Method and system for predicting extreme precipitation events, device and medium

Abstract

The application relates to the technical field of weather forecasting, and provides a prediction method and system, equipment and a medium for an extreme precipitation event based on a space-time attention mechanism, the method comprising the following steps: S1: acquiring meteorological background features; S2: classifying extreme weather according to the meteorological background features; S3: obtaining joint attention by adopting a CNN-LSTM and a joint attention mechanism according to the meteorological background features and a classification result; and S4: obtaining a predicted precipitation condition according to the joint attention, physical constraint embedding and adaptive threshold learning. The scheme fuses multi-scale space-time features, embeds physical constraints and has adaptive attention to extreme events, fills the technical blank of fine classification and prediction of extreme weather, and improves the precision of extreme precipitation prediction and the reliability of disaster early warning. The innovative model with the space-time attention mechanism as the core can solve the above problems, realizes efficient and adaptive intelligent modeling of the extreme precipitation event, and promotes the practical application progress of the meteorological disaster intelligent early warning system.

Description

Technical Field

[0001] This invention relates to the field of meteorological forecasting and processing technology, and in particular to a method, system, device and medium for predicting extreme precipitation events based on a spatiotemporal attention mechanism. Background Technology

[0002] Extreme precipitation events pose a significant threat to people's lives and property, and accurate prediction of extreme precipitation is crucial for disaster prevention and mitigation. However, existing deep learning technologies have many limitations in extreme precipitation prediction, particularly the lack of specialized prediction capabilities for different types of extreme precipitation.

[0003] First, the application of traditional neural networks in precipitation prediction faces the following challenges: Convolutional Neural Networks (CNNs) primarily focus on spatial feature extraction while neglecting the continuity of the temporal dimension; Long Short-Term Memory Networks (LSTMs) improve upon Recurrent Neural Networks (RNNs) in enhancing temporal modeling capabilities, but they can only linearly model spatial and temporal relationships and cannot fundamentally solve the complex multi-scale, strongly nonlinear spatiotemporal coupling problem of precipitation occurrence, performing poorly with high-dimensional spatial data sampling and high-resolution precipitation field input. In summary, traditional CNNs and LSTMs struggle to effectively capture the spatiotemporal dependencies of precipitation.

[0004] Secondly, due to the low frequency and scarcity of extreme precipitation data, existing models perform poorly when handling highly imbalanced data, leading to severe underreporting or false alarms of extreme events. Attention mechanisms, however, can effectively improve the performance of neural networks. By adaptively allocating model computational resources, they enable the network to focus on key spatiotemporal locations, feature channels, and important segments in the input. In extreme precipitation prediction scenarios, attention mechanisms allow models to dynamically identify and highlight extreme precipitation signals, regardless of their sparse and irregular spatiotemporal distribution, thereby effectively improving the recognition rate and prediction accuracy of extreme events.

[0005] Furthermore, purely data-driven deep learning models lack physical constraints. This can lead to predictions that violate fundamental atmospheric physics laws, such as water vapor conservation and energy balance, directly impacting the reliability and practicality of disaster early warning systems. Existing deep learning models exhibit significant shortcomings in modeling extreme events and ensuring physical consistency. For instance, even structures like UNet, which excel in image segmentation, struggle to provide a reasonable description of extreme precipitation when directly applied to precipitation prediction without incorporating the physical mechanisms of weather processes.

[0006] Finally, the weak multi-scale spatiotemporal feature fusion mechanism is another key problem. Extreme precipitation events often involve multiple couplings from local strong convection and topographic influence to large-scale weather systems. Existing models (such as ordinary CNN, Unet, etc.) lack an effective mechanism for fusing features at different spatial and temporal scales, resulting in models only being able to capture features at a single scale and ignoring the interactions between different scales [4]. This directly affects the accurate description of the location, intensity and evolution of extreme precipitation, ultimately weakening the spatial positioning and quantitative capabilities of disaster early warning.

[0007] The main shortcomings of existing technologies also lie in the lack of ability to identify and differentiate extreme weather types. Current precipitation forecasting systems generally use a uniform model structure to handle all types of extreme precipitation, failing to fully consider the unique physical mechanisms of different weather systems such as typhoon precipitation, frontal precipitation, and monsoon precipitation. This "one-size-fits-all" approach leads to a significant decrease in forecast accuracy, especially under complex terrain and variable climate conditions.

[0008] These technical deficiencies severely impact the reliability of extreme precipitation forecasts and hinder the improvement of disaster early warning systems. Therefore, there is an urgent need to develop a prediction method, system, equipment, and medium for extreme precipitation events based on a spatiotemporal attention mechanism. This method should integrate multi-scale spatiotemporal features, embed physical constraints, and adaptively focus on extreme events to fill the technical gap in refined classification and prediction of extreme weather, thereby improving the accuracy of extreme precipitation forecasts and the reliability of disaster early warnings. An innovative model centered on the spatiotemporal attention mechanism can address these challenges by achieving efficient and adaptive intelligent modeling of extreme precipitation events, thus advancing the practical application of intelligent meteorological disaster early warning systems.

[0009] The information disclosed in the background section is only intended to enhance the understanding of the background of this application, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0010] The main objective of this invention is to overcome the problem of low reliability in extreme precipitation forecasts by providing a method, system, device, and medium for predicting extreme precipitation events based on a spatiotemporal attention mechanism. This method integrates multi-scale spatiotemporal features, embeds physical constraints, and adaptively focuses on extreme events, filling the technological gap in refined classification and prediction of extreme weather, and improving the accuracy of extreme precipitation forecasts and the reliability of disaster early warnings. The innovative model centered on the spatiotemporal attention mechanism can address the aforementioned challenges by achieving efficient and adaptive intelligent modeling of extreme precipitation events, thus promoting the practical application of intelligent meteorological disaster early warning systems.

[0011] To achieve the above objectives, the first aspect of this invention provides a method for predicting extreme precipitation events based on a spatiotemporal attention mechanism, comprising the following steps: S1: Obtain meteorological background characteristics; S2: Classify extreme weather based on meteorological background characteristics; S3: Based on meteorological background characteristics and classification results, joint attention is obtained using CNN-LSTM and a joint attention mechanism; S4: Precipitation forecasts are obtained based on joint attention, physical constraint embedding, and adaptive threshold learning.

[0012] According to an exemplary embodiment of the present invention, in step S1, the meteorological background features include: dynamic meteorological variables, static topographic information, and large-scale circulation background field.

[0013] According to an exemplary embodiment of the present invention, in step S2, classifying the weather based on meteorological background characteristics includes: A feature matrix is ​​constructed based on meteorological background characteristics, and dimensionality reduction is performed. Classification was performed using the K-means algorithm and the random forest classification method. The classification results include typhoon precipitation, peak precipitation, monsoon precipitation, and localized severe convection.

[0014] According to an exemplary embodiment of the present invention, the dimensionality reduction is performed using principal component analysis. The random forest classification method uses formula (2): P=Softmax(RF(Fenhanced))(2); Where P represents the probability distribution vector of the current input sample belonging to multiple extreme precipitation types, and F enhanced represents the feature vector after dimensionality reduction and enhancement by PCA, Softmax represents the normalization exponential function, and RF represents the decision function output of the random forest classifier.

[0015] According to an exemplary embodiment of the present invention, in step S3, obtaining joint attention based on meteorological background features and classification results using CNN-LSTM and a joint attention mechanism includes: Based on meteorological background characteristics, the CNN-LSTM method is used to calculate temporal attention, spatial attention, and channel attention respectively; Joint attention is calculated based on classification results, temporal attention, spatial attention, and channel attention; Perform CNN decoding.

