Multi-task precipitation forecasting model and method fusing adaptive convolution kernel and ConvLSTM-ConvGRU

Abstract

This invention discloses a multi-task precipitation prediction model that integrates adaptive convolutional kernels and ConvLSTM-ConvGRU. It includes a sequence input layer, a multi-scale adaptive convolutional feature extraction layer, a ConvLSTM-ConvGRU parallel dual-branch layer, a multi-task decoding branch and cross-attention layer, a multi-task loss function and model training layer, and a model output and evaluation layer. The multi-scale adaptive convolutional feature extraction layer extracts features from the input data of the sequence input layer. The ConvLSTM-ConvGRU parallel dual-branch layer is used to implement weight allocation between the current input and historical states, historical information filtering, output control, and long-term dependency storage. The multi-task decoding branch and cross-attention layer are used to construct the multi-task decoding branch based on the output of the shared feature extraction layer. This invention has the advantage of achieving accurate prediction of short-term precipitation. This invention also discloses a method for predicting short-term precipitation using a multi-task precipitation prediction model that integrates adaptive convolutional kernels and ConvLSTM-ConvGRU.

Description

Technical Field

[0001] This invention relates to the fields of weather forecasting and neural network prediction technology, and in particular to a multi-task precipitation prediction model combining adaptive convolutional kernels with ConvLSTM and ConvGRU, and its short-term precipitation prediction method. Background Technology

[0002] The core challenge of short-term precipitation forecasting lies in predicting small- and medium-scale weather systems, especially severe convective weather systems. Compared to large-scale weather systems, these small- and medium-scale or severe convective weather systems are characterized by their suddenness, complex and diverse formation mechanisms, and short lifespans, making short-term precipitation forecasting a highly challenging and complex problem in the meteorological field. Disasters such as rainstorms and floods caused by short-term precipitation often result in significant losses to people's lives and property, industrial and agricultural production, and the ecological environment. Therefore, conducting research on high-precision short-term precipitation forecasting has important practical significance and application value.

[0003] The causes of short-term heavy precipitation events are extremely complex. Their occurrence and development are not only closely related to the transport, convergence, and changes of water vapor, but also require comprehensive consideration of various meteorological observation information such as topography, wind direction, and atmospheric dynamics. Furthermore, water vapor observation data is characterized by uneven station distribution, strong spatial heterogeneity of water vapor, and diverse weather system scales. Single-dimensional data analysis or traditional fixed feature extraction methods are insufficient to meet the needs of accurate prediction. In recent years, deep learning theory, with its powerful feature extraction and spatiotemporal modeling capabilities, has been rapidly applied to short-term precipitation prediction research, providing a new path to overcome the limitations of traditional forecasting methods. Related research by scholars at home and abroad has shown that introducing advanced technologies such as machine learning and deep learning can effectively mine potential correlation features in meteorological data, significantly improving the accuracy and precision of short-term precipitation forecasts. However, single neural network models often have limitations when processing complex meteorological data. They struggle to accurately capture the directionality, multi-scale characteristics, and station irregularities of water vapor distribution, and also cannot simultaneously consider the modeling needs of long-term feature dependence and short-term dynamic changes, thus failing to achieve accurate prediction of short-term precipitation.

[0004] Therefore, it is necessary to develop a high-precision short-term precipitation prediction model and method that can accurately capture the directionality, multi-scale characteristics and station irregularities of water vapor distribution, while taking into account both long-term characteristic dependence and short-term dynamic changes. Summary of the Invention

[0005] The primary objective of this invention is to provide a multi-task precipitation prediction model that integrates adaptive convolutional kernels and ConvLSTM-ConvGRU, which can accurately capture the directionality, multi-scale characteristics, and station irregularities of water vapor distribution, while taking into account both long-term feature dependence and short-term dynamic changes, and can achieve accurate prediction of short-term precipitation.

[0006] The second objective of this invention is to provide a method for predicting short-term precipitation using a multi-task precipitation prediction model that integrates adaptive convolutional kernels and ConvLSTM-ConvGRU. This method can significantly improve the accuracy, precision, and comprehensiveness of short-term precipitation forecasts, providing reliable technical support for meteorological operational applications and thus promoting the development of related fields.

