A spatiotemporal coordinated attention radar nowcast method based on long-term evolution characteristics

Abstract

The application provides a spatiotemporal cooperative attention radar nowcasting method based on long-term evolution characteristics, relates to the technical field of radar data processing, and comprises the following steps: obtaining a historical radar echo image sequence of a to-be-monitored target, and dividing the historical radar echo image sequence into a historical input image sequence and a future real image sequence; inputting the historical input image sequence into a bidirectional spatiotemporal attention recurrent network, generating a predicted echo image sequence through feature extraction and fusion of each layer; obtaining a total loss value of the current iteration by calculating a pixel-level error, a spatiotemporal distribution deviation and a potential divergence penalty term; iteratively adjusting the parameters of the bidirectional spatiotemporal attention recurrent network in the gradient direction generated by the total loss value of the current iteration, so as to obtain a radar nowcasting model; obtaining a real-time radar echo image sequence, and predicting a future radar echo image by using the radar nowcasting model. The application reduces error accumulation, echo ambiguity and morphological distortion from the root.

Description

Technical Field

[0001] This invention relates to the field of radar data processing technology, and more particularly to a spatiotemporal collaborative attention radar near-term prediction method based on long-term evolution characteristics. Background Technology

[0002] Radar echo extrapolation is a core technology for short-term and nowcasting weather forecasting. By mining the patterns in historical radar echo time series data, it quantitatively predicts the future evolution trend of echoes, providing core support for early warning of sudden severe weather events such as short-term heavy precipitation and severe convection.

[0003] Radar echo observation data is characterized by strong nonlinearity and complex spatiotemporal dynamic coupling. Meteorological operations place stringent requirements on the extrapolation of long-term radar echoes with a duration of over 60 minutes in terms of forecast accuracy, stability, and operational adaptability.

[0004] Currently, radar echo extrapolation is divided into two main technical routes: traditional numerical calculation models and deep learning models. Neither of them can meet the requirements for long-term, high-precision prediction.

[0005] The closest existing technology to this invention is the radar echo extrapolation method based on hybrid neural networks (representative models: MRP-Net, SA-ConvLSTM, PredRNN++). Taking MRP-Net as an example, the steps are as follows: 1. Data input: Input a continuous time-series radar echo image sequence; 2. Spatial feature extraction: Use a convolutional neural network to extract the local spatial features of the radar echo; 3. Temporal feature extraction: Use a recurrent neural network (LSTM / ConVLSTM) to extract time-series features; 4. Feature fusion: Fuse high- and low-level spatial-temporal features through skip connections and residual structures; 5. Result output: Output the radar echo prediction image for the future time period; 6. Loss optimization: Use mean squared error (MSE) / mean absolute error (MAE) as the loss function, and only perform pixel-level fitting optimization.

[0006] Traditional numerical calculation models (optical flow method, echo mutual tracking method, centroid tracking method) follow these steps: 1. Calculate the motion vectors (optical flow / centroid) of adjacent radar echoes; 2. Extrapolate the future echo position based on the motion vectors; 3. Output short-term prediction results for stationary echoes. Specifically, existing technologies have the following drawbacks: 1. Traditional numerical calculation models can only fit stationary radar echoes and cannot model complex spatiotemporal evolution patterns over long periods. They have poor adaptability to severe convective weather, leading to rapid decay of echo trends and distortion of prediction morphology in long-term predictions. 2. Convolutional Neural Networks (CNNs) are constrained by the inherent receptive field of the convolutional kernel, allowing only the extraction of local spatial features and failing to capture long-distance spatial dependencies within radar echo images. 3. Recurrent Neural Networks (RNNs / LSTMs) suffer from inherent defects such as gradient vanishing and long-term feature forgetting, resulting in blurred and distorted prediction images. 4. Hybrid Neural Networks have complex network structures, single feature transmission paths, unstable training processes, and poor model convergence. 5. Existing loss function: It only focuses on pixel-level fitting error, does not consider the spatiotemporal correlation characteristics of radar echo, exacerbates the error accumulation of long-term prediction, and causes echo details and strong core areas to be lost quickly.

[0007] No effective solutions have yet been proposed to address the problems in the relevant technologies. Summary of the Invention

[0008] In view of this, the present invention provides a spatiotemporal collaborative attention radar near-term prediction method based on long-term evolution characteristics to solve the aforementioned problems.

[0009] To solve the above problems, the specific technical solution adopted by the present invention is as follows:

[0010] The spatiotemporal collaborative attention radar near-term prediction method based on long-term evolution characteristics includes the following steps:

[0011] S1. Acquire the historical radar echo image sequence of the target to be monitored, and divide it into historical input image sequence and future real image sequence according to the preset number of frames;

[0012] S2. Construct a bidirectional spatiotemporal attention recurrent network and input the historical input image sequence into the bidirectional spatiotemporal attention recurrent network. Through feature extraction and fusion at each layer, generate a predicted echo image sequence. The feature extraction and fusion includes the calculation of the proportion of zero values ​​in the local neighborhood grid and the causal flow intervention analysis based on the backdoor adjustment rule.

[0013] S3. Input the predicted echo image sequence and the future real image sequence into the pre-constructed total objective function, and obtain the total loss value of the current iteration by calculating the pixel-level error, spatiotemporal distribution deviation and advection divergence penalty term;

[0014] S4. Based on the gradient direction generated by the total loss value of the current iteration, iteratively adjust the parameters of the bidirectional spatiotemporal attention recurrent network in reverse order to obtain the radar near-term prediction model.

[0015] S5. Acquire real-time radar echo image sequences and use the radar nowcasting model to predict future radar echo images.

[0016] The beneficial effects of this invention are as follows:

[0017] 1. This invention solves the problems of insufficient long-range spatial dependency capture and insufficient temporal modeling in existing models through spatiotemporal collaborative attention mechanism. The bidirectional information flow transmission effectively suppresses gradient vanishing and information decay. STDC Loss and the physical constraint of convective divergence break through the limitation of traditional loss functions that only fit pixels. It not only ensures pixel accuracy, but also constrains the spatiotemporal distribution of echoes and meteorological physical laws, thereby reducing error accumulation, echo blurring and morphological distortion from the root.

[0018] 2. Compared with existing mainstream models, this invention significantly improves prediction accuracy, image quality, model stability and generalization in long-term radar echo extrapolation, and can better support long-term early warning services for short-term severe convective weather. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly described below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings:

[0020] Figure 1 This is a flowchart according to an embodiment of the present invention;

[0021] Figure 2 This is an overall structural diagram of the Bi-STAN weather radar echo extrapolation network according to an embodiment of the present invention;

[0022] Figure 3 This is a schematic diagram illustrating the working principle of the spatiotemporal collaborative attention mechanism according to an embodiment of the present invention;

[0023] Figure 4 This is a schematic diagram of the spatiotemporal collaborative attention mechanism according to an embodiment of the present invention;

[0024] Figure 5 This is a structural diagram of the Intra-Spatial Long-Range Attention (ILSAM) mechanism according to an embodiment of the present invention;

[0025] Figure 6 This is a structural diagram of the cross-frame temporal evolution attention mechanism (TEAM) according to an embodiment of the present invention;

[0026] Figure 7 This is a flowchart of the operation of the Bi-STAN loop unit according to an embodiment of the present invention;

[0027] Figure 8 This is a time-step score comparison chart of MSE and SSIM indices in the Guangzhou station radar echo dataset according to an embodiment of the present invention. Figure 8 (a) shows a comparison of the scores of the Mean Squared Error (MSE) index. Figure 8 (b) is a comparison chart of the scores of the Structural Similarity Index (SSIM);

[0028] Figure 9 This is a comparison chart of visualization results based on typical sequences of radar echo dataset from Guangzhou Station according to an embodiment of the present invention;

[0029] Figure 10 This is a time-step score comparison chart of MSE and SSIM evaluation metrics under the CIKM 2017 radar echo dataset according to an embodiment of the present invention. Figure 10 (a) shows the time-step scores of the MSE evaluation index. Figure 10 (b) is a time-step score plot of the SSIM evaluation index;

[0030] Figure 11 This is a comparison chart of visualization results of a typical sequence under the CIKM 2017 radar echo dataset according to an embodiment of the present invention. Detailed Implementation

[0031] To enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this application.

[0032] According to embodiments of the present invention, a spatiotemporal cooperative attention radar near-term prediction method based on long-term evolution characteristics is provided.