[0016] According to an exemplary embodiment of the present invention, in step S3, the calculation of temporal attention, spatial attention, and channel attention using the CNN-LSTM method based on meteorological background features includes: Features are extracted based on meteorological background characteristics; Perform feature fusion; The fused feature computation time attention mechanism and LSTM temporal modeling are used for feature computation spatial attention and channel attention after LSTM temporal modeling.

[0017] According to an exemplary embodiment of the present invention, in step S4, the physical constraint embedding includes: adding a physical constraint term to the loss function, wherein the physical constraint term adopts formula (10): L total =L prediction +λ1L conservation +λ2L gradient +λ3L extrene (10); Among them, L total L represents the total loss function value during model training. prediction To predict losses, the mean squared error (MSE) model is used to calculate the difference between the predicted precipitation and the actual observed precipitation, L. conservation L represents the water vapor conservation constraint. gradient L represents the spatial gradient smoothing constraint. extrene λ1, λ2, and λ3 represent the weights of the physical constraint terms, indicating the weights of the extreme event weights.

[0018] According to an exemplary embodiment of the present invention, in step S4, the adaptive threshold learning adopts formula (11): τ(x, y, t) = f threshold ([H(x,y),S(t),C(x,y)])(11); Where τ(x, y, t) represents the dynamic threshold (unit: mm / h) for determining whether an event is an extreme precipitation event at spatial location (x, y) and time t, H(x, y) represents historical statistical characteristics, S(t) represents seasonal characteristics, C(x, y) represents climate background, and f threshold This represents a pre-trained multilayer perceptron (MLP) or nonlinear regression function, where x represents the spatial coordinate index in the longitude direction, y represents the spatial coordinate index in the latitude direction, and t represents the time index of the prediction time.

[0019] According to a second aspect of the present invention, the present invention provides a prediction system for extreme precipitation events based on a spatiotemporal attention mechanism, comprising: a data acquisition module, an extreme weather classification module, a feature analysis module, and a constraint module; The data acquisition module is used to acquire meteorological background characteristics; The extreme weather classification module is used to classify extreme weather based on meteorological background characteristics; The feature analysis module is used to obtain joint attention based on meteorological background features and classification results using CNN-LSTM and a joint attention mechanism; The constraint module is used to obtain the predicted precipitation based on joint attention, physical constraint embedding, and adaptive threshold learning.

[0020] As a third aspect of the present invention, the present invention provides an electronic device comprising: One or more processors; Storage device for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method for predicting extreme precipitation events based on the spatiotemporal attention mechanism.

[0021] As a fourth aspect of the present invention, the present invention provides a computer-readable medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned method for predicting extreme precipitation events based on a spatiotemporal attention mechanism.

[0022] The advantages of this invention are: This solution significantly improves the prediction accuracy, physical consistency, and computational efficiency of rare extreme precipitation events, providing reliable technical support for meteorological disaster early warning systems. It can be widely applied in areas such as rainstorm warnings, urban flood control, and flash flood prevention. Specifically: 1. The extreme weather classification and preprocessing technology fills the technical gap of the existing technology's "one-size-fits-all" treatment of different types of extreme precipitation. It automatically identifies precipitation types such as typhoons, fronts, and monsoons by meteorological background features, and matches the optimal deep learning algorithm for each type, achieving a technical breakthrough from general prediction to accurate classification prediction.

[0023] 2. The innovative multi-scale spatiotemporal attention mechanism, through 1×1 to 7×7 multi-branch convolution and three-dimensional joint attention, comprehensively surpasses the traditional CNN / LSTM, greatly improving the spatiotemporal feature extraction capability, and can effectively capture different rainfall patterns from local convection to weather scale.

[0024] 3. The multi-constraint loss function ensures that the prediction results conform to atmospheric physical laws, reducing the error of water vapor conservation. Embedding atmospheric physical constraints into deep learning ensures the physical rationality of the prediction results. At the same time, specialized learning strategies for rare extreme events, such as imbalanced learning and adaptive thresholding, optimize extreme value prediction. The adaptive thresholding function improves the accuracy of extreme event identification, thereby improving disaster early warning capabilities. Attached Figure Description

[0025] The above and other objects, features, and advantages of this application will become more apparent from the detailed description of exemplary embodiments with reference to the accompanying drawings. The drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0026] Figure 1 The diagram illustrates the structure of a prediction system for extreme precipitation events based on a spatiotemporal attention mechanism.

[0027] Figure 2 The diagram illustrates the steps of a method for predicting extreme precipitation events based on a spatiotemporal attention mechanism.

[0028] Figure 3 The flowchart illustrates a method for predicting extreme precipitation events based on a spatiotemporal attention mechanism.

[0029] Figure 4 The diagram illustrates the architecture of CNN-LSTM.

[0030] Figure 5 The diagram illustrates the architecture of the joint attention mechanism.

[0031] Figure 6 The diagram illustrates a comparison of extreme precipitation prediction results.

[0032] Figure 7 A schematic diagram of the electronic device is shown.

[0033] Figure 8 A schematic diagram of the structure of a computer medium is shown. Detailed Implementation

[0034] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the embodiments set forth herein; rather, they are provided so that this application will be thorough and complete, and will fully convey the concept of the exemplary embodiments to those skilled in the art. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted.

[0035] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. Numerous specific details are provided in the following description to give a thorough understanding of embodiments of this application. However, those skilled in the art will recognize that the technical solutions of this application can be practiced without one or more of the specific details, or other methods, components, apparatuses, steps, etc., can be employed. In other instances, well-known methods, apparatuses, implementations, or operations are not shown or described in detail to avoid obscuring various aspects of this application.

[0036] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0037] The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be broken down, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0038] It should be understood that although the terms first, second, third, etc., may be used herein to describe various components, these components should not be limited by these terms. These terms are used to distinguish one component from another. Therefore, the first component discussed below may be referred to as the second component without departing from the teachings of this application. As used herein, the term "and / or" includes all combinations of any one and more of the associated listed items.

[0039] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of exemplary embodiments, and the modules or processes in the drawings are not necessarily essential for implementing this application, and therefore cannot be used to limit the scope of protection of this application.

[0040] According to a first specific embodiment of the present invention, the present invention provides a prediction system for extreme precipitation events based on a spatiotemporal attention mechanism, such as... Figure 1 As shown, it includes: a data acquisition module, an extreme weather classification module, a feature analysis module, a constraint module, and a post-processing module.

[0041] The data acquisition module is used to acquire meteorological background characteristics.

[0042] The extreme weather classification module is used to classify extreme weather based on meteorological background characteristics.

[0043] The feature analysis module is used to obtain joint attention based on meteorological background features and classification results using CNN-LSTM and joint attention mechanism.

[0044] The constraint module is used to obtain predicted precipitation based on joint attention, physical constraint embedding, and adaptive threshold learning.

[0045] The post-processing module is used to visualize the prediction results.

[0046] According to a second embodiment of the present invention, the present invention provides a method for predicting extreme precipitation events based on a spatiotemporal attention mechanism, employing the extreme precipitation event prediction system based on a spatiotemporal attention mechanism of the first embodiment. For example... Figure 2 and Figure 3 As shown, it includes the following steps: This scheme employs a unified CNN-LSTM serial hybrid architecture to process all types of extreme precipitation. The specific data flow is as follows: Standardized multi-source meteorological data first enters the extreme weather classification module, outputting classification probability vectors P for four precipitation types; subsequently, the data enters the multi-scale CNN encoder, which extracts multi-scale spatial features in parallel for the spatial data at each time step using four different sizes of convolutional kernels (1×1, 3×3, 5×5, 7×7), and obtains the encoded feature sequence after channel concatenation and fusion; this feature sequence is compressed by a fully connected layer and then enters a multi-layer LSTM module for temporal modeling, learning the movement trajectory and intensity evolution of precipitation spatial patterns in the time dimension, and outputting temporal features.