[0007] To achieve the first objective of the present invention, the technical solution of the present invention is as follows: a multi-task precipitation prediction model that integrates adaptive convolutional kernel and ConvLSTM-ConvGRU, including a sequence input layer, a multi-scale adaptive convolutional feature extraction layer, a ConvLSTM-ConvGRU parallel dual-branch layer, a multi-task decoding branch and cross-attention layer, a multi-task loss function and model training layer, and a model output and evaluation layer. The sequence input layer is used to receive GNSS water vapor data and meteorological data at multiple time steps; The multi-scale adaptive convolutional feature extraction layer is used to extract features from each data input to the sequence input layer, and to achieve concatenation and weighted fusion of the data features. The ConvLSTM-ConvGRU parallel dual-branch layer is used to implement weight allocation between the current input and historical states, historical information filtering, output control, and long-term dependency storage. Multi-task decoding branches and cross-attention layers are used to construct multi-task decoding branches based on the output of the shared feature extraction layer, enabling multi-dimensional prediction of precipitation-related data. The multi-task loss function and model training layer are used to introduce an uncertainty-weighted dynamic loss adjustment strategy to achieve the training objective of balancing multiple tasks. The model output and evaluation layer is used to output multi-dimensional predictions related to precipitation and to evaluate the accuracy of the prediction results.

[0008] In the above technical solution, the multi-scale adaptive convolutional feature extraction layer includes: Layers 1-2: 3×3 convolutional layers to extract local features, with 64 convolutional kernels. ReLU activation layers and 2×2 max pooling layers are connected after each of the two convolutional layers. Layers 3-4: 5×5 convolutional layers, extracting small-to-medium scale features, with 128 convolutional kernels, also connected to ReLU activation layers and 2×2 max pooling layers; Layers 5-6: 7×7 convolutional layers, extracting mid-scale features, with 256 convolutional kernels, connecting ReLU activation layers and 2×2 max pooling layers.

[0009] In the above technical solution, the ConvLSTM-ConvGRU parallel dual-branch layer includes: ConvLSTM branch: Set up 1 LSTM layer, whose structure includes input gate, forget gate, output gate and memory unit, respectively to implement the weight allocation of current input and historical state, historical information filtering, output control and long-term dependency storage. The branch output dimensions are: batch, time step, 512, H / 8, W / 8. ConvGRU branch: contains update gate, reset gate and candidate hidden state. The update gate determines how much historical information is retained, the reset gate controls the degree of influence of historical information on the current state, and the candidate hidden state combines the current input and the reset historical state. The branch output dimension is consistent with ConvLSTM. Adaptive fusion weight module: The input is the output features of ConvLSTM and ConvGRU. First, the global average pooling features are calculated, and then the fusion weights α and β are learned through two fully connected layers, where α+β=1. The final fusion output is α×ConvLSTM output+β×ConvGRU output. The fusion feature dimensions are: batch, time step, 512, H / 8, W / 8.

[0010] In the above technical solution, the multi-task decoding branch and cross-attention layer include task-specific branches and inter-task cross-attention mechanisms; the inter-task cross-attention mechanism uses the intermediate layer features of each task branch as queries, keys, and values; the attention weight of task i on task j = Where d represents the feature dimension; Q i K represents the intermediate layer features of task i; j T The intermediate layer features represent the features of task j; the enhanced features of task i equal the original features. Among them, V j Represent the characteristics of task j; Task-specific branches include main tasks and side tasks; The main task is to predict precipitation, including: Layer 1: 1×1 convolution from 512 channels to 256 channels to reduce dimensionality; Layer 2: 3×3 convolution upsampling to restore spatial resolution; Layer 3: Continue upsampling to the original resolution; Layer 4: 1×1 convolution to output single-channel precipitation data. The auxiliary tasks include auxiliary task 1, auxiliary task 2 and auxiliary task 3; auxiliary task 1 is the precipitation probability prediction branch, auxiliary task 2 is the precipitation type identification branch, and auxiliary task 3 is the precipitation start time prediction branch. Among them, auxiliary task 1: precipitation probability prediction includes: Layer 1: 1×1 convolution from 512 channels to 128 channels; Layer 2: Global average pooling to obtain spatial feature vectors; Layer 3: Fully connected layer, outputting five probabilities: no rain, light rain, moderate rain, heavy rain, and torrential rain, with the activation function being Softmax; Auxiliary Task 2: Precipitation Type Identification, including: Layer 1: 1×1 convolution from 512 channels to 128 channels; Layer 2: Extracts convection features and stratus features; Layer 3: Fully connected layer, outputs three probabilities: convection, stratus, and mixed type, with Softmax as the activation function; Auxiliary Task 3: Prediction of precipitation start time, including: Layer 1: Attention mechanism in the time dimension, focusing on feature changes before precipitation begins; Layer 2: LSTM layer to process time series features; Layer 3: Fully connected layer, outputting the probability of precipitation starting every 15 minutes in the next 0-3 hours.

