Short temporary rainfall forecasting method and system based on self-attention ConvLSTM generative adversarial network

Abstract

The invention discloses a short temporary rainfall forecasting method and system based on a self-attention ConvLSTM generative adversarial network, and belongs to the technical field of weather forecasting. According to the method, a self-attention mechanism and a ConvLSTM network are combined, and the precision and timeliness of rainfall forecasting are improved through multi-scale feature fusion; on a model structure, a self-attention memory model is adopted, and long-range dependence in a space-time sequence is captured by enhancing relevance between input features; in addition, a generative adversarial network framework is introduced, and the generation effect of the model is further optimized through an adversarial training mode; in order to solve the problem of mode collapse in training, a WGAN-GP algorithm is adopted, a gradient penalty term constraint discriminator is utilized, and a stable training process is kept. The method can be effectively applied to the fields of weather forecast and the like, and provides more accurate and real-time forecast support for early warning of meteorological disasters.

Description

Technical Field

[0001] The present invention relates to a short-term and imminent precipitation forecasting method, specifically a short-term and imminent precipitation forecasting method and system based on a self-attention ConvLSTM generative adversarial network, belonging to the technical field of weather forecasting. Background Technique

[0002] Short-term and imminent precipitation forecasting is used to predict precipitation conditions within the next 1 to 6 hours and is an important short-term warning means in meteorological forecasting. However, due to the high spatio-temporal variability of precipitation in the short term and local areas, traditional numerical weather forecasting methods are prone to large errors in predicting complex terrain and small-scale precipitation, making it difficult to meet the requirements of high precision and timeliness.

[0003] In recent years, the application of deep learning in the meteorological field has achieved remarkable results. Among them, the convolutional neural network (CNN) can regard meteorological data as images for spatial feature extraction, and the long short-term memory network (LSTM) is good at capturing time series dependencies. However, CNN is difficult to effectively capture long-term dependencies, and LSTM faces efficiency bottlenecks and the risk of gradient disappearance when processing long sequences. To overcome the above limitations, the self-attention mechanism is introduced to calculate the correlation between positions in the sequence and capture global dependencies, showing excellent performance in fields such as natural language processing and computer vision.

[0004] In addition, generative adversarial networks (GANs) have also begun to be applied in meteorological forecasting and can generate more realistic precipitation prediction images. However, traditional GANs are prone to mode collapse or instability during the training process.

[0005] In summary, although existing deep learning methods have made certain progress in short-term and imminent precipitation forecasting, there are still deficiencies in aspects such as model complexity, long-range dependence modeling, and the quality of generated images. Summary of the Invention

[0006] Object of the Invention: Aiming at the above problems, the object of the present invention is to provide a short-term and imminent precipitation forecasting method and system based on a self-attention ConvLSTM generative adversarial network, which comprehensively uses the self-attention mechanism, WGAN-GP, and multi-scale feature fusion technology, takes into account spatio-temporal features and global dependencies, and can generate high-quality precipitation prediction images more accurately and stably, meeting the requirements of short-term forecasting for accuracy and timeliness.

[0007] Technical Solution: On the one hand, the short-term and imminent precipitation forecasting method based on a self-attention ConvLSTM generative adversarial network provided by the present invention includes the following steps: Obtain meteorological data, perform preprocessing, and divide it into a training set and a test set according to a ratio; Construct a generative adversarial network, including a generator and a discriminator; The generative adversarial network is trained using a phased training strategy; The trained generative adversarial network is used for short-term and imminent precipitation forecasting; Among them, the generator includes a multi-scale feature fusion module, an image encoder, and an image decoder; The image encoder includes a feature extraction module and a spatio-temporal modeling module; The feature extraction module includes multiple convolutional layers; The spatio-temporal modeling module includes multiple layers of SA-ConvLSTM units.

[0008] Furthermore, the SA-ConvLSTM unit includes a self-attention module and a ConvLSTM gating structure.

[0009] Furthermore, the discriminator includes two branches. The first branch adopts a two-dimensional convolutional neural network structure, and the second branch adopts a three-dimensional convolutional structure or a temporal modeling structure.

[0010] Furthermore, the steps of training the generative adversarial network using a phased training strategy include: The first stage is the autoencoder-style pre-training stage, where the generator is trained; The second stage is the adversarial training stage, where the generator and the discriminator are alternately trained.

[0011] Furthermore, in the second stage, an asymmetric training method is used to alternately train the generator and the discriminator.