[0047] The temporal features output by the LSTM are unfolded into spatial feature maps by fully connected layers and then enter a 3D joint attention module to calculate spatial attention, temporal attention, and channel attention respectively. These three are then fused to obtain attention-enhanced features. During this process, the classification probability vector P is multiplied by the channel attention weights to achieve classification-guided feature modulation, enabling the network to dynamically emphasize relevant feature channels according to different precipitation types. Finally, the attention-enhanced features enter the CNN decoder, where transposed convolutions are used to progressively upsample and restore the original spatial resolution. Skip connections are then combined to preserve the encoder's detailed features, outputting the final precipitation prediction field.

[0048] S1: Obtain meteorological background features.

[0049] Meteorological background features include: dynamic meteorological variables, static topographic information, and large-scale circulation background field. The large-scale circulation background field includes the 500hPa geopotential height field, the 850hPa wind field, and the overall water vapor flux.

[0050] In the input data phase, this scheme uses 10-minute resolution meteorological data stored in NetCDF format, fully leveraging the advantages of this format in processing multidimensional scientific data. The input data is divided into three categories: The first category is dynamic meteorological variables, including precipitation P(x,y,t), temperature T(x,y,t), relative humidity RH(x,y,t), and wind field components U(x,y,t) and V(x,y,t). These continuously changing meteorological variables, after quality control and spatial interpolation, can reflect the real-time evolution of atmospheric conditions. The second category is static topographic variables, including digital elevation model DEM(x,y) and slope Slope(x,y). This geographic information is crucial for understanding the lifting and blocking effects of topography on precipitation. The third category is the large-scale circulation background field, including the 500hPa geopotential height field Z500, the 850hPa wind field UV850, and the whole-layer water vapor flux IVT. These data are used to identify weather system types and determine the large-scale driving factors of precipitation. In terms of model input construction, the system integrates dynamic meteorological variables, static topographic information, and large-scale circulation background field into a unified input. Precipitation is the sole target variable for prediction, while the remaining variables are used to assist in characterizing the physical environment in which precipitation forms. The large-scale circulation background field is also primarily used in the extreme weather classification module, and its output class probabilities serve as conditional information to guide subsequent feature modulation.

[0051] In terms of data preprocessing, the system employs the Z-score standardization method, converting the original data into standardized data by calculating the mean μ and standard deviation σ of the training set, and incorporating the minimum value ε = 1 × 10⁻⁶ in the standard deviation calculation. -8 To ensure numerical stability, the time series is organized using a sliding window, with each input series containing 12 time steps corresponding to 120 minutes of historical observations, used to predict precipitation conditions for the next 12 time steps. This design ensures sufficient historical information input while meeting the timeliness requirements of short-term forecasts.

[0052] S2: Classify extreme weather based on meteorological background characteristics.

[0053] Based on the integration of atmospheric circulation pattern recognition theory and machine learning technology, intelligent classification and differentiated processing of different types of extreme precipitation can be achieved.

[0054] In the classification stage, a feature matrix F is constructed based on the large-scale circulation background field. After dimensionality reduction using principal component analysis, a three-order strategy combining physical rule definition, K-means clustering, and random forest algorithm is employed for extreme weather classification. The specific classification is based on the following physical criterion: when the vorticity ζ is greater than 2 × 10⁻⁶... -5 s - ¹When the 850 hPa wind speed V850 is greater than 15 m / s and has a spiral structure, it is determined to be typhoon precipitation; when the temperature gradient | T| greater than 3K / 100km and water vapor flux IVT greater than 250kg·m - ¹·s - ¹ When the wind speed is consistently greater than 10 m / s and exhibits obvious seasonality, it is classified as frontal precipitation; when the convective effective potential energy (CAPE) is greater than 1500 J / kg and lacks large-scale system organization, it is determined to be localized severe convection. K-means clustering is then used to optimize label quality and extract class center features, followed by random forest for fast and stable online classification. The classification result P is used as a conditional vector input to the subsequent CNN-LSTM prediction network to achieve adaptive feature modulation for different weather systems.

[0055] S21: Construct a feature matrix based on meteorological background characteristics and perform dimensionality reduction.

[0056] The characteristic matrix is ​​a multidimensional matrix containing elements such as the 500 hPa geopotential height field, the 850 hPa wind field, and the overall water vapor flux. The 500 hPa geopotential height field, located in the middle troposphere, reflects the location and intensity of upper-level troughs and ridges, and can be used to calculate the vorticity field ζ= v / x- u / The y-squared wind field is used to identify vortex systems such as typhoons and plays a central role in controlling weather dynamics. Located at the top of the atmospheric boundary layer, the 850 hPa wind field reflects the location and intensity of the low-level jet stream. Based on wind speed, jet stream intensity can be determined, and the divergence field can be calculated. •V is used to determine the lower-level convergence region. Overall vapor flux (IVT) comprehensively reflects both vapor content and transport intensity; when IVT is greater than 250 kg·m⁻², it indicates a high concentration of vapor. - ¹·s - ¹ indicates the existence of a strong water vapor transport channel, and water vapor flux divergence can be calculated based on IVT to determine the precipitation area. The above three types of elements provide information on large-scale circulation background, low-level dynamic convergence, and water vapor supply, respectively, and are key elements that provide physical constraints for extreme weather classification and prediction.

[0057] Dimensionality reduction was performed using Principal Component Analysis (PCA) with formula (1): (1); Among them, Z 500 The geopotential height field is 500 hPa, u850 and v850 are the wind fields at 850 hPa, and IVT is the total water vapor flux. Auxiliary meteorological features can be selected from: SLP (sea level pressure) and RH (relative humidity). T is the temperature gradient, CAPE is the convective available potential energy, and Tau is the convective settling time.

[0058] S22: Classify using the K-means algorithm and the random forest classification method.

[0059] A four-category weather pattern classification system was established using physical rule definition, K-means clustering labeling, and random forest classification strategies.

[0060] First, derived diagnostic quantities are calculated based on the original meteorological elements in the feature matrix. Then, physical criteria for four types of extreme precipitation are defined according to atmospheric dynamics principles: when vorticity ζ > 2 × 10⁻⁶. -5 s -1 And wind speed V 850 A velocity greater than 15 m / s and possessing a spiral structure is classified as a cyclonic system, corresponding to typhoon precipitation; when the temperature gradient | T>3K / 100km and water vapor flux IVT>250 kg m - 1 s -1 When the wind speed is consistently greater than 10 m / s and has obvious seasonal characteristics, it is classified as a frontal system; when the convective adjustment time Tau>6h and CAPE>1500J / kg, it is classified as a local severe convection.

[0061] The above physical criteria are used for preliminary labeling of historical long-sequence data. Then, the K-means algorithm is used to perform unsupervised clustering optimization on the labeled samples to minimize the intra-class distance. The role of K-means clustering is to correct the blurred boundary samples in the physical rule labeling and obtain more stable cluster centers of the four weather patterns, which are used as training labels for the random forest. Then, the random forest is used as a classifier for real-time discrimination, and the output classification probability vector P is calculated using formula (2): P = Softmax(RF(F) enhanced ))(2); Where P represents the probability distribution vector of the current input sample belonging to multiple (preferably four) extreme precipitation types, and F enhanced represents the feature vector after dimensionality reduction and enhancement by PCA, RF represents the decision function output of the random forest classifier, and Softmax is the normalized exponential function that converts the output into a normalized probability distribution.