[0011] In the above technical solution, the multi-task loss function and model training layer include the multi-task loss function and model training; Multi-task loss functions include: Main task losses: ,in Indicates sample weights, N represents the sample size; y p y represents the model's predicted value; t Represents the actual value; Losses in Auxiliary Task 1: ; Losses in Auxiliary Mission 2: ; Losses in Auxiliary Mission 3: ; Total loss function: Introduces an uncertainty parameter for the science system for each task. The total loss formula is: .

[0012] To achieve the second objective of this invention, the technical solution is as follows: a method for predicting short-term precipitation using the multi-task precipitation prediction model that integrates adaptive convolutional kernels and ConvLSTM-ConvGRU, characterized by comprising the following steps. Step 1: First, acquire GNSS atmospheric water vapor data (PWV) and surface meteorological data; Step 2: Perform data preprocessing on GNSS atmospheric water vapor data (PWV) and surface meteorological data to obtain preprocessed data. Remove observation errors, outliers, and invalid data, unify data format and dimensions, and retain the core correlation characteristics between water vapor and surface meteorological parameters. Step 3: Establish a multi-task neural network model that integrates an adaptive convolutional kernel system and a ConvLSTM-ConvGRU dual-branch system. The adaptive convolutional kernel system includes adaptive anisotropic convolutional kernels, deformable convolutional kernels, multi-scale convolutional kernels, and orientation-sensitive convolutional kernels. The ConvLSTM branch and the ConvGRU branch are adaptively fused through learnable weights and combined with a cross-attention mechanism between tasks to achieve multi-task feature enhancement. Step 4: Train the multi-task neural network model with the fusion of adaptive convolutional kernel system and ConvLSTM-ConvGRU dual branch using the preprocessed data, and use the trained multi-task neural network model with the fusion of adaptive convolutional kernel system and ConvLSTM-ConvGRU dual branch to predict the short-term precipitation in the target area to verify the effectiveness of the model. Step 5: Evaluate the accuracy of the prediction results of the model described in Step 4 based on the recent actual precipitation results in the target area; Specifically, the multi-dimensional forecast results are compared with the recent actual precipitation results of the target area. Meteorological professional indicators such as TS score, POD (probability of detection), and CSI (critical success index) are used to comprehensively evaluate the accuracy of the forecast results from multiple dimensions such as forecast accuracy, success index, and error control.

[0013] In the above technical solution, in step one, the ground meteorological data includes 2m air temperature, 2m relative humidity, 10m wind speed, ground air pressure, and hourly precipitation.

[0014] In the above technical solution, in step two, the preprocessing includes data gross error processing, data normalization, data pruning, and data denoising.

