A short-term extreme precipitation prediction method based on abnormal driving residual dynamic diffusion

Abstract

This invention discloses a short-term extreme precipitation prediction method based on anomaly-driven residual dynamic diffusion, comprising: acquiring a continuous radar echo image sequence of a target area, inputting it into an anomaly hotspot extraction module, using a student-teacher network to learn normal weather characteristics, and combining Gaussian temporal weighted attention with multi-resolution feature fusion to generate anomaly hotspot images; performing attention analysis on the anomaly hotspot images to obtain a spatial weight matrix, and calling an image dynamic regeneration module to perform diffusion enhancement on historical and current data; inputting the enhanced samples into a weighted residual long short-term memory network, fusing the spatial weight matrix in a forget gate to guide attention to anomaly regions, and introducing residual connections between units to output radar echo prediction images for multiple future time periods, thereby providing extreme precipitation evolution results and early warning information; this invention improves the ability to focus on anomaly regions and the stability of prediction.

Description

Technical Field

[0001] This invention relates to the field of meteorological radar echo extrapolation technology, specifically to a method for predicting short-term extreme precipitation based on anomaly-driven residual dynamic diffusion. Background Technology

[0002] Extreme precipitation events are characterized by their suddenness, high uncertainty, and significant destructiveness. In short-term forecasting tasks, they typically exhibit scarce samples, rapid evolution, and complex spatial distribution. Existing short-term extreme precipitation prediction methods usually rely on extensive historical meteorological data for training, making them highly dependent on anomalous samples and computational resources, and struggling to balance timely warnings with accurate predictions of future evolution. Anomaly detection methods can learn the distribution of normal samples even in the absence of anomalous samples, thereby identifying anomalous samples that deviate from the normal state. Student-Teacher Networks (STNs) can identify anomalous samples by learning the differences in features of normal samples through Teacher Networks and Student Networks, thus reducing reliance on anomalous data. However, simple anomaly detection models are mainly used for early warning and lack the ability to continuously simulate the dynamic evolution of future weather images. Furthermore, traditional time-series forecasting models typically require incremental training for fine-tuning when facing the latest weather changes, but insufficient incremental training data can easily lead to overfitting, resulting in a decline in predictive ability for rapidly changing radar echo images. Summary of the Invention

[0003] Purpose of the invention: The purpose of this invention is to provide a short-term extreme precipitation prediction method based on anomaly-driven residual dynamic diffusion, which utilizes anomaly hotspot information, enhances time-series prediction capabilities, and takes into account the robustness of short-term extreme precipitation prediction through small sample incremental training, thereby solving the problems existing in the background technology.

[0004] Technical solution: The present invention provides a method for predicting short-term extreme precipitation based on anomaly-driven residual dynamic diffusion, comprising the following steps:

[0005] (1) Obtain radar echo image sequences of the target area at multiple consecutive times, and construct the input image sequence according to the preset time window;

[0006] (2) Input the input image sequence into the abnormal hotspot extraction module, utilize the feature expression differences between the pre-trained teacher network and the student network to be trained, combine the Gaussian temporal weighting mechanism to enhance the spatial attention map, and generate abnormal hotspot images representing potential abnormal regions by multi-resolution fusion of feature maps output by different convolutional layers.

[0007] (3) Perform attention analysis on the abnormal hotspot images to generate a spatial attention weight matrix to indicate the degree of abnormality in each pixel region;

[0008] (4) Call the image dynamic regeneration module, and perform a regeneration process of gradually adding noise and then removing noise from the original image to enhance the historical time window data and the current abnormal weather data to obtain enhanced training samples.

[0009] (5) Input the enhanced training samples into the residual time-series prediction module, fuse the spatial attention weight matrix into the forget gate, guide the forget gate to selectively discard information according to the importance of the spatial abnormal region, and introduce residual connections between the residual time-series prediction modules to pass the original output information of the previous unit to the next unit in order to generate radar echo prediction image sequences for multiple future time steps.

[0010] (6) Output the future evolution results and early warning information of extreme precipitation based on radar echo prediction image sequences at multiple future time steps.