[0012] Furthermore, the loss of the generator consists of adversarial loss, reconstruction loss, and structural similarity loss. The total loss function is: , In the formula, represents the adversarial loss, represents the reconstruction loss, represents the structural similarity loss, and are the loss term weights.

[0013] Furthermore, the loss of the discriminator consists of a sampled Wasserstein distance loss and a loss based on the WGAN-GP algorithm. The total loss function is: , In the formula, represents the spatial consistency loss, represents the temporal coherence loss, and are adjustable weight coefficients.

[0014] Furthermore, the steps of using the trained generative adversarial network for short-term and nowcasting precipitation include: Taking the trained generator as the short-term and nowcasting precipitation prediction model, and using the short-term and nowcasting precipitation prediction model to perform short-term and nowcasting precipitation prediction on the meteorological data in the test set.

[0015] Furthermore, the steps of obtaining meteorological data and performing preprocessing include: Obtaining a sequence of radar reflectivity images of multiple consecutive frames, performing normalization processing on the sequence of radar reflectivity images, and scaling the pixel values of each image to a specified range; Adjusting the image size of the sequence of radar reflectivity images to a specified size.

[0016] On the other hand, the short-term and nowcasting precipitation prediction system based on the self-attention ConvLSTM generative adversarial network provided by the present invention includes: A model construction module for designing and constructing the architecture of the entire generative adversarial network, including a generator, a discriminator, and a multi-scale feature fusion module; A prediction module for performing time series prediction, generating precipitation prediction, and verifying the results through evaluation metrics; An optimization module for ensuring the stable training of the generator and the discriminator, and performing optimization by applying alternating training, WGAN-GP, and multiple loss functions.

[0017] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: 1. By combining the self-attention ConvLSTM generative adversarial network, the present invention successfully overcomes the problems of insufficient accuracy and unstable training in traditional short-term and nowcasting precipitation prediction methods; 2. By introducing a multi-scale feature fusion module, the present invention can simultaneously extract spatio-temporal features at multiple scales, significantly improving the model's ability to capture the details of precipitation distribution and global patterns, thereby improving the generation quality and prediction accuracy of precipitation images; 3. The self-attention mechanism introduced by the present invention enables the model to efficiently model long-term dependencies in spatio-temporal sequences, solving the problem of gradient disappearance or explosion that may occur in traditional LSTM when dealing with long time sequences; 4. Combining with the ConvLSTM unit, the present invention effectively extracts the dynamic information in spatio-temporal data, ensuring the accurate capture of temporal changes; 5. By using the WGAN-GP algorithm, the present invention avoids the unstable problems in the training of traditional generative adversarial networks, ensures the stable optimization of the generator and the discriminator, and generates higher-quality precipitation prediction images; 6. Overall, the method of the present invention achieves higher accuracy in short-term and nowcasting precipitation prediction, especially showing stronger robustness and adaptability under complex meteorological patterns. Description of the Drawings

[0018] Figure 1 It is a flowchart of a short-term and imminent precipitation forecasting method based on a self-attention ConvLSTM generative adversarial network; Figure 2 It is a schematic structural diagram of a multi-scale feature fusion module; Figure 3 It is a schematic structural diagram of a spatio-temporal modeling module; Figure 4 It is a schematic structural diagram of a SA-ConvLSTM cell; Figure 5 It is a schematic structural diagram of a discriminator. Detailed Implementation Manner

[0019] In order to make the objectives, technical solutions and advantages of this application clearer and more understandable, the following further elaborates on this application in conjunction with the drawings and embodiments.

[0020] Embodiment 1 The short-term and imminent precipitation forecasting method based on a self-attention ConvLSTM generative adversarial network described in this embodiment has a flowchart as Figure 1 shown, and this method includes the following steps: Step 1: Obtain meteorological data, perform preprocessing, and divide it into a training set and a test set according to a ratio.

[0021] Specifically, the steps of obtaining meteorological data and performing preprocessing include: Obtain a sequence of radar reflectivity images of multiple consecutive frames, perform normalization processing on the radar reflectivity image sequence, and scale the pixel values of each image to a specified range; Adjust the image size of the radar reflectivity image sequence to a specified size, and divide the preprocessed meteorological data into a training set and a test set according to a ratio such as 6:4. The training set is used to train the generative adversarial network, and the test set is used to detect the precipitation forecasting performance of the generator after training.

[0022] Step 2: Construct a generative adversarial network, including a generator and a discriminator.