[0062] Physical rules provide interpretable classification criteria, K-means clustering optimizes label quality and extracts class center features, and random forest achieves fast and stable online classification. The classification result P is used as a conditional vector input to the subsequent CNN-LSTM prediction network, guiding the model to perform differential feature modulation.

[0063] As a preferred implementation, for the subsequent analysis algorithm calls, an adaptive algorithm is used to switch the four types of weather pattern matching to a deep learning algorithm library, and the weights are adjusted adaptively using the following formula (3): W final =α1W typhoon +α2W frontal +α3W monsoon +α4W convective (3); Among them, W final W represents the final fusion weight matrix, used to integrate the prediction results of various types of precipitation. typhoon The weight matrix of the typhoon precipitation model emphasizes the spiral structure and cyclonic circulation characteristics. W frontal W represents the weight matrix of the frontal precipitation model, emphasizing temperature gradient and meridional water vapor transport characteristics. monsoon W represents the weight matrix of the monsoon precipitation model, focusing on the long-term dependence of large-scale seasonal circulation. convective α1 represents the weight matrix of the local severe convection model, focusing on the rapid development characteristics of small-scale convective cells. α1, α2, α3, and α4 represent the dynamic fusion coefficients of various precipitation models, which are determined by the probability vector P output by the extreme weather classification module and satisfy α1+α2+α3+α4=1.

[0064] Based on the extreme weather classification results, the system employs a unified multi-scale CNN-LSTM architecture to process all types of precipitation, and achieves differentiated processing through a classification-guided feature modulation mechanism. Specifically, all types of precipitation data undergo feature extraction using a unified multi-scale CNN encoder. This encoder uses convolutional kernels of four different scales (1×1, 3×3, 5×5, and 7×7) in parallel processing, capable of simultaneously capturing point features, local features, mesoscale features, and large-scale features, adapting to the spatial scale characteristics of different precipitation types. The extracted spatial features are then used by an LSTM module to model temporal dependencies, learning the temporal evolution patterns of precipitation processes.

[0065] Differentiated processing for different types of precipitation is achieved through the following mechanism: The probability vector P output by the classification module is processed by an embedding layer to generate modulation coefficients, which are then used to calculate the channel weights in the attention module. For typhoon precipitation, the modulation coefficients enhance the feature channel weights related to large-scale spatial structure (corresponding to 5×5 and 7×7 convolution outputs), allowing the model to focus on the spatial distribution characteristics of spiral cloud systems. For frontal precipitation, the modulation coefficients enhance the feature channel weights related to temporal evolution, focusing on capturing the temporal evolution of linear precipitation bands along the front. For monsoon precipitation, the modulation coefficients enhance the feature channel weights related to long-range dependence, learning the persistence characteristics of seasonal circulation through the attention mechanism. For localized severe convection, the modulation coefficients enhance the feature channel weights related to small-scale spatial structure (corresponding to 1×1 and 3×3 convolution outputs) to capture the rapid development process of small-scale convective cells.

[0066] The mathematical expression of this modulation mechanism is: A'_channel = A_channel ⊙ σ(W_embed · P + b_embed); Where A'_channel is the channel attention weight after classification modulation, A_channel is the original channel attention weight, P is the classification probability vector, W_embed is the learnable embedding weight matrix, b_embed is the bias vector, σ is the Sigmoid activation function, and ⊙ represents element-wise multiplication.

[0067] S3: Based on meteorological background characteristics and classification results, joint attention is obtained using CNN-LSTM and joint attention mechanism.

[0068] In the analysis phase, this scheme employs a unified CNN-LSTM hybrid architecture to process all types of extreme precipitation, and achieves differentiated processing through a classification-guided feature modulation mechanism. For example... Figure 4 As shown, the input data first passes through a multi-scale CNN encoder to extract spatial features. The encoder uses four types of convolutional kernels, 1×1, 3×3, 5×5, and 7×7, to process in parallel, which can simultaneously capture point features, local features, regional features, and large-scale features. Then, the temporal dependency is modeled through an LSTM module to capture the temporal evolution of the precipitation process. Finally, the prediction result is output by a CNN decoder, and the skip connections preserve the detailed features of the encoder.

[0069] The spatiotemporal attention mechanism calculates temporal attention, spatial attention, and channel attention separately, and achieves adaptive weighted fusion of three-dimensional features through a joint attention formula. The classification probability vector P is multiplied by the channel attention weights to form a classification-aware attention distribution, enabling the network to dynamically emphasize relevant feature channels according to different precipitation types, achieving targeted prediction of different types of extreme precipitation while maintaining architectural consistency. The specific steps are as follows: S31: Based on the meteorological background characteristics, the CNN-LSTM method is used to calculate the temporal attention, spatial attention, and channel attention respectively.

[0070] S311: Extract features based on meteorological background characteristics.

[0071] Multi-scale feature extraction employs a parallel multi-branch convolutional method, with each branch having a kernel size of 1×1, 3×3, 5×5, and 7×7. This approach effectively captures the diverse characteristics of extreme precipitation at different spatial scales, significantly enhancing the joint perception capability of localized severe convection and large-scale circulation. Feature maps at each scale are normalized before fusion, and channel stitching and dimensionality reduction operations further improve the compactness and discriminative power of feature representation.

[0072] Multi-scale feature extraction uses formula (4): F s =Conv s ([X) t-nt ]), s∈{1,3,5,7}(4; Among them, F s The feature map X represents scale s. t-nt Conv represents the input sequence within the time window. s This represents a convolution operation with a kernel size of s×s.

[0073] This invention innovatively employs a lightweight convolution design, achieving a balance between computational efficiency and feature extraction capability through depthwise separable convolution technology. Depthwise separable convolution decomposes standard convolution into two steps: depthwise convolution and pointwise convolution, significantly reducing the number of model parameters and computational complexity. This design enables the model to be deployed in resource-constrained environments while maintaining excellent feature extraction performance.

[0074] The encoder employs a dual convolutional block structure, with each block consisting of a 3×3 convolutional layer, a batch normalization layer, a ReLU activation function, and a lightweight convolutional layer. This cascaded structure progressively extracts and refines features, the batch normalization layer ensures the stability of the training process, and the ReLU activation function introduces necessary non-linearity. By stacking multiple such convolutional blocks, the model can learn hierarchical feature representations from low to high levels.

[0075] The multi-scale spatial attention mechanism is one of the core innovations of this invention. The system uses four different scale convolutional kernels—1×1, 3×3, 5×5, and 7×7—in parallel to capture point features, local features, regional features, and large-scale features, respectively. This multi-scale design stems from a deep understanding of the physical processes of extreme precipitation: extreme precipitation often involves multi-scale interactions ranging from microscale convection to synoptic-scale systems. Through the attention mechanism, the model can adaptively learn the importance weights of features at different scales.

[0076] S312: Perform feature fusion.

[0077] Multi-scale feature fusion adopts formula (5); F multi =Concat([F1, F3, F5, F7])(5); Among them, F multi F represents the features after fusion. s The feature map representing scale s.

[0078] S313: The fusion of features and the timing attention mechanism for computation, as well as the LSTM temporal modeling, and the spatial and channel attention for feature computation after LSTM temporal modeling.