[0033] Specifically, the spatiotemporal collaborative attention radar echo extrapolation network constructed in this invention is composed of four layers of stacked Bi-STAN cyclic units. As the core unit of the deep spatiotemporal feature extraction network, each Bi-STAN integrates an intra-frame spatial long-range attention mechanism (ILSAM) and a cross-frame temporal evolution attention mechanism (TEAM). This network implements memory units. Hidden state Memory unit The overall architecture for bidirectional information transmission in both horizontal (temporal) and vertical (spatial) directions is as follows: Figure 2 As shown. Among them, , , These represent the original radar echo images input at times t-1, t, and t+1, respectively. , , The model predicts the radar echo image at the corresponding time step. `Conv` represents the convolutional output layer used to map deep spatiotemporal features to a single-channel predicted echo image; `M` represents the inter-layer memory unit features of the Bi-STAN unit, carrying historical memory information for spatial long-range attention modeling. The superscript numbers 1 to 4 correspond to the 1st to 4th layers of Bi-STAN units, respectively, and the subscript numbers correspond to different time steps, such as... , , These represent the memory cell characteristics of the 1st, 2nd, and 3rd layer units at time t-1. The memory unit characteristics of the third layer at time t. , The memory cell features of the second and third layers at time t+1; H represents the hidden state of the Bi-STAN cell, which is the output feature after gating at the current time, such as... , This represents the hidden states of the first and fourth layer cells at time t-1. Let be the hidden state of the first-layer cell at time t. This represents the hidden state of the third-layer cell at time t+1. This represents the cell state of the first layer Bi-STAN unit at time t, used to store the temporal cell information of the recurrent network.

[0034] The network uses radar echo image sequences As input tensors, the encoder first fuses ILSAM to extract long-range spatial dependencies, then updates cell and hidden states through gating operations. After processing by the context feature fusion module, it is fed into the decoder to fuse TEAM to extract time-series evolution features, and finally outputs the predicted radar echo image. .

[0035] Furthermore, to ensure the network possesses accurate predictive capabilities, this invention requires offline model training before system deployment. The STDC joint loss function serves as the objective function to guide parameter optimization during this training phase, and its specific processing and usage steps are as follows:

[0036] Dataset preparation: Select historical radar echo image sequences (such as the Guangzhou station dataset) and pair the images with 10 historical input frames and 10 future ground truth frames.

[0037] Forward inference: The input sequence is fed into a Bi-STAN recurrent network, and after feature extraction and reconstruction at each layer, a predicted echo image sequence is generated. .

[0038] Bias calculation (loss function processing): This involves processing the generated predicted image... The image I is fed into the STDC Loss calculation module simultaneously with the real image I. This module calculates the pixel-level error (MSE) and spatiotemporal distribution deviation (differential entropy DE) in parallel.

[0039] Parameter inverse update: using the gradient descent algorithm, based on The resulting gradient direction is used to reverse-adjust all convolution weights and gating parameters in the Bi-STAN unit and encoder / decoder.

[0040] Iterative loop: Repeat the above process until the loss function converges and the metrics (such as CSI, SSIM) on the validation set are optimal, thereby obtaining a prediction model with long-term evolution capture capability.

[0041] In establishing optimization objectives with multidimensional physical constraints Subsequently, a high-capacity network topology capable of handling complex spatiotemporal computations needs to be constructed. To effectively address the technical pain points of insufficient modeling of time-series dependencies and weak ability to capture long-distance spatial dependencies, this invention designs a deep spatiotemporal feature extraction network architecture composed of four stacked Bi-STAN recurrent units.

[0042] Bi-STAN consists of four layers of stacked cyclic units. At different times, units within the same layer share a time dimension transmission path, enabling feature analysis of the radar echo's time dimension. At the same time, units in different layers share a spatial dimension transmission path, enabling feature extraction of the radar echo's spatial dimension. Its spatiotemporal information transmission principle is as follows: Figure 3 As shown.

[0043] The direction of transmission between spatial dimension memory units and hidden states is represented by red dashed arrows, while black arrows represent the direction of transmission between unit input and output states. Relying on the transmission of bidirectional spatiotemporal information, efficient extraction of long-range spatiotemporal dependent features of radar echoes is achieved, suppressing the attenuation of echo trends during long-term prediction. While minimizing losses during feature computation and transmission, accurate prediction of the long-term motion evolution of radar echoes is achieved.

[0044] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figure 1 As shown, according to a first embodiment of the present invention, a spatiotemporal cooperative attention radar near-term prediction method based on long-term evolution characteristics is provided, comprising the following steps:

[0045] S1. Acquire the historical radar echo image sequence of the target to be monitored, and divide it into historical input image sequence and future real image sequence according to the preset number of frames;

[0046] It should be noted that after dividing the historical input image sequence and the future real image sequence, a temporal consistency screening method based on motion vector pre-estimation is used to preprocess the historical input image sequence and the future real image sequence, specifically including:

[0047] Step 1: Acquire consecutive frame image pairs and extract spatial grid points for the motion vector to be estimated;

[0048] Specifically, for each frame of the historical input image sequence and the future real image sequence, the t-th frame and the (t-1)-th frame are taken in chronological order to form a continuous frame image pair. For each pair of continuous frames, a uniformly distributed set of spatial grid sampling points is constructed in the spatial dimension. The spatial resolution of the grid sampling points is consistent with the resolution of the original image, which is used for point-by-point estimation of subsequent motion vectors.

[0049] Step 2: Estimate the motion velocity vector of each grid point based on the cross-correlation tracking method, and construct the inter-frame motion vector field;

[0050] Specifically, for each pair of consecutive frame images and their corresponding spatial grid sampling point set, a reference sub-region (M×M pixels in size) centered on each grid point is selected within a preset search window. Then, in frame t-1, a matching sub-region with the maximum similarity to this reference sub-region is searched. The spatial matching degree between the two sub-regions is measured by calculating the normalized cross-correlation coefficient between the reference sub-region and all candidate sub-regions within the search window. The spatial displacement (Δx, Δy) corresponding to the maximum value of the cross-correlation coefficient is determined as the pre-estimated value V of the motion velocity vector of the grid point within the time step of the current frame. pre(t) =(u pre ,v pre ), where u pre =Δx / Δt, v pre =Δy / Δt, where Δt is the time interval between adjacent frames; after traversing all grid sampling points, an inter-frame motion vector field covering the entire spatial domain can be constructed, which quantitatively characterizes the two-dimensional advection motion of radar echoes between adjacent time points.

[0051] The calculation of the normalized cross-correlation coefficient between the reference sub-region and all candidate sub-regions within the search window specifically includes: taking a square region of size M×M pixels centered at the grid point in frame t-1 as the reference sub-region; simultaneously, within the search window centered at the same grid point in frame t, extracting each square region of the same size as the reference sub-region as a candidate sub-region; for both the reference sub-region and each candidate sub-region, calculating the mean of all pixel values ​​within their respective regions, and subtracting the mean of the pixel value of each pixel within the region... The mean is used to obtain the mean-reduced pixel value sequence for each region. Then, the mean-reduced pixel value sequence of the reference sub-region and the mean-reduced pixel value sequence of the candidate sub-region are multiplied pixel by pixel and summed to obtain the covariance measure of the two. The sum of squares of the mean-reduced pixel value sequences of the reference sub-region and the candidate sub-region are calculated respectively. The square root of the product of the two is used as the normalized denominator. The covariance measure obtained above is divided by the normalized denominator to obtain the normalized cross-correlation coefficient between the reference sub-region and the candidate sub-region.

[0052] Furthermore, the correlation coefficient ranges from -1 to 1. A value closer to 1 indicates a higher spatial structural similarity and better matching between the two sub-regions; a value closer to 0 or negative indicates a lower similarity. By traversing all candidate sub-regions within the search window and calculating the normalized cross-correlation coefficient between each candidate sub-region and the reference sub-region, the candidate sub-region corresponding to the maximum correlation coefficient can be located. Its spatial displacement relative to the reference sub-region is then determined as the estimated motion velocity vector value of that grid point within the current inter-frame time step. Through the above grid-point calculation and traversal, an inter-frame motion vector field covering the entire spatial domain is finally constructed.

[0053] Step 3: Calculate the inter-frame temporal consistency index based on the motion vector field to quantify the smoothness of motion evolution between adjacent frames;

[0054] It should be noted that the calculation of the inter-frame temporal consistency index is based on the physical principle of divergence-free fluid dynamics. Its core lies in using the inter-frame motion vector field to quantitatively evaluate whether the advection motion of radar echoes between adjacent frames satisfies the continuity equation.