[0011] Furthermore, in step (2), the Gaussian temporal weighting mechanism enhances the spatial attention map as follows: adaptive pooling and upsampling are performed on the input image sequence to align the spatial dimensions, spatial features are extracted through convolutional layers to generate a spatial attention map, and the spatial attention map is temporally weighted using Gaussian distribution weights centered on the current time step, so that the model gives higher attention to the image at the new time step, and the weighted spatial attention map is multiplied element-wise with the original input feature map to obtain the output feature map that integrates temporal attention.

[0012] Furthermore, in step (2), multi-resolution fusion specifically involves: extracting feature maps from multiple convolutional layers of different depths in the student network, and then adding and fusing the low-level feature maps representing local edge textures with the high-level feature maps representing global structural semantics to obtain image information at different scales.

[0013] Furthermore, in step (4), the image dynamic regeneration module uses a diffusion model to perform the regeneration process. The forward process is to gradually add random noise to the original image, and the reverse process is to gradually predict and remove noise from the noisy image to restore the image. The training objective is composed of the noise prediction loss and the prediction loss of the residual temporal prediction module, so that the generated enhanced image maintains the consistency of the overall structure.

[0014] Furthermore, in step (5), the residual time-series prediction module is a long short-term memory network.

[0015] Furthermore, when calculating the activation value of the forget gate, the weighted value of the current input, the hidden state of the previous time step, and the spatial attention weight matrix is ​​used as the input of the activation function, so that the output of the forget gate depends on both the temporal features and the importance of the spatially anomalous region.

[0016] Furthermore, the residual connection is specifically as follows: the gated output of the current network unit is added to and fused with the hidden state of the previous time step or the hidden state of the previous network unit, and the fusion result is used as the final hidden state output of the current unit.

[0017] The present invention discloses a short-term extreme precipitation prediction system based on anomaly-driven residual dynamic diffusion, comprising:

[0018] Data acquisition module: used to acquire radar echo image sequences of the target area at multiple consecutive time points;

[0019] Anomaly hotspot extraction module: This module utilizes the feature representation differences between the pre-trained teacher network and the student network to be trained, combined with a Gaussian temporal weighting mechanism and a multi-resolution feature fusion structure, to generate anomaly hotspot images from radar echo image sequences.

[0020] Spatial weight generation module: used to perform attention analysis on anomalous hotspot images and generate a spatial attention weight matrix;

[0021] Image dynamic regeneration module: used to perform a regeneration process of progressively adding noise and then reverse denoising on historical time window data and current abnormal weather data to generate enhanced training samples;

[0022] The residual temporal prediction module is used to fuse the spatial attention weight matrix into the forget gate and introduce residual connections between its multiple network units. It is used to perform temporal modeling on the enhanced training samples and output the radar echo prediction image sequence for multiple future time steps.

[0023] The results output module is used to output the future evolution of extreme precipitation and early warning information based on the radar echo prediction image sequence at multiple future time steps.

[0024] Beneficial Effects: Compared with existing technologies, this invention has the following significant advantages: This invention uses front-end abnormal hotspot extraction as a guide, transforming the anomaly detection results into a spatial attention weight matrix, providing prior information on anomaly regions for subsequent time-series prediction, thereby reducing reliance on a large number of extreme precipitation anomaly samples; This invention integrates the spatial attention weight matrix in the forget gate, enabling the time-series prediction process to more effectively focus on abnormal hotspot regions, improving the targeting and accuracy of future radar echo image prediction; This invention improves the information transmission efficiency in deep time-series modeling through residual connections, alleviating the gradient vanishing problem and improving training efficiency and model stability; This invention expands incremental training samples through an image dynamic regeneration module, introducing appropriate random variations while ensuring spatiotemporal structural consistency, thereby improving the model's robustness and generalization ability in complex weather scenarios; This invention realizes an integrated processing flow from abnormal hotspot extraction, weight guidance, dynamic enhancement to future evolution prediction, taking into account both the timeliness of extreme precipitation warnings and the ability to predict future evolution. Attached Figure Description

[0025] Figure 1 This is an overall flowchart of the method of the present invention. Detailed Implementation

[0026] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0027] like Figure 1 As shown, this embodiment of the invention provides a method for predicting short-term extreme precipitation based on anomaly-driven residual dynamic diffusion, including the following steps:

[0028] Step 1: Acquire radar echo image sequences and perform preprocessing.