[0023] The generator includes a multi-scale feature fusion module, an image encoder, and an image decoder; Among them, the image encoder includes a feature extraction module and a spatio-temporal modeling module; The feature extraction module includes multiple convolutional layers; The spatio-temporal modeling module includes multiple layers of SA-ConvLSTM cells.

[0024] The SA-ConvLSTM cell includes a self-attention module and a ConvLSTM gating structure.

[0025] In order to enhance the ability of the generative adversarial network to express precipitation patterns of different spatial sizes, a multi-scale feature fusion module is integrated into the generator. Figure 2 As shown in Figure 2, the multi-scale feature fusion module uses convolution kernels of different sizes, such as , , , extract the features of the initial radar echo sequence data F1 in parallel, and obtain the sub-feature maps of each branch after convolution, batch normalization, linear rectification activation, and random inactivation operations in turn. Then, these multi-scale sub-feature maps are fused by weighting or splicing to obtain a richer feature map. The fused feature map is further subjected to additional convolution operations for further high-level feature abstraction and detail enhancement, providing comprehensive and detailed information support for the subsequent encoding-decoding process of the generator. The output feature map of the multi-scale feature fusion module is denoted as F2.

[0026] The image encoder includes a feature extraction module and a spatiotemporal modeling module, whose main function is to further extract spatial features from the input historical meteorological image, i.e., the feature map F2. The feature extraction module includes multiple convolutional layers. The first layer uses Convolution maps the number of channels of the original image to the predetermined hidden dimension and uses the LeakyReLU activation function to enhance the nonlinear characteristics. The convolutional layers are gradually downsampled to obtain a low-resolution but high-dimensional feature map, denoted as F3, which provides a compact feature representation for subsequent time series modeling.

[0027] Next, based on the feature map output by the image encoder, a multi-layer SA-ConvLSTM unit is used to perform spatiotemporal modeling on the encoded feature map F3. Figure 3 The spatiotemporal modeling module shown in the figure models 1 to n layers according to actual needs, and sets the time step t from 0 to k according to the number of input sequence frames. In this example, n can be set to 3, that is, a three-layer stacked SA-ConvLSTM is used, and k is set to 14. Finally, the low-resolution feature map processed by the multi-layer SA-ConvLSTM unit is denoted as F4.

[0028] Combination Figure 4 As shown in the figure, the SA-ConvLSTM unit includes a self-attention module and a ConvLSTM gating structure, where the Q, K, and V matrices in the self-attention module are generated by the convolution layer. Each SA-ConvLSTM unit first converts the current input The hidden state at the previous time step Concatenate, pass through the convolution layer, and use group normalization to stabilize training. Then pass the convolution weight matrix of each gate , , , The pre-activation values of the four gates are calculated and split into the input gate i, forget gate f, output gate o, and candidate state g. Subsequently, according to the formula: , update the cell state and the hidden state . The updated hidden state and the memory state enter the built-in self-attention module SAM. That is, the four gates first perform standard ConvLSTM gating calculations to obtain a new hidden state . After that, this hidden state is input into SAM for global dependency extraction and update. This self-attention module does not directly rewrite the four gates or the cell state. It performs self-attention calculations on the already updated hidden state . By generating query Q, key K, and value V, it calculates the global dependencies between positions in the input data and combines this global information with local temporal features to further update the hidden state and memory state, thereby improving the ability to capture complex spatio-temporal dependencies. The hidden state output by this self-attention module is the final hidden state , and this is used as the input hidden state for the next layer / next time step.

[0029] Finally, in the image decoder, the deconvolution layer ConvTranspose2d is used to upsample the low-resolution feature map F4 to gradually restore it to the original image size. The last layer convolutional layer restores the number of channels to the number of channels of the input image, thereby generating a preliminary precipitation forecast image, and the output feature is denoted as F5.

[0030] Furthermore, the discriminator includes two branches, namely the first branch and the second branch. The first branch adopts a two-dimensional convolutional neural network structure, and the second branch adopts a three-dimensional convolutional structure or a temporal modeling structure.

[0031] Combined with Figure 5 , the discriminator adopts a spatio-temporal decoupled architecture, aiming to more accurately distinguish the images generated by the generator from real images, while taking into account both single-frame spatial features and sequence spatio-temporal coherence. First, the input image sequence F5 is processed into a tensor format suitable for 3D convolution operations and is respectively input into the two parallel branches of the discriminator. The first branch is the spatial consistency branch, and the second branch is the temporal coherence branch.