[0055] Specifically, after obtaining the two-dimensional motion vector field covering the entire spatial domain, spatial differentiation is required for the motion vector at each grid point. For the horizontal velocity component, spatial partial derivatives are calculated along the horizontal direction; for the vertical velocity component, spatial partial derivatives are calculated along the vertical direction. Adding these two derivatives yields the divergence value at that grid point. The physical meaning of this divergence value lies in characterizing the source and sink characteristics of radar echoes within a local spatial region: when the divergence value is zero, it indicates that the echo motion satisfies the divergence-free constraint of incompressible fluids, meaning that the flux of echo material flowing into and out of the local region is balanced, and the motion evolution between adjacent frames fully conforms to the law of conservation of atmospheric mass; when the divergence value is not zero, it indicates the presence of echo convergence or divergence in the local region, reflecting a non-physical abrupt change or discontinuity in the motion field in space.

[0056] Furthermore, to further quantify the temporal consistency of the entire frame image sequence, the absolute divergence values ​​of all grid points are spatially averaged to obtain the temporal consistency index for that frame. The smaller the value of this index, the more the echo motion between adjacent frames satisfies the divergence-free constraint in the entire spatial domain, meaning the smoother and more continuous the temporal evolution between frames, and the more it conforms to the motion laws of the real atmosphere. Conversely, the larger the value, the more it indicates the existence of abnormal jumps or data mutations between frames that violate the principles of fluid kinematics. Such samples may have a negative impact on the physical consistency learning of the model in subsequent training.

[0057] Step 4: Perform sample screening and abnormal frame repair based on temporal consistency indicators, and output a clean image sequence that has passed the temporal consistency screening.

[0058] It should be noted that after calculating the temporal consistency index, quality assessment and selective processing of each frame in the image sequence are required based on this index. Specifically, a temporal consistency threshold based on prior knowledge of radar echo kinematics is preset. This threshold can be determined by statistically analyzing the distribution of temporal consistency indices under normal evolution conditions in a large amount of historical radar echo data. Its physical meaning lies in defining the critical boundary between "inter-frame evolution that satisfies the continuity of fluid motion" and "inter-frame evolution with abnormal abrupt changes." The calculated temporal consistency index of each frame is compared with this threshold: if the index value of a frame is lower than or equal to the threshold, it indicates that the echo motion between this frame and its predecessor is physically reasonable and continuous, and this frame is judged as a temporally qualified sample and retained in its entirety; if the index value is higher than the threshold, it indicates that there is an abnormal jump between this frame and its predecessor that violates the laws of fluid motion. In this case, the continuous frame pair formed by this frame and its predecessor is judged as a temporally abnormal sample, and the entire frame pair is removed from the training set.

[0059] Meanwhile, to maintain the continuity and integrity of the training samples, after removing anomalous frame pairs, the position index of the anomalous frame in the original time series needs to be recorded, and bidirectional linear interpolation is performed between the normal frames adjacent to the anomalous frame for repair. That is, based on the actual time interval between the two frames before and after the anomalous frame, the repaired image at the time of the anomalous frame is calculated using linear interpolation, and the original anomalous frame is inserted into the corresponding position in the sequence. This maintains the constant length and temporal continuity of the image sequence in the time dimension while removing anomalous samples. Finally, the clean image sequence output after the above screening and repair processes has passed the physical consistency test without divergence constraints for the motion evolution between all adjacent frames, fundamentally ensuring the input data quality of the subsequent Bi-STAN network when performing cross-frame temporal evolution attention calculations, and avoiding the contamination of the spatial attention weight matrix by anomalous frames.

[0060] S2. Construct a bidirectional spatiotemporal attention recurrent network (Bi-STAN) and input the historical input image sequence into the bidirectional spatiotemporal attention recurrent network. Through feature extraction and fusion at each layer, generate a predicted echo image sequence. The feature extraction and fusion includes the calculation of the proportion of zero values ​​in the local neighborhood grid and the causal flow intervention analysis based on the backdoor adjustment rule.

[0061] Specifically, a Bi-STAN unit is constructed by introducing the ILSAM intra-frame spatial long-range attention mechanism and the TEAM cross-frame temporal evolution attention mechanism. This design aims to enhance the model's performance in modeling spatiotemporal features, enabling it to more accurately capture spatial dependencies and dynamic changes in the temporal dimension, thereby effectively improving long-term prediction performance in radar echo extrapolation tasks. Its structure is as follows: Figure 4 As shown.

[0062] To address the limitations of traditional convolutional neural networks in local feature extraction, this invention proposes an Intra-Spatial Long-Range Attention (ILSAM) mechanism deployed after the convolutional layers of the network encoder. ILSAM uses the hidden state H at the current time step... t And the memory unit M from the previous time step t-1 As input, the algorithm executes five steps sequentially: feature extraction, attention weight calculation, weighted feature aggregation, feature fusion, and memory unit update. This allows it to accurately capture long-range spatial dependencies within radar images. Its specific structure is as follows: Figure 5 As shown.

[0063] In a preferred embodiment, the construction of a bidirectional spatiotemporal attention recurrent network (Bi-STAN), and the input of historical input image sequences into the Bi-STAN, generating a predicted echo image sequence through feature extraction and fusion at each layer, includes the following steps:

[0064] S21. Construct a bidirectional spatiotemporal attention recurrent network that includes an intra-frame spatial long-range attention mechanism and a cross-frame temporal evolution attention mechanism;

[0065] S22. Based on the intra-frame spatial long-range attention mechanism, a lightweight structural causal model is constructed to extract features for zero-value background region identification and causal flow intervention analysis of historical input image sequences. Combined with spatial self-attention calculation, spatial long-range dependency features are obtained.

[0066] It should be noted that in traditional spatial attention mechanisms, feature extraction often relies on direct convolutional mapping. This makes the network prone to misidentifying large areas of clear-sky low-value background regions in radar echoes as long-range spatial correlations, resulting in statistical spurious correlations. To completely overcome this technical bottleneck, this method first constructs a lightweight structural causal model (SCM) in the feature extraction stage: First, by explicitly introducing meteorological transformation physical values, the large-scale zero-value background region X is accurately determined. RAW This strictly represents the original radar reflectivity factor value (in dBZ) corresponding to each spatial grid pixel in the input radar echo image. To ensure the determination of the large-scale zero-value background region is fully feasible and deterministic, this invention provides a precise method for calculating the proportion of zero values ​​in the local neighborhood grid. When the input radar echo image tensor I... t ∈R C*H*W The reflectance of the pixels in the image satisfies the meteorological conversion physical value X. RAW When the value is <20 dBZ and the proportion of zero values ​​exceeds 85% within an 8*8 local neighborhood grid, the region is determined to be a normal clear-sky low-value background. Therefore, this region is defined as the spatial confusion factor Z that hinders the modeling of the long-range evolution characteristics of the convective core. cf ;

[0067] In the constructed lightweight structural causal model G=(V,E), Z cfAs a common factor acting simultaneously on the shallow representation of the feature extractor and the dependent variable for future extrapolation, conventional models fall into mean regression in pursuit of minimizing global pixel error, inducing non-physical rapid decay and dissipation of the strong convective core in long-term prediction. To completely cut off its statistical pseudo-correlation interference in the feature space, this method uses the backdoor adjustment rule in causal inference to intervene in the causal flow. In the implementation architecture of the neural network, this invention uses three dilated convolutional branches with a kernel size of 3*3 and a fixed channel dimension of C=128 as carriers, with their dilation rates explicitly set to r1=1, r2=2, and r3=3, respectively. This inter-layer connection design allows the network to expand the grid receptive field exponentially to 3*3, 5*5, and 7*7, respectively, without increasing the number of additional parameters. Under the Doppler radar grid with a resolution of 1km, it perfectly covers the physical span of small-to-medium scale strong convective single-cell systems of 3km-7km. Input the current step hidden state H t Interlayer memory state M from the previous step t-1 Then, the three branches are mapped in parallel to generate the original query, key, and value tensors. Subsequently, the module utilizes the soft mask M. cf Global average pooling is performed on the global context to extract the representation vector Z from the confusion factor set. cg ∈R C*1*1 The original tensor was then subjected to interventional reparameterization based on causal modulation, involving element-wise multiplication and residual stripping. This intervention involved multiplicative gating suppression of the original tensor by generating channel weights from the global clear-sky background vector, followed by explicit subtraction of the background bias to achieve residual stripping. This process thoroughly purified the spatial causal features of the strong convection core region. The specific result of the interventional reparameterization process was the acquisition of a high-purity tensor Q, which thoroughly removed global clear-sky background noise and exhibited spatial causal dynamic features of the strong convection core region. h K h and V h The logical connection between this stage and the feature extraction stage is as follows: these three high-purity feature tensors will be directly used as the matrix multiplication operator in the subsequent attention weight calculation stage, causing the subsequent spatial self-attention matrix formula to cascade and evolve into A. hThis ensures that subsequent attention operators can strongly focus on the causal dynamic evolution between the core regions of strong convective echoes. In high-time-sensitivity operational scenarios such as the South China Meiyu front rainstorm and typhoon nowcasting, which last for more than 60 minutes, the spatial proportion of the drastically evolving strong convective core is usually less than 5%. Conventional models are easily affected by the statistical drag of the dominant 95% clear sky background area, leading to long-term forecast distortion and core dissipation. However, this solution, through causal intervention steps, forcibly injects physical topology selectivity into the model driven by statistical data. This is demonstrated by the Doppler ultrasound at the Guangzhou Station of the China Meteorological Administration. Ablation comparison experiments on the radar data set show that, compared to the variant network with the causal module removed (w / oILSAM), this scheme achieves a 3.65% net deterministic improvement in the Critical Success Index (CSI-AVG) and a significant 7.07% increase in Structural Image Similarity (SSIM). It fundamentally overcomes the technical challenge of rapid dissipation of echo kernels and boundary blurring caused by mean regression induced by background noise in long-term extrapolation, demonstrating outstanding substantial features and significant progress in radar echo detail preservation and morphological fidelity.