[0029] The system acquires radar echo image sequences of the target area at multiple consecutive time points and constructs an input image sequence according to a preset time window. To improve the quality of the training data, preprocessing is performed on the original image sequence, including fixed-location outlier identification, abnormal noise filtering, and pixel value normalization, to reduce the impact of surface reflection clutter, ocean scattering interference, and anomalous electromagnetic propagation on subsequent predictions.

[0030] Step 2: Generate an abnormal hotspot image using the abnormal hotspot extraction module.

[0031] The preprocessed image sequence is then input into the anomaly hotspot extraction module. This module is built upon a student-teacher anomaly detection network, where the teacher network uses a pre-trained ResNet model, and the student network uses an untrained model. During training, both the teacher and student networks jointly learn the features of normal weather samples. In application, the teacher network, with its stronger generalization ability, forms stable representations of normal samples, while the student network has a weaker ability to represent anomalous samples, thus enabling the identification of anomalous regions.

[0032] To make the abnormal hotspot extraction module temporally sensitive, a Gaussian spatiotemporal attention mechanism is introduced into the student network. Let the... The input feature map at each time step is The spatial attention map can then be represented as:

[0033] ;

[0034] in, Represents the current time step. Representing the The feature map is input to this module at each time step; This indicates that adaptive pooling is used to downsample the feature map and extract global or local salient features; This indicates a bilinear interpolation upsampling operation, used to restore the pooled feature map to a specified spatial size; Indicates the kernel size as The convolution operation is used to extract local spatial features; This represents a non-linear activation function, which increases the non-linear expressive power of the network (sets it to 0 if it is less than 0, and leaves it unchanged if it is greater than 0); Indicates the kernel size as Convolutional operations are used for channel dimensionality reduction and cross-channel feature fusion; This represents the activation function, whose function is to map the output value to... Intervals are used to generate attention weights that represent the degree of importance; Based on input The calculated spatial attention map.

[0035] Furthermore, Gaussian time series weights are introduced:

[0036] ;

[0037] in, This is the current time step; The center position parameter of the Gaussian function; The width parameter (standard deviation) of the Gaussian distribution controls the rate at which the weights decay over time. A function of natural numbers; For the first Each time step uses a Gaussian temporal weight to measure the importance of the image at that moment to the current prediction. This Gaussian weight is used to temporally weight the spatial attention map, making the model pay more attention to newer image information. The weighted feature map is represented as follows:

[0038] ;

[0039] in, This indicates that time-series attention has been incorporated. ) and spatial attention ( The output feature map after that; This indicates element-wise multiplication, which means multiplying each value in the attention weight matrix by its corresponding feature map pixel.

[0040] Furthermore, to enhance the representation ability of abnormal regions, the feature maps output by different convolutional layers of the student network are fused pairwise. Let the first... Layer and first The layer output feature maps are respectively and The fusion result is expressed as:

[0041] ;

[0042] in Represents the sequence number of convolutional layers of different depths in the student network; Representing the Feature maps output by the layer; Representing the Feature maps output by the layer; This represents a feature map with multi-scale information obtained by adding and fusing features from two layers.

[0043] This fusion method preserves both the texture edge information extracted by the low-level convolutional layers and the global structural information extracted by the high-level convolutional layers, ultimately outputting an abnormal hotspot image.

[0044] Step 3: Generate a spatial attention weight matrix based on the abnormal hotspot images.

[0045] The output anomalous hotspot images are input into the spatial weight generation module to quantize the anomalous hotspot regions and form a spatial attention weight matrix. The spatial attention weight matrix is ​​used to characterize the degree of extreme weather or potential danger occurring in each pixel region at the current moment, and serves as guiding information for the forget gate within the subsequent residual time-series prediction module.

[0046] Step 4: Perform diffusion enhancement using the image dynamic regeneration module.