[0079] Time attention uses formula (6): A temporal =Sigmoid(LSTM(F) cnn ))(6); Channel attention uses formula (7): A channel =Sigmoid(FC(GAP(F)) lstm )))(7); Spatial attention uses formula (8): A spatial =Conv(Concat(AvgPool(F lstm ),MaxPool(F lstm )))(8; Among them, F cnn F represents the feature sequence output by the multi-scale CNN encoder after feature fusion. lstm This represents the feature map output by the LSTM temporal modeling module. LSTM stands for Long Short-Term Memory Network, GAP stands for Global Average Pooling, FC stands for Fully Connected Layer, AvgPool and MaxPool stand for Average Pooling and Max Pooling, respectively, and Sigmoid stands for Sigmoid activation function.

[0080] A hybrid CNN-LSTM architecture is employed, fully leveraging the advantages of both network structures. The CNN feature extractor first processes the spatial features at each time step, extracting spatial patterns from the original five-dimensional tensor (batch, time, channel, height, and width) through convolutional operations. Subsequently, adaptive pooling layers compress the spatial dimensions to 1×1, yielding the feature vector representation for each time step. The LSTM network is responsible for modeling temporal dependencies, employing a 512-dimensional hidden state and a two-layer stacked structure. The two-layer LSTM design enables the model to capture more complex temporal dynamics; the first LSTM layer learns short-term dependencies, while the second layer focuses on long-term trends. The batch-first approach ensures training efficiency, allowing multiple sequences to be processed in parallel.

[0081] The fusion of spatiotemporal information follows a hierarchical processing principle. First, the multi-scale spatial feature map F extracted by the CNN encoder... multi ∈R B×C×H×W Reorganized into a feature sequence along the time dimension, the sequence is input into a two-layer LSTM module (512 hidden units) for time series modeling. From the feature sequence of 12 consecutive time steps, the movement trajectory, intensity change trend, and range evolution features of the precipitation system are extracted, and the time series feature tensor H is output. time Then, H time Channel attention weights A are generated through global average pooling and a two-layer fully connected network (512→64→512). channel The algorithm adaptively enhances key feature channels related to extreme precipitation, such as water vapor flux, vorticity, and temperature gradient, while suppressing noise channels. Then, the refined features are used to generate a time attention map A through 1×1 convolution and a sigmoid function. temporal This identifies the most critical historical moment for the current prediction. Finally, spatial attention A_spatial, temporal attention A_temporal, and channel attention A_channel are fused through tensor product to obtain joint attention features, achieving synergistic weighting across the three dimensions of "space-time-channel" and completing the comprehensive integration of spatiotemporal information.

[0082] The spatiotemporal attention mechanism not only independently computes spatial and temporal attention but also innovatively introduces a three-dimensional joint attention mechanism, which innovatively integrates attention from three dimensions: space, time, and channel. This achieves precise capture of key features of extreme precipitation. Through tensor multiplication or gating mechanisms, it implements synergistic weighting of the spatial and temporal dimensions, effectively improving the model's response to the suddenness and spatial heterogeneity of extreme precipitation. As a preferred implementation, spatial attention employs a fusion of global average pooling and convolution, temporal attention uses an autoregressive gating structure, and joint attention ensures gradient and feature stability through residual connections and normalization operations.

[0083] Overall Figure 5As shown, the joint attention module contains three parallel processing paths, which respectively calculate spatial attention, channel attention, and temporal attention, and finally output an enhanced feature map by splicing and fusing them.

[0084] Spatial Attention Path: The input feature map undergoes both average pooling and max pooling. The results of these two pooling operations are concatenated, and then two convolutional layers (Conv→ReLU→Conv→ReLU) are used to extract spatially salient features, generating a spatial attention map. Spatial attention assigns weights to each spatial location on the feature map, enabling the model to focus on key areas where precipitation occurs, such as typhoon eyewalls and frontal rainbands.

[0085] Channel attention path: The input feature map is compressed in spatial dimension through average pooling, and then passed through two fully connected layers (fully connected layer → ReLU → fully connected layer) and a sigmoid activation function to generate channel attention weights. The outputs of the channel attention and spatial attention path are multiplied and fused to achieve feature purification. Since different channels represent specific learned features (such as water vapor accumulation regions, temperature gradient edges, vortex structure textures, etc.), channel attention can dynamically enhance the feature channels most important for extreme precipitation prediction and suppress noise and secondary features.

[0086] Temporal attention path: The input feature map is fed into the LSTM module for temporal modeling. The LSTM output is processed through two convolutional layers (Conv→LeakyReLU→Conv→LeakyReLU) and a sigmoid function to generate a temporal attention map. Temporal attention assigns weights to each time step of the input sequence, causing the model to focus on the historical moments that have the most impact on the current prediction, such as critical periods of rapid development or change in the direction of precipitation systems. The temporal attention is then added to and fused with the aforementioned fusion results.

[0087] S32: Calculate joint attention based on classification results, temporal attention, spatial attention, and channel attention.

[0088] Finally, the output consists of three attention fusions: the spatial-channel multiplication fusion result and the temporal addition fusion result are concatenated to produce an enhanced feature map. The overall design logic enables the model to adaptively allocate resources across the three dimensions of "space, channel, and time": spatial attention identifies "which location" requires the most attention, channel attention determines "which features" are most critical, and temporal attention identifies "which moment" is most important. These three aspects complement each other, achieving a comprehensive capture of the spatiotemporal characteristics of extreme precipitation.

[0089] like Figure 5 As shown, temporal attention is based on the fused features F output by the multi-scale CNN encoder. cnnThe calculation involves using LSTM to perform temporal modeling on the feature sequence and then outputting the attention weights at each time step. Spatial attention and channel attention are based on the feature F output by the LSTM temporal modeling module. lstm calculate: The joint attention is calculated; Joint attention uses formula (9): (9); Among them, A joint A represents the joint attention output feature map. temporal Representing a time attention graph, A channel A represents the channel attention weight. spatial Represents a spatial attention map.

[0090] S33: Perform CNN decoding.

[0091] The high-dimensional hidden features output by the aforementioned LSTM temporal modeling are mapped and reshaped through fully connected layers to restore a basic spatiotemporal feature tensor F with spatial dimensions. Based on this, to capture multidimensional key information about the precipitation process, attention weights in three dimensions are calculated: spatial attention A_spatial, focusing on the location information of key precipitation areas; channel attention A_channel, filtering features strongly correlated with extreme precipitation; and temporal attention A_temporal, mapping the global context information of the historical time sequence to channel weights. Subsequently, the basic spatiotemporal feature tensor F is deeply fused with the three attention weights using element-wise multiplication to generate a joint attention feature A_joint. Finally, A_joint is input into a CNN decoder. The decoder employs a multi-layer deconvolution or upsampling structure symmetrical to the encoder, restoring the feature map layer by layer to its original spatial resolution. The feature channels are compressed through a final 1×1 convolutional layer, outputting the predicted precipitation field P_pred for the target time. The value of each grid point in the tensor represents the precipitation intensity at that spatial location at the predicted time.

[0092] S4: Precipitation forecasts are obtained based on joint attention, physical constraint embedding, and adaptive threshold learning.

[0093] The loss function is the core of deep learning model training, used to measure the difference between the model's predicted values ​​and the actual values. In the constraint phase of the CNN-LSTM model training process of this invention, the constraint object of the physical constraint embedding is the predicted precipitation field P_pred output by the CNN decoder. P_pred and the actual observed precipitation P_obs are substituted into the physical constraint loss function. To ensure that the prediction results conform to atmospheric physics, the system embeds multiple physical constraints into the loss function. The water vapor conservation constraint ensures that the total precipitation in the prediction area remains balanced with the input water vapor flux, preventing the model from generating or eliminating water vapor out of thin air. The spatial gradient smoothing constraint penalizes excessively drastic spatial changes in the prediction field through the Laplacian operator, ensuring the spatial continuity of precipitation distribution. The extreme event weighting constraint applies a larger loss weight to extreme precipitation areas exceeding a threshold, forcing the model to focus on the accurate prediction of rare extreme events. The total loss function consists of the prediction loss and the weighted sum of the above three physical constraint losses, with the weights of each constraint term determined through Bayesian optimization. This design not only improves the physical consistency of the model but also significantly reduces false positives and false negatives in extreme precipitation predictions.