[0068] In the feature extraction stage, ILSAM did not use traditional convolution, but instead employed three dilated convolutions with a kernel size of 3×3 in parallel, using a parallel architecture to extract features from H. t and M t-1 Radar echo representations at different scales were extracted from deep space. To seamlessly cover a wider geographical area, the expansion rates were set to 1, 2, and 3 sequentially, allowing the network to exponentially expand its receptive field to 3×3, 5×5, and 7×7 respectively without increasing the number of additional parameters. Through convolutional transformation of the multi-scale receptive field, the network computed and generated the core tensor for spatial self-attention matching: the query vector Q. h Key vector K h and K m and the value vector V h and V m In this process, the memory units and hidden states share a query vector in the computation logic to ensure alignment of the feature space. The rigorous mathematical definition of the computation process is as follows:

[0069]

[0070]

[0071] In the formula, , , This indicates that the hidden state is affected in the current step. The convolutional filter matrix is ​​used for convolution mapping to generate hidden state query vectors, respectively generating... , , , and It is used for the previous memory unit. The convolution filter matrix is ​​used to generate the memory cell key vector through convolution mapping. and .

[0072] During the attention weight calculation stage, the attention values ​​are normalized using the Softmax function to obtain the spatial attention matrix of the hidden state and the memory state, preventing gradient explosion. The calculation process is shown below:

[0073] ; ;

[0074] In the weighted feature aggregation stage, the attention matrix and its corresponding value vector are weighted and summed to obtain the feature map Z after global spatial feature extraction. h With Z m The calculation method is as follows:

[0075] ; ;

[0076] In the feature fusion stage, Z h With Z m Merging is performed along the 1×1 convolutional channels to obtain the global spatial fusion feature map Z, calculated using the following formula:

[0077]

[0078] In the formula, d represents the channel dimension. The spatial attention weight matrix represents the hidden state; A value vector representing the hidden state; This represents the matrix product operation, used to perform a weighted summation of the attention matrix and the value vector; The spatial attention weight matrix represents the memory state; A vector representing the values ​​of a memory cell; This represents an intermediate fusion feature map that initially integrates information from hidden states and memory units. This represents a specific convolutional filter matrix used for weight transformation during feature fusion. This represents the convolution operation; This indicates a channel fusion operation, which involves splicing along the channel dimensions of the feature map.

[0079] During the memory unit update phase, the fused feature map and the original input are fed into the gating structure through a 1×1 convolution operation, and the first input gate is obtained after processing. First input modulation gate and the first output gate The memory unit is updated and a new hidden state is output based on the gating mechanism. The calculation process is as follows:

[0080] ; ;

[0081] ; ;

[0082] ;

[0083] In the formula, denoted by sigmoid, and tanh by hyperbolic tangent activation function. This indicates an element-wise multiplication operation. This represents the learnable weight matrix corresponding to the input gate, used to perform linear transformations on the input features and calculate the feature mapping of the input gate; This represents the learnable weight matrix corresponding to the cell state update gate, used to generate candidate cell features to be updated; This represents the learnable weight matrix corresponding to the output gate, which is used for linear transformation to obtain the output gate control features and determine which cell state information is output to the hidden layer. This represents the bias term corresponding to the input gate, and... This will help complete the linear calculation of the input gate; This represents the bias term corresponding to the candidate cell update gate, which helps generate candidate cell states; This represents the bias term corresponding to the output gate, in conjunction with... Calculate the output gate activation value.

[0084] As a preferred embodiment, the step of using an intra-frame spatial long-range attention mechanism to extract features for zero-value background region identification and causal flow intervention analysis of historical input image sequences by constructing a lightweight structural causal model, and combining spatial self-attention calculation to obtain spatial long-range dependent features includes the following steps:

[0085] S221. Based on the calculation of meteorological converted physical values ​​and the proportion of zero values ​​in local neighborhood grids, the historical input image sequence is used to identify the normalized clear sky low value background area and determine the spatial confusion factor.

[0086] As a preferred embodiment, the step of identifying normalized clear-sky low-value background regions and determining spatial confusion factors based on the calculation of meteorological converted physical values ​​and the proportion of zero values ​​in local neighborhood grids includes the following steps: S2211, determining the original radar reflectivity factor value of each spatial grid pixel according to the historical input image sequence; S2212, when the original radar reflectivity factor value meets the preset meteorological converted physical value and the zero value in the local neighborhood grid exceeds the preset threshold, determining that the spatial grid region is a normalized clear-sky low-value background region; S2213, defining the normalized clear-sky low-value background region as a spatial confusion factor that hinders the modeling of long-range evolution characteristics of the convection core space.

[0087] S222. Based on the spatial confusion factor, construct a lightweight structural causal model. First, complete the overall framework construction of the lightweight structural causal model, defining the model as including four major functional modules: dilated convolutional branch, global context pooling, causal intervention modulation, and spatial self-attention calculation.

[0088] S223. Obtain the current hidden state and the previous interlayer memory state of the current bidirectional spatiotemporal attention recurrent network, and generate an original tensor containing query vector, key vector and value vector through lightweight structure causal model mapping; Specifically, based on the lightweight structure causal model constructed in step S222, call its built-in dilated convolution branch module to perform feature mapping on the current hidden state and the previous interlayer memory state output by the bidirectional spatiotemporal attention recurrent network, and generate an original tensor containing query vector, key vector and value vector, which serves as the input basis for subsequent confusion factor extraction and causal modulation operation.

[0089] S224. Global average pooling is performed on the global context based on a soft mask to extract the concentrated representation vector of confusion factors. The soft mask is generated based on the normalized clear sky low-value background region (spatial confusion factor) identified in this step: first, each spatial pixel is classified and labeled according to the reflectance threshold and the proportion of zero values ​​in the neighborhood to generate a binary mask with the same spatial size as the input image, and then Gaussian smoothing is performed to obtain a soft mask with a value range of [0,1]. When extracting the concentrated representation vector of confusion factors based on the soft mask, the soft mask is first interpolated and adapted to the spatial size consistent with the current global context feature tensor, and then multiplied element-wise with the global context feature map to weight and highlight the background region features. Subsequently, global average pooling is performed along the spatial dimension to aggregate the two-dimensional spatial features into a channel-dimensional representation vector, and finally, the concentrated representation vector of confusion factors is obtained after normalization calibration.

[0090] S225. Interventional reparameterization of the original tensor based on the causal modulation of the representation vector of the confusion factor set, combined with spatial self-attention calculation, to determine the spatial long-distance dependency features.

[0091] As a preferred embodiment, the interventional reparameterization of the original tensor based on the representation vector of the confusion factor set, which modulates causality, and combined with spatial self-attention computation to determine long-range spatial dependency features, includes the following steps:

[0092] S2251. Using the confounding factor concentrated representation vector, the original tensor is subjected to interventional reparameterization based on causal modulation, element-wise multiplication, and residual stripping to obtain a spatial causal dynamic feature tensor stripped of global clear-sky background noise. The specific execution process of this interventional reparameterization is as follows: First, the confounding factor concentrated representation vector of dimension C×1×1 is broadcast and expanded along the height and width spatial dimensions to make its spatial size completely aligned with the original query, key, and value tensors; then, the expanded representation vector is multiplied element-wise with the original tensor, and channel-level multiplication gating is used to apply weighted suppression to the feature channels dominated by clear-sky background, completing the weighted filtering of causal modulation; then, based on the feature mean of the original tensor in the normal clear-sky low-value background region, the global clear-sky background bias tensor is calculated; finally, the causal modulation tensor and the background bias tensor are subjected to element-wise subtraction residual stripping operation to completely remove the global clear-sky background noise component, and finally obtain a spatial causal dynamic feature tensor stripped of background interference and focused on the core of strong convection.