[0032] In the spatial consistency branch, the input data is processed through multiple 3D convolutional layers, corresponding to Figure 5In the Conv3D, after each layer of convolution, it is immediately followed by layer normalization and the Leaky ReLU activation function to extract and enhance local spatial features frame by frame. After extraction, the feature maps of each frame are compressed to a fixed size through adaptive average pooling, and then a single-frame authenticity score is output through a fully connected layer. The spatial consistency branch is used to evaluate the authenticity of each frame in the input image sequence, and the temporal coherence branch is used to evaluate the temporal continuity and evolution rationality between image sequences. The discrimination results of the two branches are fused as the final discrimination output to guide the generator to optimize the spatial quality and temporal consistency of the predicted images.

[0033] In the temporal coherence branch, the discriminator uses a 3D convolutional network to model the entire image sequence. This branch first passes the input spatio-temporal data through multiple 3D convolutional layers in sequence. Except for the first convolutional layer, layer normalization and the Leaky ReLU activation function are also configured after each layer of convolution to extract spatio-temporal features. Subsequently, adaptive average pooling (AdaptiveAvg Pool) is used to reduce the dimension of the convolved features to obtain a low-dimensional vector, and finally, a global temporal coherence score is output through a fully connected layer (FC).

[0034] Finally, the outputs of the two branches are fused through weighted fusion to form the comprehensive score of the discriminator, which is used to distinguish whether the input data is a real image or a generated image. This improvement not only utilizes the comprehensive expression ability of the 3D convolutional network for spatio-temporal data, but also strengthens the single-frame details and sequence dynamic information of the image through branch decoupling, thereby improving the stability and discrimination accuracy of the discriminator in adversarial training.

[0035] Step 3: Train the generative adversarial network using a staged training strategy.

[0036] Furthermore, the steps of training the generative adversarial network using a staged training strategy include: The first stage is the autoencoder-style pre-training stage, where the generator is trained. The second stage is the adversarial training stage, where the generator and the discriminator are alternately trained.

[0037] The first stage is the autoencoder-style pre-training stage. During the training of the generator, the input data first passes through a multi-scale feature fusion module and an image encoder to extract spatial features, then enters a multi-layer self-attention ConvLSTM unit for spatio-temporal modeling, captures global dependency information through the self-attention mechanism, and finally, the image decoder upsamples the low-resolution features to the predicted image.

[0038] The loss of the generator consists of adversarial loss, reconstruction loss, and structural similarity loss. The total loss function is: , In the formula, represents the adversarial loss, represents the reconstruction loss, represents the structural similarity loss, and are the loss term weights.

[0039] Among them, the adversarial loss The calculation formula is: , In the formula, is the distribution of the generated samples, is the discriminator's score for the generated samples . Among them, the reconstruction loss uses the mean squared error MSE, and the calculation formula is: , In the formula, represents the predicted image generated by the generator for the input sample , represents the input sample The corresponding true precipitation image, i.e., the label, i represents the sample index, and N represents the total number of samples in the training batch; The structural similarity loss is expressed as: , In the formula, represents the structural similarity index between the generated image and the true image .

[0040] The generator parameters are updated through backpropagation and an optimizer such as Adam, continuously reducing the loss, so that the generated image approaches the true image in visual effect and pixel accuracy, and finally achieving high-precision short-term and nowcasting precipitation forecasts.

[0041] Furthermore, the discriminator's loss consists of the sampled Wasserstein distance loss and the loss based on the WGAN-GP algorithm, and the total loss function is: , In the formula, represents the spatial consistency loss, represents the temporal coherence loss, and are adjustable weight coefficients.

[0042] In the spatial consistency branch of the discriminator, each frame in the input image sequence is processed independently. Each frame passes through multiple convolutional layers and normalization layers, combined with a non-linear activation function such as LeakyReLU, to achieve feature extraction, and a single-frame score is output by the subsequent fully connected layer. The spatial consistency loss function part adopts the WGAN-GP idea, and its expression is: , where, is the spatial branch scoring function, and represent the distributions of real images and generated images respectively, is the interpolation sample of real images and generated images; represents the discriminator for the interpolation sample gradient, represents the L2 norm of the gradient, used to measure the magnitude of the gradient; represents the expected value of the sample under the corresponding distribution, reflecting the average output of the function on this distribution; the third term is the gradient penalty term, which is used to ensure that the discriminator is a 1-Lipschitz function, is the gradient penalty term coefficient.