[0047] Considering the insufficient incremental training samples due to extreme weather conditions, the image dynamic regeneration module is invoked to perform data augmentation on the training data. Let the original image be... , No. The image at each time step is The noise intensity is The forward noise addition process is represented as:

[0048] ;

[0049] in This represents the image state at the previous time step; This represents the image state after adding noise at the current time step; Represents in a given Under the conditions, The conditional probability distribution; This represents a normal distribution; the variable before the semicolon is random, and the variable after the semicolon is the mean. Covariance Matrix ; The preset noise intensity is a minimum value that increases with time step. The identity matrix represents the fact that the noise added to each pixel is independent and identically distributed.

[0050] To improve training efficiency, a noisy image at any time step can be directly represented as:

[0051] ;

[0052] in, Represents the initial, noise-free original image; This represents the cumulative retention rate, from step 1 to step t. The product of two products; The definition is , here That is, the preset noise level; This represents random noise that follows a standard Gaussian distribution.

[0053] The diffusion model learns the inverse denoising process by predicting noise, and its loss function can be expressed as:

[0054] ;

[0055] in, This represents the loss value during the denoising process of the diffusion model; Represents the mathematical expectation operation, indicating the representation of all possible original images. Sampling noise and time step The average value on; This indicates that the deep learning model, given the current noisy image, can... and time step At that time, the predicted model noise, It is the set of all parameters in the model; The mean squared error is used to calculate the difference between the actual noise and the predicted noise. Furthermore, the task loss of the residual temporal prediction module is fed back to the image dynamic regeneration module to construct a joint optimization objective:

[0056] ;

[0057] in, This represents the combined loss of the entire system; This represents the reconstruction loss of the image dynamic regeneration module; This represents the prediction loss of the residual time series prediction module; The weighting coefficients are pre-set hyperparameters used to balance the weight ratios of the regeneration task (diffusion model) and the prediction task (temporal model) during backpropagation. This joint optimization method makes the enhanced image more suitable for the subsequent prediction task.

[0058] Step 5: Use the residual time series prediction module to output the future radar echo sequence.

[0059] The generated enhanced training samples are input into the residual temporal prediction module. The residual temporal prediction module includes a convolutional feature extraction layer and a WRLSTM temporal modeling unit. The convolutional feature extraction layer first extracts local spatial features from the enhanced image, and then inputs the local spatial features into the WRLSTM to perform temporal modeling.

[0060] The traditional LSTM forget gate calculation method is as follows:

[0061] ;

[0062] in, Indicates the current time step; This represents the output value of the forget gate at the current moment; The weight matrix represents the learnable weights of the forget gate; This represents the learnable bias term in the forget gate; It represents the hidden state of the previous moment, containing historical memory information; This represents the network input data at the current moment; This indicates a concatenation operation, which concatenates the hidden state from the previous time step with the current input along the feature channel dimension. This represents the activation function.

[0063] In this invention, the spatial attention weight matrix is... The calculation method for the merged Forget Gate is as follows:

[0064] ;

[0065] in, A scalar coefficient representing the degree of influence of the spatial attention weights; This represents the spatial attention weight matrix, which is the abnormal hotspot region matrix mentioned earlier; Indicates prior information about spatial anomalies The revised forget gate output is then guided. In this way, the forget gate not only relies on temporal characteristics but can also perceive the importance of spatially anomalous regions.

[0066] The fused forgetting gate participates in cell state updates:

[0067] ;

[0068] in, The updated cell state at the current moment is responsible for storing the network's main memory information; This represents the cell state at the previous moment; The input gate output at the current moment determines how much new information needs to be recorded; This represents the current state of the candidate cells.

[0069] To alleviate the gradient vanishing problem in deep temporal modeling, residual connections are added between WRLSTM units, and the residual fusion output can be expressed as:

[0070] ;

[0071] in, For the current gating output, The hidden state of the previous layer or the previous moment; The residual connection is used to output the final hidden state of the current unit, which helps alleviate the vanishing gradient problem. This residual path preserves the original output information of the previous unit and improves training efficiency.

[0072] Step 6: Output future evolution results and early warning information.

