the northwest direction, and its capture of the trend of the gradually increasing area of ​​the strong echo area shows a high degree of accuracy. In contrast, although the memory decoupled ST-LSTM is similar to the actual situation in general, the area of ​​the strong echo area it predicts is relatively large, and the reflectivity factor value is high, which causes the memory decoupled ST-LSTM to have a higher prediction false alarm rate than the 3D-TCN network, and therefore its performance in the forecast score index is not as good as the 3D-TCN network. These comparisons show that the 3D-TCN network has higher accuracy in predicting the changing trend of the strong echo area and the reflectivity factor value. In order to more clearly compare the performance of the two network models, Figure 6 The CSI curves predicted by the two models are shown. CSI has a very high reference value as a scoring indicator because it comprehensively considers the prediction accuracy, false alarm rate, and false negative rate. The closer the CSI value is to 1, the higher the accuracy of the prediction result. It can be observed from the figure that although the accuracy of the predictions of the two networks gradually decreases over time, the prediction performance of the 3D-TCN model is always slightly better than the memory decoupled ST-LSTM model. This comparison shows that 3D-TCN is more stable in maintaining long-term prediction accuracy.

[0091] In thunderstorm example 2, radar echo data from another day was selected for analysis to enhance the understanding of the model’s prediction capability and verify its applicability under different meteorological conditions. The results are shown in Figure 2. Figure 7As shown in Figure 1, the thunderstorms observed on that day were characterized by fragmented and scattered echoes, accompanied by obvious squall lines. Since the thunderstorm system on that day was more complex overall, the prediction results of both network models were worse than those of Case 1. Specifically, although both models failed to predict the slow emergence of thunderstorm blocks in the southeast, the 3D-TCN model predicted the squall line phenomenon relatively better. In contrast, the memory-decoupled ST-LSTM model failed to capture the squall line in its prediction and incorrectly predicted a large area of ​​strong echoes in the west that did not actually exist. Figure 8 The CSI curves of the two models in Example 2 are given, which show that the prediction accuracy of the two network models gradually decreases over time. However, the prediction performance of the 3D-TCN model is always better than that of the memory decoupled ST-LSTM model, and the performance gap between the two further widens in the later stages of the prediction. This phenomenon indicates that 3D-TCN may have a more stable structural advantage in terms of continuous prediction ability.

[0092] From the analysis results of the two cases, the 3D-TCN model outperforms the memory-decoupled ST-LSTM model in CSI scores under different meteorological conditions and at multiple time points. In addition, from the perspective of training efficiency, when processing data of the same size, i.e., 100 sample sequences, the training time of the memory-decoupled ST-LSTM model is usually between 4 and 6 minutes, while the training time of the 3D-TCN model is significantly reduced to no more than 1 minute. This significant difference not only highlights the advantage of the 3D-TCN model in prediction accuracy, but also shows its significant superiority in training efficiency.

[0093] The above are only preferred embodiments of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should be regarded as the protection scope of the present invention.

Claims

1. A thunderstorm prediction method based on a lightweight three-dimensional time-domain convolutional network, characterized in that: The implementation process is as follows: Acquire a radar echo data set; preprocess the radar echo data set, and divide the preprocessed data set into a training set, a validation set, and a test set; Construct a 3D-TCN thunderstorm prediction network model, including multiple residual blocks with 3D causal dilation convolutional layers, to capture the temporal and spatial characteristics of radar data; The thunderstorm prediction network model is trained by using a reverse scheduling sampling strategy, and the difference between the predicted image and the real image is measured by using mean square error and perceptual loss to obtain a trained thunderstorm prediction network model; The actual radar echo data is input and predicted using the trained 3D-TCN model to obtain the thunderstorm evolution prediction result.

2. The thunderstorm prediction method based on a lightweight three-dimensional time-domain convolutional network according to claim 1 is characterized in that: The preprocessing process of the radar echo data includes quality control, image cropping, normalization, invalid image elimination, and time series synchronous extraction.

3. The thunderstorm prediction method based on lightweight three-dimensional time-domain convolutional network according to claim 1 is characterized in that: The 3D-TCN thunderstorm prediction network model includes an initial convolution layer, a first residual block, a second residual block, and an output convolution layer; the residual block includes a first 3D causal expansion convolution layer, a first normalization layer, a first ReLU layer, a first Dropout layer, a second 3D causal expansion convolution layer, a second normalization layer, a second ReLU layer, and a second Dropout layer residual connection.

4. The thunderstorm prediction method based on lightweight three-dimensional time-domain convolutional network according to claim 1 is characterized in that: The number of output channels of the residual block gradually increases.

5. The thunderstorm prediction method based on lightweight three-dimensional time-domain convolutional network according to claim 1 is characterized in that: The 3D causal dilated convolution is: Among them, Y(n,h,w) is the output tensor, which represents the feature map of the radar echo sequence after 3D convolution; n, h, w represent the index of the convolution operation in the time dimension, image height and width respectively; X(ni,hj,wl) is the input of the convolution, which is the feature map of the radar echo sequence or the last convolution operation; d is the expansion factor; i, j, l represent the number of sliding steps of the convolution kernel in time, spatial height and spatial width respectively; W(i,j,l) is the weight of the convolution kernel.

6. The thunderstorm prediction method based on lightweight three-dimensional time-domain convolutional network according to claim 1 is characterized in that: The implementation process of using mean square error and perceptual loss to measure the difference between the predicted image and the real image is as follows: The mean square error MSE is as follows: Among them, x i Represents the pixel value of the real radar echo image, represents the pixel value of the radar echo image predicted by the model, and N represents the total number of pixels; The perceptual loss LPIPS is as follows: Among them, x represents the real image, represents the generated image; φ l Represents the features extracted in the lth layer of the convolutional neural network; Indicates the difference between features; w l The weights representing the differences in features at different levels; L total =α·L MSE +β·L LPIPS Among them, α and β are loss weights used to balance the impact of MSE and LPIPS in the total loss.

7. The thunderstorm prediction method based on lightweight three-dimensional time-domain convolutional network according to claim 1 is characterized in that: The reverse scheduling sampling strategy achieves a smooth transition from full reliance on real data to full reliance on model predictions by gradually adjusting the probability of using real data in the training step.

8. The thunderstorm prediction method based on lightweight three-dimensional time-domain convolutional network according to claim 3 is characterized in that: The residual connection includes adding the input directly to the convolution output. When the number of channels of the input and output are inconsistent, adjustment is performed through a 1×1 convolution.

9. The thunderstorm prediction method based on lightweight three-dimensional time-domain convolutional network according to claim 3 is characterized in that: The initial convolution layer uses a 3×3×3 convolution kernel, a step size of 1, and a padding of 1 to ensure that the output size is consistent with the input size.

10. The thunderstorm prediction method based on lightweight three-dimensional time-domain convolutional network according to claim 3 is characterized in that: The convolution kernel size of the output convolution layer is 1×1×1.

Images