[0015] In the above technical solution, the method for constructing the multi-task precipitation prediction model that integrates adaptive convolutional kernels and ConvLSTM-ConvGRU in step three includes the following steps: First, by rearranging the dimensions of the sequence input layer, the preprocessed GNSS atmospheric water vapor data and surface meteorological data are adjusted into a standard data structure; Subsequently, features are extracted sequentially through a multi-scale adaptive convolution module: first, water vapor and meteorological features at different spatial scales are captured by 3×3, 5×5, and 7×7 multi-scale convolution kernel groups; then, the weights and sampling positions are dynamically adjusted by adaptive anisotropic convolution kernels and deformable convolution kernels to enhance the extraction capabilities of features such as the directionality of water vapor distribution and the sparsity of stations. Next, the extracted multi-scale spatial features are input into the ConvLSTM-ConvGRU parallel dual-branch module, where the ConvLSTM branch is used to store the long-term dependence of precipitation features and the ConvGRU branch is used to capture short-term dynamic changes. The dual-branch feature fusion is completed through the adaptive fusion weight module, and a random deactivation layer is connected to alleviate overfitting. Furthermore, a multi-task decoding branch is constructed based on the fused spatiotemporal features to realize precipitation regression, precipitation probability classification, precipitation type identification, and precipitation start time prediction, respectively. The multi-task feature interaction is enhanced through a cross-attention mechanism between tasks. Finally, the classified meteorological parameters, atmospheric water vapor, and long-term precipitation data are input into the model. The model is then iteratively trained using an uncertainty-weighted multi-task loss function (including precipitation MSE loss, precipitation probability / type cross-entropy loss, and precipitation time binary cross-entropy loss) to complete the construction of the short-term nowcasting precipitation model.

[0016] In the above technical solution, in step four, the trained fusion adaptive convolutional kernel system and the ConvLSTM-ConvGRU dual-branch multi-task neural network model are used to predict short-term precipitation. The specific method is as follows: The preprocessed GNSS atmospheric water vapor data and surface meteorological data are input into the model, and multi-dimensional prediction results (i.e., multi-task precipitation prediction results) are output simultaneously, including: grid-by-grid precipitation, unit mm, 5-level precipitation probability, 3 types of precipitation, and the probability of precipitation starting every 15 minutes in the next 0-3 hours, forming a complete short-term precipitation prediction report. The five levels of precipitation include no rain, light rain, moderate rain, heavy rain, and rainstorm; the three types of precipitation include convective precipitation, stratiform precipitation, and mixed precipitation.

[0017] The present invention has the following advantages: (1) This invention integrates an adaptive convolution kernel system (including adaptive anisotropic convolution kernel, deformable convolution, multi-scale convolution kernel group, and direction-sensitive convolution kernel) with a ConvLSTM-ConvGRU dual-branch model: On the one hand, it constructs an adaptive feature extraction system including adaptive anisotropic convolution kernel, deformable convolution kernel, multi-scale convolution kernel group, and direction-sensitive convolution kernel, dynamically adjusting the receptive field, sampling location, and weight allocation according to the spatial distribution characteristics of GNSS water vapor data, accurately capturing effective information under the characteristics of weather systems at different scales, the directionality of water vapor transport, and the irregular distribution of stations; on the other hand, it fully leverages the modeling advantages of ConvLSTM for long-term spatiotemporal dependence and the efficiency of ConvGRU in capturing short-term dynamic features, improving the model's adaptability to the complex evolution process of short-term precipitation through complementary advantages, and realizing the spatial and temporal integration of adaptive feature extraction and dual-branch spatiotemporal modeling through the collaborative design of adaptive feature extraction and dual-branch spatiotemporal modeling. The deep integration of the two technologies is as follows: the front-end adaptive convolutional kernel system (anisotropic, deformable, multi-scale, and direction-sensitive) reconstructs the non-uniform distribution and directional characteristics of GNSS water vapor data, providing geometric and physical information input for the back-end; the back-end ConvLSTM-ConvGRU dual-branch uses an adaptive fusion mechanism to dynamically balance the tracking of long-term dependence on large-scale systems and the capture of short-term abrupt changes in small- and medium-scale convection, accurately transforming the "spatial directionality" of the front-end into "temporal path prediction"; the two technologies, combined with the cross-attention mechanism between tasks and the uncertainty-weighted loss, form a closed loop from accurate feature extraction to spatiotemporal logical consistency and then to consistent multi-dimensional output, effectively breaking through the prediction bottleneck caused by the fixed receptive field, the separation of long-term and short-term modeling and the isolation of tasks in traditional models in complex terrain, significantly improving the accuracy and reliability of short-term precipitation in location of precipitation area, intensity forecast and start time prediction, and realizing multi-dimensional prediction of short-term precipitation; (2) This invention designs a ConvLSTM-ConvGRU dual-branch adaptive fusion mechanism, which improves the spatiotemporal feature modeling capability by dynamically balancing long-term dependencies and short-term dynamic features through learnable weights, and realizes high-precision, multi-dimensional forecasts of short-term precipitation (0-6 hours). At the same time, it outputs precipitation amount, precipitation probability, precipitation type, and precipitation start time, providing decision support for disaster prevention and mitigation, hydrological scheduling, etc.; and solves the core problems such as insufficient spatial feature extraction, single spatiotemporal dependency modeling, fragmented multi-task prediction, difficulty in balancing task weights, and inaccurate prediction of heavy precipitation. (3) This invention constructs a multi-task framework of "main task + auxiliary task", introduces a cross-attention mechanism between tasks, and realizes synchronous prediction of precipitation, precipitation probability, precipitation type and precipitation start time, so as to meet the comprehensive forecast requirements. (4) This invention proposes an uncertainty-weighted multi-task loss function, which automatically adjusts the training weights of each task, strengthens the focus on key scenarios such as heavy precipitation, and improves the core forecast accuracy. Attached Figure Description