[0073] The system outputs a sequence of radar echo prediction images for multiple future time steps, and outputs the future evolution of extreme precipitation and early warning information based on the predicted image sequence.

[0074] experiment:

[0075] To verify the effectiveness of this invention, experiments were conducted using the HKO-7 radar echo dataset. The dataset covers data from 2009 to 2015, with a temporal resolution of 6 minutes and an image resolution of 480×480 pixels, covering a 512-kilometer radius around Hong Kong. For the abnormal hotspot extraction module, in the extreme precipitation warning experiment, data from the day before ten random three-day extreme precipitation events from 2009 to 2012 were selected as the training set, and data from the day of the extreme precipitation event were used as the test set. Data from 6 hours before the warning to the warning start time were considered abnormal samples, and 40 dBZ was used as the extreme precipitation threshold. Abnormal hotspot images were compared with radar echo images 100 minutes after the test images during evaluation, and a pixel anomaly value of 0.5 was used as the anomaly classification threshold. For the experiment predicting future evolution, extreme weather data from 2009 to 2012 were used as pre-training data. The daily sequences were divided into 40 non-overlapping blocks, of which 4 blocks were used as the training set, 1 block as the test set, and 1 block as the validation set. The incremental training data used data from May 2013, with a sample length of 20 frames, of which 5 frames were used as input and 15 frames were used as the prediction target. The initial learning rate of the WRLSTM module was set to 0.001, the maximum number of iterations was 80,000, and an early stopping mechanism was configured.

[0076] Dataset: The HKO-7 real radar echo dataset was used as the experimental dataset. The dataset information is shown in Table 1.

[0077] Table 1 HKO-7 Dataset ;

[0078] To demonstrate that the selected experimental data possesses spatiotemporal variation characteristics suitable for conducting research on extreme precipitation early warning and forecasting, the structural similarity of radar echo images at different months and time intervals was analyzed. The experiments show that during Hong Kong's typhoon season (May to October), the average SSIM of images at adjacent time steps significantly decreases, indicating that radar echo images change more rapidly during this period. Comparing three weather types—no precipitation, ordinary precipitation, and extreme precipitation—extreme precipitation weather exhibits a lower SSIM at the same time interval, indicating that its image changes faster and is more likely to appear as an "anomaly" within a short time window. These results demonstrate the rationale for using anomaly detection to extract rapidly changing areas and for use in front-end early warning and subsequent forecasting.

[0079] To ensure data quality, the HKO-7 dataset underwent rigorous preprocessing, with missing data properly handled to ensure the continuity of the time series. Furthermore, the dataset was standardized to make it more suitable as input for deep learning models.

[0080] The experimental environment is configured as follows: operating system is Ubuntu 18.04, processor is Intel Xeon Gold6128 CPU @ 3.40GHz, memory is 64GB, graphics card is NVIDIA-A100 40GB, Python version is 3.7, development tool is PyCharm, and deep learning framework is PyTorch 1.13.

[0081] Experimental evaluation indicators:

[0082] SSIM (Structural Similarity Index Measure):

[0083] ;

[0084] in and These are local window pixel blocks representing the real radar echo image and the model-predicted image, respectively. and It is an image and The average pixel value; and It is an image and The pixel variance; It is an image and Covariance between them; and It is a constant used to prevent the denominator from being zero. Simply put, SSIM is mainly used to evaluate how much image quality is lost when a compressed, transmitted, or processed image is compared to the original high-definition image.

[0085] CSI (Critical Success Index):

[0086] ;

[0087] in, (True Positive) represents a situation where extreme weather was predicted and actually occurred. (False Positive) represents a situation where extreme weather was predicted but did not actually occur. (FalseNegative) represents a situation where extreme weather was predicted not to occur, but actually did.

[0088] CSI (Conformity Index) is a commonly used metric for measuring the accuracy of weather forecasts, primarily used to assess the degree of match between forecasts and actual observations. For extreme weather forecasting, CSI effectively reflects the model's accuracy in identifying extreme weather events (such as heavy rain and typhoons). In extreme weather forecasting, These are often particularly serious problems because they can lead to catastrophic consequences. Therefore, a higher CSI value not only means accurate predictions of extreme weather, but also that there are fewer instances of underreporting.