[0043] In the temporal coherence branch, this branch uses 3D convolution or temporal convolutional network to process the entire image sequence, and extracts the features of frame-to-frame changes and dynamic continuity in the sequence. After 3D convolutional layer, normalization and activation processing, the final output represents the score of overall temporal coherence. Its loss function expression is: , where, is the temporal branch scoring function, and represent the distributions of real images and generated images respectively.

[0044] In order to comprehensively reflect the characteristics of both space and time, in this embodiment, the losses of the spatial branch and the temporal branch are weighted and fused, and the total loss of the discriminator is defined as: , where, and are adjustable weight coefficients, and the contribution sizes of the two branches can be adjusted according to actual needs.

[0045] In the design of this discriminator, the Wasserstein distance loss is obtained by calculating the difference in the average scores of real samples and generated samples on each branch. The gradient penalty term is added to the loss formulas of the spatial and temporal branches respectively to ensure that each branch network satisfies the 1-Lipschitz condition, thereby stabilizing the training process and improving the generation effect.

[0046] Furthermore, in the second stage, an asymmetric training method is adopted to alternately train the generator and the discriminator.

[0047] In the second stage, the generator and the discriminator are optimized using an alternating training strategy, where the update frequency of the discriminator is multiple times that of the generator. Specifically, in each training cycle, the discriminator is updated multiple times first to ensure that it can accurately distinguish real images from generated images, and then the generator is updated once to improve the generated images using the discriminator's feedback. The parameter update can be expressed as: , , where, and are the parameters of the discriminator and the generator respectively, and are their respective learning rates, represents the gradient.

[0048] In the self-encoder style pre-training of this example, only the generator is trained. By inputting a sequence of real images, the goal is to minimize the reconstruction loss and the structural similarity loss of the images to initialize the parameters of the generator. The generator is trained as a self-encoder, receiving input data and attempting to reconstruct the data. By minimizing the reconstruction loss, the parameters of the generator are optimized so that it can effectively extract the features of the input data. The goal of this stage is to initialize the feature expression ability of the generator and provide good initial weights for subsequent adversarial training.

[0049] In the adversarial training strategy of the generative adversarial network, an asymmetric training method is adopted. By increasing the training frequency of the discriminator, the generator is guided to generate precipitation images with better spatial consistency and physical rationality. An alternating training strategy is carried out until the preset number of training times or convergence criteria are reached. This alternating training strategy ensures that the discriminator can always provide effective feedback to the generator, thereby accelerating the optimization process of the generator and improving the prediction accuracy and stability of the entire network.

[0050] Step 4, use the trained generative adversarial network for nowcasting of short-term precipitation.

[0051] Specifically, the steps of using the trained generative adversarial network for nowcasting of short-term precipitation include: After training, the generator is used as a short-term and nowcasting precipitation forecasting model, and the short-term and nowcasting precipitation forecasting model is used to perform short-term and nowcasting precipitation forecasting on meteorological data in the test set.

[0052] The effectiveness and superiority of the method described in the present invention are further illustrated by the following examples.

[0053] The method is experimented on the CIKM AnalytiCup 2017 short-term quantitative precipitation forecasting dataset. The evaluation of all models is shown in Table 1. The comparative models used include Convolutional Long Short-Term Memory Network (ConvLSTM), Trajectory Gated Recurrent Unit Network (TrajGRU), Predictive Recurrent Neural Network (PredRNN), and Memory-in-Memory Network (MIM). Mean Squared Error (MSE), and Heidke Skill Score (HSS) and Critical Success Index (CSI) with dBZ thresholds of 5 dBZ, 25 dBZ, and 45 dBZ respectively are used as evaluation metrics.

[0054] Table 1 Index comparison of short-term and nowcasting precipitation forecasting models

[0055] From the experimental data in Table 1, it can be seen that the method comprehensively surpasses the existing models in multiple key indicators: the MSE drops to 0.015, which is better than 0.021 of the comparative model ConvLSTM and 0.018 of the PredRNN model, indicating that the precipitation images generated by it have lower pixel-level errors; especially it performs well in strong precipitation forecasting, with the HSS reaching 0.233, and the effect is improved by 25% compared with MIM, and the CSI is 0.134, which is improved by 21.8% compared with TrajGRU, verifying the accurate capture ability of the model constructed in the present invention for strong precipitation events. At the same time, while maintaining the prediction accuracy of weak precipitation, the method significantly improves the prediction performance of medium and strong precipitation conditions, as can be seen from the data of 25 dBZ and 45 dBZ in the table. And the average performance of the method under different thresholds also shows the best performance.