[0093] S2252. Use the Softmax function to perform spatial self-attention calculation on the spatial causal dynamics feature tensor to obtain a global spatial fusion feature map;

[0094] S2253. The global spatial fusion feature map and the historical input image sequence are input into the gate structure through a convolution operation. After processing, the first input gate, the first input modulation gate and the first output gate are obtained.

[0095] S2254, based on the first input gate, the first input modulation gate and the first output gate, the gating mechanism is used to update the memory unit and output a new hidden state, and the updated memory unit and the new hidden state are used as spatial long-distance dependent features.

[0096] S23. Based on spatial long-distance dependency features and through contextual feature fusion, global features in the time dimension are obtained;

[0097] Specifically, the memory unit M output by ILSAM t The hidden state H generated by the gating operation t The result is fed into the context feature fusion module. and As input, the evolution of radar echoes along the time axis is captured through stages such as feature extraction, feature fusion, and memory unit update. Feature extraction and attention weight calculation are also performed. Its structure is as follows: Figure 6 As shown.

[0098] It should be noted that during the feature extraction stage, and By performing element-wise addition, we obtain the global features in the time dimension. The calculation process is as follows: ;

[0099] As a preferred embodiment, obtaining global features in the time dimension based on spatial long-distance dependency features and through context feature fusion includes the following steps:

[0100] S231. Perform context feature fusion on the updated memory units and new hidden states in the long-distance spatial dependency features to obtain the fused memory units and hidden states.

[0101] S232. The merged memory units and hidden states are added element by element to obtain the global features in the time dimension.

[0102] S24. Based on the cross-frame temporal evolution attention mechanism and combined with the squeezing-excitation mechanism, perform temporal attention calculation and analysis on the global features in the time dimension to extract the time series evolution features.

[0103] Specifically, in the attention recalculation stage, a squeeze-excitation (SE) mechanism is used (this mechanism first uses global average pooling to compress temporal features into a global context vector, then uses a two-layer fully connected network to model nonlinear channel dependencies to capture dynamic evolution patterns, and finally generates an adaptive temporal attention weight matrix M between 0 and 1 through Sigmoid activation). T To achieve precise enhancement and redundancy suppression of key temporal evolution information, adaptive temporal attention weights are calculated, first for... Average pooling is performed, followed by dimensionality reduction through a fully connected layer. Attention weights are then calculated using a shared MLP, and finally activated by a sigmoid function to output the temporal attention weight matrix. The calculation process is as follows:

[0104] ; ; ;

[0105] In the formula, For the dimension reduction mapping matrix, For bias terms, For average pooling operation, This represents the feature vector after average pooling. This represents the feature vector after dimensionality reduction mapping.

[0106] In the feature fusion stage, the temporal attention weight matrix is ​​used. Compared with the original global features By performing element-wise multiplication, we obtain features incorporating the time attention mechanism. The calculation method is as follows:

[0107] ;

[0108] During the memory unit update phase, Interlayer memory features to be updated The input is passed to the gating structure to obtain the second input gate. Second input modulation gate Second output gate The memory unit is updated through a gating mechanism to obtain the updated inter-layer memory unit. After further gating and feature filtering, the final hidden state is output. , , , This represents the corresponding convolution filter matrix. , , For the corresponding bias term, tanh is the hyperbolic tangent activation function, which maps the output to the range [-1, 1] for data normalization. The calculation process is as follows:

[0109] ; ;

[0110] ; ;

[0111] ;

[0112] As a preferred embodiment, the step of performing time attention calculation and analysis on global features in the time dimension based on the cross-frame temporal evolution attention mechanism and combined with the squeeze-excitation mechanism to extract time series evolution features includes the following steps: S241, calculating the time attention weight matrix based on the global features in the time dimension and using the squeeze-excitation mechanism; S242, performing element-wise multiplication of the time attention weight matrix with the global features in the time dimension to obtain features integrated with the time attention mechanism; S243, inputting the features integrated with the time attention mechanism and the inter-layer memory features to be updated into a preset gating structure to obtain a second input gate, a second input modulation gate, and a second output gate; S244, based on the second input gate, the second input modulation gate, and the second output gate, updating the memory unit and outputting the final hidden state through the gating mechanism; using the updated memory unit and the final hidden state as time series evolution features.

[0113] S25. Based on the time series evolution features, the predicted echo image sequence is obtained by frame-by-frame output after multi-level upsampling reconstruction and pixel mapping by the decoder. The time series evolution features are the deep abstract spatiotemporal features output by the TEAM module at the decoder end. The complete generation process from features to image sequence is as follows: (1) After the time series evolution features are extracted by TEAM, the decoder performs multi-level transposed convolution upsampling on the features to gradually restore the spatial resolution of the feature map; at the same time, the spatial detail features of the corresponding scale on the encoder side are fused by skip connections to make up for the loss of detail information caused by the downsampling process. (2) After upsampling to the same spatial size as the original input radar echo, the multi-channel features are compressed into a single channel through a 1×1 convolutional layer, and then a single-frame predicted echo image with the same numerical range as the original echo is obtained through linear mapping of pixel values. (3) The above decoding operation is repeated step by step along the time dimension, and the predicted echo images output at each time point are concatenated in time order to finally obtain a complete predicted echo image sequence.

[0114] Specifically, the Bi-STAN recurrent unit integrates the intra-frame spatial long-range attention mechanism (ILSAM) and the cross-frame temporal evolution attention mechanism (TEAM), and the computational flow is as follows: Figure 7 As shown.

[0115] First, the interlayer hidden state output by the Bi-STAN unit at the previous time step is... Compared with the raw radar echo image input from the outside at the current moment Both are fed into an encoder equipped with ILSAM. The ILSAM mechanism is used to extract long-range spatial dependency features, resulting in the encoded feature I'. t With H' t-1 The calculation process is as follows: ; ;

[0116] The gating structure operation is performed on the encoded features. In the Bi-STAN cyclic cell gating operation described in the following formula, the calculation process is as follows:

[0117] ; ;

[0118] ;

[0119] ; ;

[0120] ;

[0121] In the formula, I t H represents the original radar echo image tensor input at the current time step t, while H t-1 and C t-1These represent the hidden state and the cell state from the previous time step t-1, respectively. The character f t i t g t and o t These represent the forget gate, input gate, input modulation gate, and output gate, respectively, which together constitute the gating structure controlling the flow of information. The computation of these gates depends on a specific convolutional filter matrix (weights) W. f W i W g W o and the corresponding bias term b f b i b g b o Ultimately, through element-wise multiplication operations, the model uses these gating mechanisms to update the current cell state C. t Output the new hidden state H at the current moment. t .

[0122] Update the hidden state after gating calculation Memory cells generated by ILSAM The fused intermediate features are fed into the context feature fusion module to obtain the fused intermediate features. and It is then fed into a decoder integrated into the team to extract time-series evolution features, ultimately yielding the output memory unit for the next layer. and hidden state The calculation process is as follows:

[0123] ; ;

[0124] S3. Input the predicted echo image sequence and the future real image sequence into the pre-constructed total objective function, and obtain the total loss value of the current iteration by calculating the pixel-level error, spatiotemporal distribution deviation and advection divergence penalty term;

[0125] In a preferred embodiment, the step of inputting the predicted echo image sequence and the future real image sequence into a pre-constructed total objective function, and obtaining the total loss value of the current iteration by calculating pixel-level error, spatiotemporal distribution deviation, and advection divergence penalty term includes the following steps:

[0126] S31. Based on the pre-constructed overall objective function, calculate the pixel-level error and spatiotemporal distribution deviation based on the predicted echo image sequence and the future real image sequence, respectively.

[0127] In a preferred embodiment, the calculation of the spatiotemporal distribution deviation includes the following steps:

[0128] Based on the predicted echo image sequence, a non-parametric Gaussian kernel density estimation method is used, combined with the bandwidth of the Gaussian kernel function, to calculate the true probability distribution;

[0129] Based on the true probability distribution, perform differential entropy calculation between it and the probability distribution of the true echo to obtain the spatiotemporal distribution deviation.

[0130] S32. Based on the predicted echo image sequence, determine the motion velocity component of the target to be monitored and calculate the advection divergence penalty term;

[0131] As a preferred embodiment, the step of determining the motion velocity component of the target to be monitored based on the predicted echo image sequence and calculating the advection divergence penalty term includes the following steps: S321, determining the motion velocity component of the target to be monitored in a preset direction based on the predicted echo image sequence; S322, constructing a divergence-free constraint based on the mass conservation principle and the motion velocity component of the target to be monitored in the preset direction; S323, calculating the advection divergence penalty term based on the total number of pixels in the predicted echo image sequence and in combination with the divergence-free constraint.