[0089] FAR (False Alarm Ratio):

[0090] ;

[0091] FAR measures the false alarm rate of a model. It describes the proportion of extreme weather events predicted by the model that did not actually occur; in other words, the percentage of cases predicted as "existing" but not actually happening. A higher FAR indicates that the model has more false alarms, which may interfere with the reliability of practical applications and reduce the model's usability. Therefore, FAR is a key parameter for measuring a model's "false alarm tendency."

[0092] POD (Probability of Detection):

[0093] ;

[0094] POD measures a model's ability to capture actual events. It reflects the percentage of all real extreme weather events that the model successfully detects or predicts. In other words, a higher POD indicates that the model has a stronger response capability when facing real extreme weather scenarios and can effectively capture most actual events. It is a key reference standard for measuring a model's "missed event capability".

[0095] MAE (Mean Absolute Error):

[0096] ;

[0097] in It is the first The actual radar echo value of each pixel; It is the first The model predicts radar echo values ​​for each pixel; It is the total number of all pixels in the image.

[0098] MAE is a basic and intuitive regression evaluation metric that measures the mean absolute error between model predictions and actual values. It represents the average magnitude of the absolute values ​​of all prediction errors. The smaller the MAE, the smaller the overall error between the model's predictions and the actual values, and the higher the model's prediction accuracy.

[0099] MSE (Mean Squared Error):

[0100] ;

[0101] MSE calculates the average of all squared errors by squaring the error values. This method gives higher weight to larger errors during the calculation, so when the model has a large bias, MSE can more significantly reflect the decline in the model's predictive performance.

[0102] In future radar echo evolution prediction tasks, the ARDD-Net model (anomaly-driven residual dynamic diffusion model) proposed in this invention is compared with ConvLSTM (convolutional long short-term memory model), TrajGRU (trajectory-gated recurrent model), PredRNN (predictive recurrent neural network model), NowcastNet (nowcast prediction model), and trajPredRNN++ (trajectory prediction recurrent neural network++). As shown in Table 2, the model of this invention achieved good prediction results on the HKO-7 dataset: critical success index of 0.662, false alarm rate of 0.328, detection probability of 0.762, mean absolute error and mean square error of 71.182 and 41.004, respectively, and structural similarity index of 0.648. Further analysis shows that compared with the performance indicators of other models, the model of this invention performs better in overall comprehensive performance.

[0103] Table 2 shows the performance of each model on the HKO-7 dataset. ;

[0104] The above results demonstrate that the main process of this invention outperforms existing mainstream time-series prediction models in terms of overall performance, particularly in three aspects: First, in terms of operational metrics, ARDD-Net maintains a high detection probability while reducing the alert rate and achieving a higher critical success index, indicating that the model has better comprehensive identification capabilities for extreme precipitation events. Second, in terms of image error metrics, ARDD-Net achieves the lowest mean square error and the highest structural similarity index, indicating that the future radar echo images generated by this invention are superior in terms of overall structural consistency and large error control. Third, although the model of this invention is slightly inferior to the NowcastNet model in terms of mean absolute error, the difference is small, indicating that this invention, while maintaining overall error controllability, enhances the prediction quality of key areas through anomaly guidance and dynamic regeneration mechanisms. In other words, this invention tends to achieve a balance between prediction reliability in extreme areas and overall error.

Claims

1. A method for predicting short-term extreme precipitation based on anomaly-driven residual dynamic diffusion, characterized in that, Includes the following steps: (1) Obtain radar echo image sequences of the target area at multiple consecutive times, and construct the input image sequence according to the preset time window; (2) Input the input image sequence into the abnormal hotspot extraction module, utilize the feature expression differences between the pre-trained teacher network and the student network to be trained, combine the Gaussian temporal weighting mechanism to enhance the spatial attention map, and generate abnormal hotspot images representing potential abnormal regions by multi-resolution fusion of feature maps output by different convolutional layers. (3) Perform attention analysis on the abnormal hotspot images to generate a spatial attention weight matrix to indicate the degree of abnormality in each pixel region; (4) Call the image dynamic regeneration module, and perform a regeneration process of gradually adding noise and then removing noise from the original image to enhance the historical time window data and the current abnormal weather data to obtain enhanced training samples. (5) Input the enhanced training samples into the residual time-series prediction module, fuse the spatial attention weight matrix into the forget gate, guide the forget gate to selectively discard information according to the importance of the spatial abnormal region, and introduce residual connections between the residual time-series prediction modules to pass the original output information of the previous unit to the next unit in order to generate radar echo prediction image sequences for multiple future time steps. (6) Output the future evolution results and early warning information of extreme precipitation based on radar echo prediction image sequences at multiple future time steps.