[0056] Embodiment 2 The short-term and nowcasting precipitation forecasting system based on the self-attention ConvLSTM generative adversarial network described in this embodiment includes: A model construction module for designing and constructing the architecture of the entire generative adversarial network, including a generator, a discriminator, and a multi-scale feature fusion module; A prediction module for performing time series prediction, generating precipitation forecasts, and verifying the results through evaluation metrics; An optimization module for ensuring the stable training of the generator and the discriminator, and applying alternating training, WGAN-GP, and multiple loss functions for optimization.

Claims

1. A short-term precipitation forecasting method based on self-attention ConvLSTM generative adversarial network, characterized in that: The steps include: Obtain meteorological data, preprocess it, and divide it into training set and test set according to the proportion; Build a generative adversarial network, including a generator and a discriminator; A phased training strategy is used to train the generative adversarial network; Use the trained generative adversarial network to forecast short-term precipitation; Among them, the generator includes a multi-scale feature fusion module, an image encoder and an image decoder; The image encoder includes a feature extraction module and a spatiotemporal modeling module; The feature extraction module includes multiple convolutional layers; The spatiotemporal modeling module consists of multiple layers of SA-ConvLSTM units.

2. The short-term precipitation forecasting method based on the self-attention ConvLSTM generative adversarial network according to claim 1 is characterized in that: The SA-ConvLSTM unit includes a self-attention module and a ConvLSTM gating structure.

3. The short-term precipitation forecasting method based on the self-attention ConvLSTM generative adversarial network according to claim 1 is characterized in that: The discriminator includes two branches, the first branch adopts a two-dimensional convolutional neural network structure, and the second branch adopts a three-dimensional convolutional structure or a temporal modeling structure.

4. The short-term precipitation forecasting method based on the self-attention ConvLSTM generative adversarial network according to any one of claims 1 to 3 is characterized in that: The steps for training a generative adversarial network using a phased training strategy include: The first stage is the autoencoder pre-training stage, which trains the generator; The second stage is the adversarial training stage, in which the generator and the discriminator are trained alternately.

5. The short-term precipitation forecasting method based on the self-attention ConvLSTM generative adversarial network according to claim 4 is characterized in that: In the second stage, the generator and discriminator are trained alternately using an asymmetric training method.

6. The short-term precipitation forecasting method based on the self-attention ConvLSTM generative adversarial network according to claim 1 is characterized in that: The loss of the generator consists of adversarial loss, reconstruction loss and structural similarity loss, and the total loss function is: ， In the formula, Represents resistance to loss, represents the reconstruction loss, represents the structural similarity loss, and is the loss term weight.

7. The short-term precipitation forecasting method based on the self-attention ConvLSTM generative adversarial network according to claim 1 is characterized in that: The loss of the discriminator consists of the sampled Wasserstein distance loss and the loss based on the WGAN-GP algorithm. The total loss function is: ， In the formula, represents the spatial consistency loss, represents the loss of temporal coherence, and is an adjustable weight coefficient.

8. The short-term precipitation forecasting method based on the self-attention ConvLSTM generative adversarial network according to claim 1 is characterized in that: The steps of using the trained generative adversarial network to forecast short-term precipitation include: The trained generator is used as a short-term precipitation forecast model, and the short-term precipitation forecast model is used to forecast short-term precipitation on the meteorological data in the test set.

9. The short-term precipitation forecasting method based on the self-attention ConvLSTM generative adversarial network according to claim 1 is characterized in that: The steps of obtaining meteorological data and preprocessing include: Acquire a continuous multi-frame radar reflectivity image sequence, perform normalization on the radar reflectivity image sequence, and scale the pixel value of each image to a specified range; Resizes the image size of a radar reflectivity image sequence to the specified size.

10. A short-term precipitation forecasting system based on self-attention ConvLSTM generative adversarial network, characterized by: include: The model building module is used to design and build the entire architecture of the generative adversarial network, including the generator and discriminator, as well as the multi-scale feature fusion module; The prediction module performs time series prediction, generates precipitation forecasts, and verifies the results through evaluation indicators; The optimization module ensures stable training of the generator and discriminator, and applies alternating training, WGAN-GP, and multiple loss functions for optimization.

Images