[0132] S33. Based on pixel-level error, spatiotemporal distribution deviation, and advection divergence penalty term, the total loss value of the current iteration is obtained through weighted calculation.

[0133] It should be noted that the original design intention of STDC Loss (Spatiotemporal Distribution Constraint Joint Loss Function) is to force the model to learn the topological structure of the meteorological field while ensuring numerical accuracy.

[0134] ;

[0135] in, and Used to calculate the i-th predicted image With the i-th real image The pixel-level error between them, where m represents the total number of samples. The calculation method is as follows: ;

[0136] The core innovation of the STDC framework lies in the introduction of differential entropy loss. In radar echo generation, by accurately calculating the probability distribution of predicted image pixel values ​​and optimizing its differential entropy, the spatial arrangement pattern of pixels within the image domain can be effectively constrained. This allows the prediction results to maintain a high gradient while better conforming to the smooth and coherent topological structure of the real meteorological field. It can effectively capture and solidify the spatiotemporal correlation of radar echoes, significantly suppressing detail blurring and strong echo core dissipation problems during the prediction process, and improving the visual quality and meteorological usability of long-term predictions. The calculation process is as follows:

[0137] ;

[0138] To achieve differentiable calculation of differential entropy in neural networks, this invention employs a nonparametric Gaussian kernel density estimation method to approximate the i-th predicted pixel value. True probability distribution , This represents the bandwidth of the Gaussian kernel function. The estimation process considers not only isolated pixels but also the similarity of global samples in the probability density space, as shown in the following equation:

[0139] ;

[0140] In the formula, This represents the value of the j-th predicted pixel.

[0141] Although STDC Loss effectively improves the fidelity of radar echo image structure through differential entropy, current cutting-edge research in the field of meteorological deep learning indicates that in order to further eliminate morphological distortions that violate meteorological laws in long-term forecasts (such as radar echo tearing and anomalous aggregation), it is necessary to explicitly inject physical constraints based on hydrodynamic partial differential equations into the model in addition to statistical loss.

[0142] In macroscopic dynamic meteorology, the short-term motion of radar echoes is mainly controlled by the advection of the atmospheric background wind field. This requires that the motion vector field V=(u,v,w) learned implicitly or explicitly by the network must satisfy the mass conservation and continuity equations, where u, v, and w are the velocity components of the object in the x, y, and z directions, respectively. Within the spatiotemporal scope of small-to-medium scale nowcasting, the compressibility of the atmosphere can be approximately ignored; in this case, the motion wind field should strictly satisfy the divergence-free condition.

[0143] ;

[0144] This invention adds an advection divergence penalty term to the total loss function. The central difference method is used to approximate the velocity field divergence at grid points, thus forcibly constraining the parameter update trajectory of the neural network. The calculation process is shown below:

[0145] ;

[0146] In the formula, N represents the total number of pixels in the predicted image; This indicates a divergence-free condition.

[0147] By incorporating this physical constraint into the aforementioned STDC Loss, not only can non-physical deformations that violate the principles of fluid kinematics be eliminated, but deep learning models can also be endowed with strong meteorological interpretability, ensuring that the radar echoes extrapolated by the model have strict physical consistency in both spatial morphology and temporal evolution.

[0148] Meanwhile, in order to organically combine meteorological physics priors with the data-driven capabilities of deep learning, this method uses a linear weighted combination to incorporate the advection divergence penalty term based on fluid dynamics principles. Compared with the aforementioned data-driven pixel-level mean square error loss and distributed differential entropy loss Seamless integration is performed to construct the overall objective function. The overall objective function is a joint function with the global optimization objectives of jointly minimizing pixel extrapolation error, maximizing the preservation of echo spatiotemporal topology, and enforcing the consistency of fluid motion physical laws. Its specific mathematical expression is as follows:

[0149] ;

[0150] In the formula, , and These are weighting coefficients used to balance the contribution ratios of different loss terms. To make the experimental evidence and academic presentation more focused, subsequent quantitative evaluation and ablation experimental analysis will focus on the actual contribution of each individual loss term to the meteorological-specific index and the improvement effect of the core attention module, rather than on the overall objective function. The numerical evolution process itself undergoes repetitive formula decomposition and redundancy analysis; during the actual offline training phase of the model, the network always operates on the overall objective function. The global minimization is the sole guiding principle for backpropagation and iterative updates of network parameters.

[0151] Among them, at the beginning of training, regarding the weighting coefficients , and The value of is determined using an improved soft optimal uncertainty weighting method, which specifically includes the following steps:

[0152] Step 1: In the training start phase, preset the sharpness control factor. The initial value range is This is used to control the smoothness of the weight distribution; at the same time, all trainable parameters are initialized, and pixel-level mean squared error loss is applied respectively. Differential entropy loss and advection divergence penalty Establish a prior statistics queue for recording its loss values;

[0153] Step 2: Collect statistical uncertainty indicators for each loss term during the warm-up training phase;

[0154] Specifically, three weighting coefficients are fixed. Run the warm-up training process until the values ​​of each loss term tend to stabilize and converge; after the warm-up is completed, statistical analysis is performed on each loss term. , and In recent The mean within each iteration step with standard deviation Standard deviation Defined as a measure of the inherent uncertainty of the task loss;

[0155] Step 3: Based on the uncertainty measure of each loss term, combine the sharpness control factor to perform weighted exponent normalization, and analyze the initial optimal value of each weighting coefficient.

[0156] It should be noted that, after completing the standard deviation statistics for each loss item in step 2, the core task of step 3 is to calculate the standard deviation σ. k (k∈{MSE,DE,div}) serves as a quantitative indicator of the inherent uncertainty of the loss for each task. Using an improved Soft Optimal Uncertainty Weighting (UW-SO) strategy, the initial optimal values ​​of the three weighting coefficients are analytically calculated. The initial values ​​of each weighting coefficient are calculated using the following formula:

[0157] ;

[0158] In the formula, Let σ represent the initial weighting coefficient of the k-th loss term, with a value range of (0,1), and the sum of the three coefficients is always equal to 1. k This represents the standard deviation of the k-th loss term. This represents the sharpness control factor, where k is a general indicator, taking values ​​of MSE, DE, and div, respectively. MSE, DE, and div correspond to the weighted coefficient sets of the three loss terms. Let represent the standard deviation of the j-th loss term.

[0159] Step 4: Based on physical priors, fine-tune the initial weighting coefficients by physical strength anchoring, and output the final initial weighting coefficients.

[0160] Specifically, the initial weights assigned to the advection divergence penalty term in step 3 are compared with a preset physical safeguard threshold. This threshold, set based on prior knowledge of the radar echo prediction mission, represents the minimum priority share of physical constraints in the total loss. If the initial weights are lower than this threshold, it means that the purely data-driven weight allocation strategy may have excessively underestimated the constraining value of fluid dynamics conservation laws for model training. In this case, a compensatory boosting mechanism is activated to raise the weights to the safeguard threshold level, and the resulting weight increment is deducted from the original proportions of each data-driven loss term. This achieves a soft correction of the data-driven weight results by physical priors while maintaining the sum of the three weights at 1. Conversely, if the initial weights have reached or exceeded this threshold, the data-driven results are considered to meet the minimum requirements of physical constraints, requiring no intervention, and the calculation results from step 3 are directly used. Through this anchoring fine-tuning, this method effectively avoids the predicament of physical constraints being suppressed by the data-driven task for a long time due to excessively low weights in the early stage of training, and only being forced to activate in the later stage of training. This ensures that the model always maintains respect for the basic physical laws of meteorology throughout the optimization process. The three weighting coefficients finally output by this step will be used as the initial values ​​in the formal training stage and fed into the total loss function to participate in the subsequent dynamic adaptive adjustment based on gradient balance. This forms a complete weight optimization chain from data-driven static initialization to physical anchoring and then to dynamic adaptation.

[0161] S4. Based on the gradient direction generated by the total loss value of the current iteration, iteratively adjust the parameters of the bidirectional spatiotemporal attention recurrent network in reverse order to obtain the radar near-term prediction model.

[0162] S5. Acquire real-time radar echo image sequences and use the radar nowcasting model to predict future radar echo images.