[0094] Physical constraint embedding includes adding physical constraint terms to the loss function, where the physical constraint terms are expressed by formula (10): L total =L prediction +λ1L conservation +λ2L gradient +λ3L extrene (10); Among them, L total L represents the total loss function value during model training. prediction To predict losses, the mean squared error (MSE) model is used to calculate the difference between the predicted precipitation and the actual observed precipitation; L conservation L represents the water vapor conservation constraint. gradient L represents the spatial gradient smoothing constraint. extrene λ1, λ2, and λ3 represent the weights of the physical constraint terms, indicating the weights of the extreme event weights.

[0095] The weights λ1, λ2, and λ3 of the physical constraint terms are dynamically learned through Bayesian optimization or adaptive adjustment mechanisms to further enhance the model's generalization ability.

[0096] The loss function design of this invention fully considers meteorological physical laws, and achieves an organic combination of data-driven and physical constraints through a multi-component loss function. The total loss consists of four parts: prediction loss, extreme event weighted loss, spatial gradient smoothing loss, and water vapor conservation loss, each with its specific physical meaning and optimization objective.

[0097] The prediction loss uses mean squared error (MSE) to measure the difference between predicted and actual values, which is the most basic supervision signal. The extreme event-weighted loss identifies areas of extreme precipitation exceeding a threshold and imposes a greater penalty weight on the prediction errors in these areas. This design directly addresses the core challenge of extreme precipitation prediction, forcing the model to focus more on accurately predicting extreme events.

[0098] Spatial gradient smoothing loss calculates the spatial second derivative of the predicted field using the Laplacian operator, penalizing overly abrupt spatial variations. This constraint is based on the principle of atmospheric motion continuity; a true precipitation field should possess a certain degree of spatial smoothness. The convolution operation of the Laplacian kernel effectively detects spatial outliers, guiding the model to generate physically more plausible predictions.

[0099] The water vapor conservation loss reflects the fundamental physical laws of the atmospheric water cycle. By comparing the predicted total precipitation with the actual total precipitation, it is ensured that the model does not generate or disappear water vapor out of thin air. This global constraint is crucial for maintaining the physical consistency of the prediction results, especially in preventing error accumulation during long-term predictions.

[0100] During the model training phase, the gradient of the total loss function with respect to the model parameters is calculated using the backpropagation algorithm, and the parameters are updated using an optimizer (such as Adam) to gradually reduce the value of the loss function until the model converges.

[0101] The threshold function is used to determine whether precipitation at a specific location at a given time constitutes an extreme precipitation event. Adaptive threshold learning operates during the model inference phase, with the constraint being the predicted precipitation field P_pred output by the CNN decoder. A dynamic threshold τ(x,y,t) is calculated for each grid point based on its spatial location (x,y) and prediction time t. The predicted value for each grid point in P_pred is compared with the corresponding dynamic threshold. When the predicted precipitation intensity P_pred(x,y,t) is greater than or equal to the dynamic threshold τ(x,y,t), the system marks it as an extreme precipitation event. Traditional extreme precipitation prediction typically uses a fixed threshold for discrimination, but the characteristics of extreme precipitation vary significantly across different regions and seasons. This invention uses a multi-factor adaptive function to dynamically adjust the extreme precipitation event discrimination threshold, incorporating terrain height, seasonal factors, and climate background, significantly improving the spatial localization and temporal recognition accuracy of extreme events. The threshold function can be jointly optimized through end-to-end backpropagation, taking into account regional differences and temporal distortions.

[0102] Adaptive threshold learning uses formula (11): τ(x, y, t) = f threshold ([H(x,y),S(t),C(x,y)])(11); Where τ(x, y, t) represents the dynamic threshold (unit: mm / h) for determining whether an event is an extreme precipitation event at spatial location (x, y) and time t, H(x, y) represents historical statistical characteristics, S(t) represents seasonal characteristics, C(x, y) represents climate background, and f threshold This represents a pre-trained multilayer perceptron (MLP) or nonlinear regression function, where x represents the spatial coordinate index in the longitude direction, y represents the spatial coordinate index in the latitude direction, and t represents the time index of the prediction time.

[0103] For areas with smaller spatial scales (such as urban blocks or small watersheds), the absolute critical value method (such as 50 mm / h or 100 mm / h) is used to define the threshold, which facilitates disaster prevention decision-making. For areas with larger spatial scales (such as provinces or watersheds), the percentile method (such as the 95th percentile) is used to define the threshold, which can reflect regional climate differences.

[0104] The definition of extreme precipitation varies significantly across different regions and seasons, and fixed thresholds cannot adapt to this spatiotemporal heterogeneity. The adaptive threshold learning mechanism proposed in this invention comprehensively considers three dimensions: historical statistical characteristics, seasonal variations, and climate background.

[0105] Historical statistical features are calculated based on training data from the past three years, typically using the 95th percentile as the criterion for extreme events. This statistical method automatically adapts to the precipitation and climate characteristics of different regions, with different threshold standards for arid and humid areas. The introduction of seasonal features takes into account the seasonal variation in precipitation; for example, the threshold will be correspondingly higher during periods of frequent severe convective weather in summer.

[0106] Climate background information integrates static geographic elements such as topographic elevation, slope, and land use type. Due to orographic uplift, extreme precipitation thresholds are typically higher in mountainous areas than in plains. Through machine learning methods, the model can automatically learn these complex geographic-climate relationships and generate a spatially continuous threshold field.

[0107] like Figure 6 As shown, the comparison results between the method of this invention and the traditional forecasting method in extreme precipitation prediction show that the traditional forecasting method has significant deviations in precipitation intensity prediction. The predicted precipitation range is smoother and more ambiguous than the observed results, and the intensity of the core area of ​​extreme precipitation is insufficiently estimated. Moreover, the error gradually accumulates as the forecast lead time increases. In contrast, the method of this invention has a significant improvement in the spatial distribution of precipitation, and can more accurately capture the morphological structure and intensity evolution of the precipitation system, verifying the effectiveness and superiority of this method in the refined prediction of extreme precipitation.

[0108] This invention employs multi-dimensional evaluation metrics to comprehensively measure model performance. In the post-processing stage, quantitative metrics include mean squared error (MSE) and critical success index (CSI). MSE, as a fundamental metric, directly reflects the average deviation between predicted and actual values; CSI is the core metric for evaluating extreme event predictions, comprehensively considering hits, misses, and false alarms. CSI calculation is based on binary prediction, transforming continuous predictions into a binary classification problem by setting an extreme event threshold. In addition to quantitative metrics (MSE and CSI), the system also provides visualization tools such as spatial distribution maps and time-series comparison maps of the prediction results to help forecasters intuitively understand the model's predictions. This multi-dimensional evaluation system ensures the reliability and practicality of the model in real-world applications.

[0109] This invention demonstrates numerous advantages in engineering implementation. Firstly, it exhibits strong data adaptability, supporting multiple standard meteorological data formats such as NetCDF, and can seamlessly integrate with existing meteorological observation and numerical weather prediction systems. A modular design philosophy permeates the entire system, with clear interfaces between functional modules, facilitating customization and expansion according to specific needs.