2. The method for predicting short-term extreme precipitation based on anomaly-driven residual dynamic diffusion as described in claim 1, characterized in that, In step (2), the Gaussian temporal weighting mechanism enhances the spatial attention map as follows: adaptive pooling and upsampling are performed on the input image sequence to align the spatial dimensions. Spatial features are extracted through convolutional layers to generate a spatial attention map. The spatial attention map is temporally weighted using Gaussian distribution weights centered on the current time step, so that the model gives higher attention to the image at the new time step. The weighted spatial attention map is then multiplied element-wise with the original input feature map to obtain an output feature map that integrates temporal attention.

3. The method for predicting short-term extreme precipitation based on anomaly-driven residual dynamic diffusion as described in claim 1, characterized in that, In step (2), multi-resolution fusion specifically involves extracting feature maps from multiple convolutional layers of different depths in the student network, and then adding and fusing the low-level feature maps representing local edge textures with the high-level feature maps representing global structural semantics to obtain image information at different scales.

4. The method for predicting short-term extreme precipitation based on anomaly-driven residual dynamic diffusion according to claim 1, characterized in that, In step (4), the image dynamic regeneration module uses a diffusion model to perform the regeneration process. The forward process is to gradually add random noise to the original image, and the reverse process is to gradually predict and remove noise from the noisy image to restore the image. The training objective is composed of the noise prediction loss and the prediction loss of the residual temporal prediction module, so that the generated enhanced image maintains the consistency of the overall structure.

5. The method for predicting short-term extreme precipitation based on anomaly-driven residual dynamic diffusion according to claim 1, characterized in that, In step (5), the residual time-series prediction module is a long short-term memory network.

6. The method for predicting short-term extreme precipitation based on anomaly-driven residual dynamic diffusion according to claim 5, characterized in that, When calculating the activation value of the forget gate, the weighted values ​​of the current input, the previous hidden state, and the spatial attention weight matrix are used together as the input to the activation function, so that the output of the forget gate depends on both the temporal features and the importance of the spatially anomalous region.

7. The method for predicting short-term extreme precipitation based on anomaly-driven residual dynamic diffusion according to claim 1, characterized in that, The residual connection is as follows: the gated output of the current network unit is added to and fused with the hidden state of the previous time step or the hidden state of the previous network unit, and the fused result is used as the final hidden state output of the current unit.

8. A short-term extreme precipitation prediction system based on anomaly-driven residual dynamic diffusion, characterized in that, include: Data acquisition module: used to acquire radar echo image sequences of the target area at multiple consecutive time points; Anomaly hotspot extraction module: This module utilizes the feature representation differences between the pre-trained teacher network and the student network to be trained, combined with a Gaussian temporal weighting mechanism and a multi-resolution feature fusion structure, to generate anomaly hotspot images from radar echo image sequences. Spatial weight generation module: used to perform attention analysis on anomalous hotspot images and generate a spatial attention weight matrix; Image dynamic regeneration module: used to perform a regeneration process of progressively adding noise and then reverse denoising on historical time window data and current abnormal weather data to generate enhanced training samples; The residual temporal prediction module is used to fuse the spatial attention weight matrix into the forget gate and introduce residual connections between its multiple network units. It is used to perform temporal modeling on the enhanced training samples and output the radar echo prediction image sequence for multiple future time steps. The results output module is used to output the future evolution of extreme precipitation and early warning information based on the radar echo prediction image sequence at multiple future time steps.

Images