[0163] The following detailed description of the above technical solution, in conjunction with specific embodiments, includes the following implementation process: The effectiveness of the method of the present invention is verified by two radar echo extrapolation datasets, including the Guangzhou Station radar echo dataset and the CIKM2017 public radar echo dataset. The time interval between the radar echo images in both datasets is 6 minutes. The present invention uses the two datasets to conduct long-term extrapolation experiments with 10 frames of image input and 10 frames of future image output, respectively, to verify the prediction performance of the model in long-term tasks.

[0164] To comprehensively evaluate the model's predictive performance, six meteorological-specific indicators and image quality evaluation indicators were introduced. The meteorological-specific indicators include the average value of the Key Success Index (CSI-AVG↑), the average value of the HydeKee skill score (HSS-AVG↑), the average value of the false alarm rate (FAR-AVG↓), and the average value of the hit rate (POD-AVG↑), all of which are average values ​​at thresholds of 20dBZ, 30dBZ, and 40dBZ. The image quality evaluation indicators include the mean squared error (MSE↓) and the structural similarity (SSIM↑). The arrows after the indicators indicate the performance trend, with the upper arrow indicating higher scores and the lower arrow indicating lower scores and better performance.

[0165] Seven mainstream radar echo extrapolation models were selected as comparison objects, including ConvLSTM, PredRNN, PredRNN++, MIM, PhyDNet, SA-ConvLSTM, and MRP-Net. Quantitative comparison experiments, time-step trend analysis, and visualization comparison were carried out on the Guangzhou Station radar echo dataset and the CIKM 2017 radar echo dataset to verify the superiority of Bi-STAN.

[0166] The quantitative comparison results of each model on the radar echo dataset of Guangzhou Station are shown in Table 1. The Bi-STAN model achieved the best performance in all evaluation indicators. Compared with the second-best model MRP-Net, it improved CSI-AVG by 3.12%, HSS-AVG by 2.54%, and POD-AVG by 8.26%, while reducing FAR-AVG by 5.78%, MSE by 6.55%, and SSIM by 2.81%. This fully demonstrates that Bi-STAN has significant advantages in prediction accuracy, false alarm control, and echo image quality.

[0167] Table 1. Comparison of quantization performance of each model on the Guangzhou station radar echo dataset.

[0168] The trends of MSE and SSIM for each model with the prediction step size are as follows: Figure 8 As shown, Figure 8The horizontal axis, Lead Time, represents the prediction step size, corresponding to the prediction duration increasing step by step. The left sub-icon (a) indicates MSE, which stands for Mean Squared Error. MSE, on the vertical axis, quantifies the pixel-level numerical deviation between the predicted and real images; a lower value indicates higher prediction accuracy. The right sub-icon (b) indicates SSIM, which stands for Structural Similarity Index. SSIM, on the vertical axis, measures the similarity of echo images in terms of structure, brightness, and contrast; a higher value indicates better echo morphology fidelity. In the legend, ConvLSTM is a convolutional long short-term memory network, PredRNN is a predictive recurrent neural network, PredRNN++ is an enhanced predictive recurrent neural network, MIM is a memory-interactive memory network, PhyDNet is a physically constrained decoupled prediction network, SA-ConvLSTM is a spatial attention convolutional long short-term memory network, MRP-Net is a multi-scale echo prediction network, and Bi-STAN is the bidirectional spatiotemporal attention recurrent network proposed in this invention. The Result of Guangzhou Station The dataset represents the results from the Guangzhou station dataset. As the prediction step size increases from 2 to 10, the MSE of all models shows an upward trend, while the SSIM shows a downward trend. Bi-STAN consistently maintains the lowest MSE and the highest SSIM, especially in the long-term prediction stage (step size > 5), where its MSE increases the slowest and its SSIM decreases the most steadily, effectively suppressing the error accumulation problem in long-term predictions. The horizontal axis represents the prediction step size (Lead Time), and the vertical axes represent the mean square error (MSE) measuring the deviation of predicted pixel values ​​and the structural similarity (SSIM) measuring the fidelity of echo morphology, respectively, to visually demonstrate the changing trend of model prediction performance with increasing prediction time. Figure 9 This comparison presents visualization results based on typical sequences from the Guangzhou station radar echo dataset, where the letter T represents a specific future time point predicted by the model. Bi-STAN maintains the structural integrity and accurate intensity distribution of the radar echoes throughout the full-time extrapolation process, without significant blurring, deformation, or diffusion. In contrast, the comparison models exhibit varying degrees of echo structure loss and blurring in high-intensity regions during long-term predictions, further validating the superior visualization and prediction performance of Bi-STAN. Figure 9 The T=6 to T=60 at the top represent the forecast times from the 6th to the 60th minute in the future, with adjacent times 6 minutes apart; the English text in the left rows are the names of different comparison models and benchmarks, with Ground Truth being the true benchmark used to measure the prediction effect; the dBZ in the color scale on the right is the unit of radar reflectivity factor (decibels), used to quantify the intensity of radar echoes, with higher values ​​indicating stronger echoes and greater precipitation potential.

[0169] Table 2 shows the quantitative comparison results of each model on the CIKM 2017 radar echo dataset. Bi-STAN still maintains the best performance in all evaluation metrics. Compared with the second-best model MRP-Net, it improves CSI-AVG by 3.98%, HSS-AVG by 1.77%, and POD-AVG by 18.33%, while reducing FAR-AVG by 19.72%, MSE by 4.61%, and SSIM by 1.98%, demonstrating that the Bi-STAN model has good generalization ability.

[0170] Table 2 Comparison of quantization results for each model on the CIKM 2017 radar echo dataset

[0171] The trends of MSE and SSIM for each model with the prediction step size are as follows: Figure 10 As shown, Figure 10 The Result of CIKM 2017 Dataset is the result of the CIKM 2017 dataset. The results are consistent with the Guangzhou dataset. Bi-STAN can still maintain the best error control and structural similarity in the long-term prediction stage. Its error growth rate is significantly lower than that of the comparison model, which proves the stability of the model in long-term prediction on different datasets.

[0172] The visualization results of a typical sequence from the CIKM 2017 radar echo dataset are as follows: Figure 11 As shown, Bi-STAN can accurately capture the evolution trend of radar echoes in both short and long-term predictions. The echo profile and intensity gradient are highly consistent with the true values. In contrast, the comparison model has problems such as echo morphology distortion and disappearance of weak echo regions in long-term predictions, which verifies the effectiveness of Bi-STAN's generalization ability.

[0173] To verify the effectiveness of the Intra-Frame Spatial Long Range Attention Mechanism (ILSAM), the Cross-Frame Temporal Evolution Attention Mechanism (TEAM), and the Spatiotemporal Distribution Constraint Joint Loss Function (STDC Loss), ablation experiments were conducted based on the Guangzhou Station radar echo dataset. The experimental subjects included PredRNN, MRP-Net, Bi-STAN with ILSAM removed (w / o ILSAM), Bi-STAN with TEAM removed (w / o TEAM), Bi-STAN with unoptimized loss function (w / o STDC), and standard Bi-STAN. The quantization results are shown in Table 3.

[0174] Table 3. Results of Bi-STAN model ablation analysis based on radar echo dataset from Guangzhou Station.

[0175] Ablation experiments show that removing any one of the ILSAM, TEAM, or STDC Loss modules leads to a decrease in all evaluation metrics of the model to varying degrees. This demonstrates that all three modules are key to improving model performance, and their synergistic effect can achieve a dual improvement in spatiotemporal dependency modeling capability and loss constraint capability, enabling Bi-STAN to exhibit the best prediction performance in long-term radar echo extrapolation.

[0176] According to a second embodiment of the present invention, a spatiotemporal cooperative attention radar near-term prediction system based on long-term evolution characteristics is provided, comprising:

[0177] The data acquisition module is used to acquire historical radar echo image sequences of the target to be monitored, and divide them into historical input image sequences and future real image sequences according to a preset number of frames.

[0178] The image prediction module is used to construct a bidirectional spatiotemporal attention recurrent network and input the historical input image sequence into the bidirectional spatiotemporal attention recurrent network. Through feature extraction and fusion at each layer, a predicted echo image sequence is generated. The feature extraction and fusion includes the calculation of the proportion of zero values ​​in the local neighborhood grid and the causal flow intervention analysis based on the backdoor adjustment rule.

[0179] The loss calculation module is used to input the predicted echo image sequence and the future real image sequence into the pre-constructed total objective function, and obtain the total loss value of the current iteration by calculating the pixel-level error, spatiotemporal distribution deviation and advection divergence penalty term;

[0180] The model building module is used to iteratively adjust the parameters of the bidirectional spatiotemporal attention recurrent network based on the gradient direction generated by the total loss value of the current iteration, so as to obtain the radar near-term prediction model.

[0181] The image prediction module is used to acquire real-time radar echo image sequences and use a radar nowcasting model to predict future radar echo images.