[0110] The system fully leverages the advantages of modern deep learning frameworks, supporting GPU-accelerated training and inference. On servers equipped with NVIDIA GPUs, a single prediction takes only a few seconds, fully meeting the timeliness requirements of real-time alerts. Simultaneously, the model supports batch prediction, enabling it to handle prediction tasks for multiple regions or at multiple time points simultaneously, significantly improving computational efficiency.

[0111] Deployment flexibility is another significant advantage. The system can be deployed on cloud servers to provide massively parallel prediction services, or on edge devices for rapid, localized response. This flexible deployment approach adapts to the needs of different application scenarios.

[0112] In summary, this scheme significantly improves the prediction accuracy, physical consistency, and computational efficiency of rare extreme precipitation events, providing reliable technical support for meteorological disaster early warning systems. It can be widely applied in areas such as rainstorm warnings, urban flood control, and flash flood prevention. Specifically: 1. The extreme weather classification and preprocessing technology fills the technical gap of the existing technology's "one-size-fits-all" treatment of different types of extreme precipitation. It automatically identifies precipitation types such as typhoons, fronts, and monsoons by meteorological background features, and matches the optimal deep learning algorithm for each type, achieving a technical breakthrough from general prediction to accurate classification prediction.

[0113] 2. The innovative multi-scale spatiotemporal attention mechanism, through 1×1 to 7×7 multi-branch convolution and three-dimensional joint attention, comprehensively surpasses the traditional CNN / LSTM, greatly improving the spatiotemporal feature extraction capability, and can effectively capture different rainfall patterns from local convection to weather scale.

[0114] 3. The multi-constraint loss function ensures that the prediction results conform to atmospheric physical laws, reducing the error of water vapor conservation. Embedding atmospheric physical constraints into deep learning ensures the physical rationality of the prediction results. At the same time, specialized learning strategies for rare extreme events, such as imbalanced learning and adaptive thresholding, optimize extreme value prediction. The adaptive thresholding function improves the accuracy of extreme event identification, thereby improving disaster early warning capabilities. Specific implementation examples: Special optimizations and applications have been developed for extreme precipitation forecasting in the East Asian monsoon region. The monsoon region has unique climatic characteristics, and extreme weather events such as short-duration heavy rainfall and persistent torrential rains frequently occur during the summer monsoon, posing a severe challenge to disaster prevention and mitigation in the region.

[0116] Adaptive Design for Monsoon Precipitation Characteristics: This embodiment fully considers the spatiotemporal distribution characteristics of monsoon precipitation. Monsoon precipitation exhibits significant seasonality, suddenness, and locality; precipitation intensity can abruptly change from light rain to torrential downpours within a short period. To address this characteristic, the system employs an enhanced temporal attention mechanism, focusing on the time points of abrupt changes in precipitation intensity. By analyzing historical monsoon precipitation data, the model learns characteristic patterns of key meteorological processes such as the passage of monsoon fronts, the establishment of low-level jet streams, and changes in water vapor transport channels.

[0117] Spatially, monsoon precipitation often manifests as the organized development of mesoscale convective complexes (MCCs) and mesoscale convective systems (MCSs). The multi-scale feature extraction module of this invention specifically enhances the weights of 5×5 and 7×7 convolutional kernels to better capture the spatial structure of these mesoscale systems. Simultaneously, the spatial attention mechanism can automatically identify and track the development, merging, and dissipation of convective cells.

[0118] Data Configuration and Preprocessing: In applications in monsoon regions, the system integrates multi-source observation data. In addition to conventional ground-based meteorological station and upper-air sounding data, wind profiler radar data is specifically added to monitor the intensity and altitude changes of the low-level jet stream. Simultaneously, the resolution of topographic data is increased to 500 meters to more accurately reflect the lifting effect of topography on monsoon airflow. During data standardization, seasonal statistical parameters are used, with different means and standard deviations for the summer monsoon period (June-September) and the winter monsoon period (December-February).

[0119] The weighting of the physical constraint loss function was also adjusted according to monsoon characteristics. The weight of the water vapor conservation constraint was increased to 0.2, reflecting the intense water vapor transport process during the monsoon season. The weighting coefficients for extreme events were set in a stepped manner according to different intensity levels: 10 for 50-100 mm / h, 15 for 100-150 mm / h, and 20 for over 150 mm / h, ensuring that the model has good predictive ability for extreme precipitation of different intensities.

[0120] Adaptive Threshold Mechanism: Extreme precipitation thresholds in monsoon regions exhibit significant spatiotemporal variability. This embodiment employs a three-dimensional lookup table method, comprehensively considering geographical location, altitude, and seasonal phase. For example, during the monsoon peak season (mid-May to mid-June), the extreme precipitation threshold for coastal areas is set at 80 mm / h, while the threshold for inland mountainous areas is 60 mm / h during the same period. During the peak monsoon season (July-August), these thresholds are increased to 100 mm / h and 80 mm / h, respectively.

[0121] The threshold also takes into account the impact of accumulated precipitation in the preceding period. When the accumulated precipitation in 72 hours exceeds 200 mm, the extreme precipitation threshold is automatically lowered by 20%, reflecting the influence of soil saturation on the intensity of disaster-causing precipitation. This dynamic adjustment mechanism makes the early warning more closely reflect the actual disaster risk.

[0122] Operational Application Effectiveness: In the rainstorm warning system, the system can provide precipitation forecasts for the next 2 hours, with a temporal resolution of 10 minutes and a spatial resolution of 1 kilometer. The precipitation data and other meteorological data all use a 10-minute temporal resolution. This high temporal resolution is designed to accurately capture the rapid evolution of extreme precipitation events within a short period, which is crucial for achieving the disaster warning function claimed in this invention. The warning threshold is dynamically adjusted according to regional characteristics, with different standards used for mountainous and urban areas. A real-time rolling update mechanism ensures that forecasters are always aware of the latest weather trends.

[0123] Technological Innovation and Promotion Value: The successful application of this embodiment verifies the effectiveness of the spatiotemporal attention mechanism in extreme precipitation prediction. In particular, the three-dimensional joint attention mechanism demonstrates strong adaptability for the multi-scale and highly time-varying characteristics of monsoon precipitation. The introduction of physical constraints not only improves prediction accuracy but also enhances the interpretability of the prediction results, which is crucial for forecasters to understand and use the model's predictions. The system's modular design makes it easy to extend to other climate zones. By adjusting the input data configuration, model parameters, and physical constraint weights, this methodological framework can adapt to different types of extreme precipitation prediction, such as tropical cyclone precipitation, frontal precipitation, and orographic precipitation. This versatility and scalability lay the foundation for the widespread application of this technology.

[0124] According to a third specific embodiment of the present invention, the present invention provides an electronic device, such as... Figure 7 As shown, Figure 7 This is a block diagram illustrating an electronic device according to an exemplary embodiment.

[0125] The following reference Figure 7 To describe an electronic device 700 according to this embodiment of the present application. Figure 7 The electronic device 700 shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0126] like Figure 7 As shown, the electronic device 700 is presented in the form of a general-purpose computing device. The components of the electronic device 700 may include, but are not limited to: at least one processing unit 710, at least one storage unit 720, a bus 730 connecting different system components (including storage unit 720 and processing unit 710), a display unit 740, etc.

[0127] The storage unit stores program code that can be executed by the processing unit 710, causing the processing unit 710 to perform the steps described in this specification according to various exemplary embodiments of this application. For example, the processing unit 710 can perform the steps shown in the second specific embodiment.

[0128] The storage unit 720 may include a readable medium in the form of a volatile storage unit, such as a random access memory unit (RAM) 7201 and / or a cache storage unit 7202, and may further include a read-only memory unit (ROM) 7203.