[0182] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A spatiotemporal collaborative attention radar near-term prediction method based on long-term evolution characteristics, characterized in that, Includes the following steps: S1. Acquire the historical radar echo image sequence of the target to be monitored, and divide it into historical input image sequence and future real image sequence according to the preset number of frames; S2. Construct a bidirectional spatiotemporal attention recurrent network and input the historical input image sequence into the bidirectional spatiotemporal attention recurrent network. Through feature extraction and fusion at each layer, generate a predicted echo image sequence. The feature extraction and fusion includes the calculation of the proportion of zero values ​​in the local neighborhood grid and the causal flow intervention analysis based on the backdoor adjustment rule. S3. Input the predicted echo image sequence and the future real image sequence into the pre-constructed total objective function, and obtain the total loss value of the current iteration by calculating the pixel-level error, spatiotemporal distribution deviation and advection divergence penalty term; S4. Based on the gradient direction generated by the total loss value of the current iteration, iteratively adjust the parameters of the bidirectional spatiotemporal attention recurrent network in reverse order to obtain the radar near-term prediction model. S5. Acquire real-time radar echo image sequences and use the radar nowcasting model to predict future radar echo images.

2. The spatiotemporal collaborative attention radar near-term prediction method based on long-term evolution characteristics according to claim 1, characterized in that, The construction of a bidirectional spatiotemporal attention recurrent network, and the input of historical input image sequences into the network, followed by the generation of predicted echo image sequences through feature extraction and fusion at each layer, includes the following steps: S21. Construct a bidirectional spatiotemporal attention recurrent network that includes an intra-frame spatial long-range attention mechanism and a cross-frame temporal evolution attention mechanism; S22. Based on the intra-frame spatial long-range attention mechanism, a lightweight structural causal model is constructed to extract features for zero-value background region identification and causal flow intervention analysis of historical input image sequences. Combined with spatial self-attention calculation, spatial long-range dependency features are obtained. S23. Based on spatial long-distance dependency features and through contextual feature fusion, global features in the time dimension are obtained; S24. Based on the cross-frame temporal evolution attention mechanism and combined with the squeezing-excitation mechanism, perform temporal attention calculation and analysis on the global features in the time dimension to extract the time series evolution features. S25. Based on the time series evolution characteristics, the image is output frame by frame after multi-level upsampling reconstruction and pixel mapping by the decoder, and then concatenated to obtain the predicted echo image sequence.

3. The spatiotemporal collaborative attention radar near-term prediction method based on long-term evolution characteristics according to claim 2, characterized in that, The intra-frame spatial long-range attention mechanism, by constructing a lightweight structural causal model, extracts features for zero-value background region identification and causal flow intervention analysis of historical input image sequences, and combines spatial self-attention computation to obtain spatial long-range dependent features, including the following steps: S221. Based on the calculation of meteorological converted physical values ​​and the proportion of zero values ​​in local neighborhood grids, the historical input image sequence is used to identify the normalized clear sky low value background area and determine the spatial confusion factor. S222. Construct a lightweight structural causal model based on spatial confusion factors; S223. Obtain the current hidden state and the previous interlayer memory state of the current bidirectional spatiotemporal attention recurrent network, and generate the original tensor containing the query vector, key vector and value vector through lightweight structure causal model mapping. S224. Perform global average pooling on the global context based on a soft mask to extract the representation vector of the confusion factor set. S225. Interventional reparameterization of the original tensor based on the causal modulation of the representation vector of the confusion factor set, combined with spatial self-attention calculation, to determine the spatial long-distance dependency features.

4. The spatiotemporal collaborative attention radar near-term prediction method based on long-term evolution characteristics according to claim 3, characterized in that, The process of identifying low-value background regions under normal clear skies and determining spatial confusion factors based on meteorological converted physical values ​​and the proportion of zero values ​​in local neighborhood grids includes the following steps: S2211. Based on the historical input image sequence, determine the original radar reflectivity factor value of each spatial grid pixel one by one; S2212. When the original radar reflectivity factor value meets the preset meteorological conversion physical value and the zero value in the local neighborhood grid exceeds the preset threshold, it is determined that the spatial grid area is a normal clear sky low value background area. S2213 defines the low-value background area of ​​normal clear sky as a spatial confusion factor that hinders the modeling of the long-range evolution characteristics of the core space of convection.

5. The spatiotemporal collaborative attention radar near-term prediction method based on long-term evolution characteristics according to claim 3, characterized in that, The interventional reparameterization of the original tensor based on the causal modulation of the representation vector of the confusion factor set, combined with spatial self-attention computation, to determine the spatial long-distance dependency features includes the following steps: S2251. Using the confusion factor to centrally represent the vector, the original tensor is subjected to interventional reparameterization based on causal modulation, element-wise multiplication and residual stripping, so as to obtain the spatial causal dynamic feature tensor stripped of global clear sky background noise. S2252. Use the Softmax function to perform spatial self-attention calculation on the spatial causal dynamics feature tensor to obtain a global spatial fusion feature map; S2253. The global spatial fusion feature map and the historical input image sequence are input into the gate structure through a convolution operation. After processing, the first input gate, the first input modulation gate and the first output gate are obtained. S2254. Based on the first input gate, the first input modulation gate and the first output gate, the gating mechanism is used to update the memory unit and output a new hidden state, and the updated memory unit and the new hidden state are used as spatial long-distance dependent features.

6. The spatiotemporal collaborative attention radar near-term prediction method based on long-term evolution characteristics according to claim 2, characterized in that, The process of obtaining global features in the time dimension based on long-distance spatial dependency features and through contextual feature fusion includes the following steps: S231. Perform context feature fusion on the updated memory units and new hidden states in the long-distance spatial dependency features to obtain the fused memory units and hidden states. S232. The merged memory units and hidden states are added element by element to obtain the global features in the time dimension.

7. The spatiotemporal collaborative attention radar near-term prediction method based on long-term evolution characteristics according to claim 6, characterized in that, The method for extracting time series evolution features by performing temporal attention calculation and analysis on global features in the time dimension based on the cross-frame temporal evolution attention mechanism and combined with the squeezing-excitation mechanism includes the following steps: S241. Calculate the temporal attention weight matrix based on the global features of the time dimension and using the squeeze-excitation mechanism; S242. Multiply the temporal attention weight matrix element-wise with the global features of the time dimension to obtain the features incorporated into the temporal attention mechanism. S243. The features incorporated into the time attention mechanism and the interlayer memory features to be updated are fed into a preset gating structure to obtain a second input gate, a second input modulation gate and a second output gate. S244. Based on the second input gate, the second input modulation gate, and the second output gate, the memory cell is updated and the final hidden state is output through the gating mechanism; the updated memory cell and the final hidden state are used as time series evolution features.

8. The spatiotemporal collaborative attention radar near-term prediction method based on long-term evolution characteristics according to claim 1, characterized in that, The process of inputting the predicted echo image sequence and the future real image sequence into a pre-constructed overall objective function, and obtaining the total loss value for the current iteration by calculating pixel-level error, spatiotemporal distribution deviation, and advection divergence penalty term includes the following steps: S31. Based on the pre-constructed overall objective function, calculate the pixel-level error and spatiotemporal distribution deviation based on the predicted echo image sequence and the future real image sequence, respectively. S32. Based on the predicted echo image sequence, determine the motion velocity component of the target to be monitored and calculate the advection divergence penalty term; S33. Based on pixel-level error, spatiotemporal distribution deviation, and advection divergence penalty term, the total loss value of the current iteration is obtained through weighted calculation.

9. The spatiotemporal collaborative attention radar near-term prediction method based on long-term evolution characteristics according to claim 8, characterized in that, The calculation of the spatiotemporal distribution deviation includes the following steps: Based on the predicted echo image sequence, a non-parametric Gaussian kernel density estimation method is used, combined with the bandwidth of the Gaussian kernel function, to calculate the true probability distribution; Based on the true probability distribution, perform differential entropy calculation between it and the probability distribution of the true echo to obtain the spatiotemporal distribution deviation.

10. The spatiotemporal collaborative attention radar near-term prediction method based on long-term evolution characteristics according to claim 8, characterized in that, The process of determining the velocity components of the target to be monitored based on the predicted echo image sequence and calculating the advection divergence penalty term includes the following steps: S321. Based on the predicted echo image sequence, determine the velocity component of the target to be monitored in the preset direction; S322. Based on the principle of mass conservation, construct divergence-free condition constraints according to the velocity components of the target to be monitored in the preset direction; S323. Based on the total number of pixels in the predicted echo image sequence and combined with the divergence-free condition constraint, calculate the advection divergence penalty term.

Images