[0129] The storage unit 720 may also include a program / utility 7204 having a set (at least one) program module 7205, such program module 7205 including but not limited to: an operating system, one or more application programs, other program modules and program data, each or some combination of these examples may include an implementation of a network environment.

[0130] Bus 730 can represent one or more of several types of bus structures, including a memory cell bus or memory cell controller, a peripheral bus, a graphics acceleration port, a processing unit, or a local bus using any of the various bus structures.

[0131] Electronic device 700 can also communicate with one or more external devices 700' (e.g., keyboard, pointing device, Bluetooth device, etc.), enabling users to communicate with devices that interact with electronic device 700, and / or any device (e.g., router, modem, etc.) that allows electronic device 700 to communicate with one or more other computing devices. This communication can be performed via input / output (I / O) interface 750. Furthermore, electronic device 700 can also communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN), and / or public networks, such as the Internet) via network adapter 760. Network adapter 760 can communicate with other modules of electronic device 700 via bus 730. It should be understood that, although not shown in the figures, other hardware and / or software modules can be used in conjunction with electronic device 700, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0132] Through the above description of the embodiments, those skilled in the art will readily understand that the exemplary embodiments described herein can be implemented by software or by combining software with necessary hardware.

[0133] Therefore, according to a fourth specific embodiment of the present invention, the present invention provides a computer-readable medium. For example... Figure 8 As shown, the technical solution according to the embodiments of the present invention can be embodied in the form of a software product. The software product can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, mobile hard drive, etc.) or on a network, and includes several instructions to cause a computing device (such as a personal computer, server, or network device, etc.) to execute the above-described method according to the embodiments of the present invention.

[0134] The software product may employ any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0135] The computer-readable storage medium may include data signals propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. The readable storage medium may also be any readable medium other than a readable storage medium, capable of transmitting, propagating, or transmitting programs for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the readable storage medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof.

[0136] Program code for performing the operations of this invention can be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java and C++, and conventional procedural programming languages ​​such as C or similar languages. The program code can execute entirely on the user's computing device, partially on the user's device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0137] The aforementioned computer-readable medium carries one or more programs, which, when executed by a device, cause the computer-readable medium to perform the functions of the second specific embodiment.

[0138] Those skilled in the art will understand that the above modules can be distributed in the device as described in the embodiments, or they can be modified accordingly and placed in one or more devices that are unique to this embodiment. The modules in the above embodiments can be combined into one module, or they can be further divided into multiple sub-modules.

[0139] Through the description of the above embodiments, those skilled in the art will readily understand that the exemplary embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solutions of the embodiments of the present invention can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, portable hard drive, etc.) or on a network, including several instructions to cause a computing device (such as a personal computer, server, mobile terminal, or network device, etc.) to execute the methods according to the embodiments of the present invention.

[0140] Exemplary embodiments of the present invention have been specifically shown and described above. It should be understood that the present invention is not limited to the detailed structures, arrangements, or implementations described herein; rather, the present invention is intended to cover various modifications and equivalent arrangements contained within the spirit and scope of the appended claims.

Claims

1. A method for predicting extreme precipitation events based on a spatiotemporal attention mechanism, characterized in that, Includes the following steps: S1: Obtain meteorological background characteristics; S2: Classify extreme weather based on meteorological background characteristics; S3: Based on meteorological background characteristics and classification results, joint attention is obtained using CNN-LSTM and a joint attention mechanism; S4: Precipitation forecasts are obtained based on joint attention, physical constraint embedding, and adaptive threshold learning.

2. The method for predicting extreme precipitation events based on a spatiotemporal attention mechanism according to claim 1, characterized in that, In step S2, classifying extreme weather based on meteorological background characteristics includes: A feature matrix is ​​constructed based on meteorological background characteristics, and dimensionality reduction is performed. Classification was performed using the K-means algorithm and the random forest classification method. The classification results include typhoon precipitation, peak precipitation, monsoon precipitation, and localized severe convection.

3. The method for predicting extreme precipitation events based on a spatiotemporal attention mechanism according to claim 2, characterized in that, Principal component analysis was used for dimensionality reduction. The random forest classification method uses formula (2): P=Softmax(RF(F enhanced ))(2); Where P represents the probability distribution vector of the current input sample belonging to multiple extreme precipitation types, and F enhanced represents the feature vector after dimensionality reduction and enhancement by PCA, Softmax represents the normalization exponential function, and RF represents the decision function output of the random forest classifier.

4. The method for predicting extreme precipitation events based on a spatiotemporal attention mechanism according to claim 2, characterized in that, In step S3, obtaining joint attention based on meteorological background features and classification results using the CNN-LSTM method and joint attention mechanism includes: Based on meteorological background characteristics, the CNN-LSTM method is used to calculate temporal attention, spatial attention, and channel attention respectively; Joint attention is calculated based on classification results, temporal attention, spatial attention, and channel attention; Perform CNN decoding.

5. The method for predicting extreme precipitation events based on a spatiotemporal attention mechanism according to claim 4, characterized in that, In step S3, the step of calculating temporal attention, spatial attention, and channel attention using the CNN-LSTM method based on meteorological background features includes: Features are extracted based on meteorological background characteristics; Perform feature fusion; The fused feature computation time attention mechanism and LSTM temporal modeling are used for feature computation spatial attention and channel attention after LSTM temporal modeling.

6. The method for predicting extreme precipitation events based on a spatiotemporal attention mechanism according to claim 5, characterized in that, In step S4, the physical constraint embedding includes: adding a physical constraint term to the loss function, wherein the physical constraint term adopts formula (10): L total =L prediction +λ1L conservation +λ2L gradient +λ3L extrene (10); Among them, L total L represents the total loss function value during model training. prediction To predict losses, a mean square error (MSE) model is used to calculate the difference between the predicted precipitation and the actual observed values, L. conservation L represents the water vapor conservation constraint. gradient L represents the spatial gradient smoothing constraint. extrene λ1, λ2, and λ3 represent the weights of the physical constraint terms, indicating the weights of the extreme event weights.

7. The method for predicting extreme precipitation events based on a spatiotemporal attention mechanism according to claim 1, characterized in that, In step S4, the adaptive threshold learning adopts formula (11): τ(x,y,t)=f threshold ([H(x,y),S(t),C(x,y)])(11); Where τ(x, y, t) represents the dynamic threshold for determining whether an event is an extreme precipitation event at spatial location (x, y) and time t, H(x, y) represents historical statistical characteristics, S(t) represents seasonal characteristics, C(x, y) represents climate background, and f threshold Let x represent a pre-trained multilayer perceptron or nonlinear regression function, where x represents the spatial coordinate index in the longitude direction, y represents the spatial coordinate index in the latitude direction, and t represents the time index of the prediction time.

8. A prediction system for extreme precipitation events based on a spatiotemporal attention mechanism, characterized in that, include: Data acquisition module, extreme weather classification module, feature analysis module, and constraint module; The data acquisition module is used to acquire meteorological background characteristics; The extreme weather classification module is used to classify extreme weather based on meteorological background characteristics; The feature analysis module is used to obtain joint attention based on meteorological background features and classification results using CNN-LSTM and a joint attention mechanism; The constraint module is used to obtain the predicted precipitation based on joint attention, physical constraint embedding, and adaptive threshold learning.

9. An electronic device, characterized in that, include: One or more processors; Storage device for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method for predicting extreme precipitation events based on the spatiotemporal attention mechanism as described in any one of claims 1-7.

10. A computer-readable medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method for predicting extreme precipitation events based on the spatiotemporal attention mechanism as described in any one of claims 1-7.

Images