[0018] Figure 1 This is a diagram of the multi-task network model architecture of the present invention, which combines adaptive convolutional kernels with ConvLSTM and ConvGRU models. Figure 2 This is a schematic diagram of a short-term precipitation prediction method based on a multi-task neural network model combining adaptive convolutional kernels with ConvLSTM and ConvGRU, as described in this invention. Figure 3 This is a comparison chart of the results of short-term precipitation prediction using this invention at a certain station in a certain city, and the actual measured results, in an embodiment of this invention. Figure 4 This is a scatter plot comparing the predicted short-term precipitation results from this invention with the measured results from a station in a certain city, as described in an embodiment of the invention. Detailed Implementation

[0019] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings. However, these descriptions do not constitute a limitation of the present invention and are merely illustrative. The advantages of the present invention will become clearer and easier to understand through this description.

[0020] In this invention, the adaptive anisotropic convolutional kernel serves as the core for feature extraction in ConvLSTM and ConvGRU networks. First, historical GNSS water vapor data is collected, and the gradient variation patterns of water vapor in the horizontal and vertical directions are statistically analyzed to identify the main directional features of water vapor distribution. Then, based on the GNSS station density distribution map, a smaller receptive field is used in densely populated areas, while the receptive field is expanded in sparsely populated areas. Combined with topographic elevation data, the vertical convolutional weights are enhanced in areas with large topographic gradients. Using reanalysis data to obtain dominant wind direction information, wind-sensitive convolutional kernels are designed, increasing weights along the wind direction. Finally, historical water vapor transport paths are statistically analyzed, and higher feature extraction weights are set in the directions of major water vapor channels. This invention calculates the spatial offset between the irregular location of each GNSS station and the standard grid points; designs an offset learning network, with the station distribution feature map as input and a two-dimensional offset vector for each sampling point as output; adds the learned offset to the sampling position of the standard convolution to obtain new sampling coordinates; and uses bilinear interpolation to obtain feature values ​​at non-integer coordinate positions, so that the receptive field of the convolution kernel can be adaptively deformed according to the station distribution. This invention sets up convolutional kernel groups of different sizes to capture weather system features at different spatial scales, with the specific configuration as follows: 3×3 convolution kernel group: specifically designed to capture local convective systems (spatial scale of about 10-30 km) and identify small-scale water vapor convergence and rapid changes in boundary layer water vapor; 5×5 convolution kernel group: captures small-to-medium scale systems (spatial scale of about 50-100 kilometers) and extracts features such as mesoscale convective complexes and squall lines; 7×7 convolution kernel group: Identifies mesoscale systems (spatial scale approximately 100-300 km) and captures features of weather systems such as frontal systems and low-pressure troughs; The implementation method is as follows: the input of the same layer is fed into convolutional kernels of different scales, and the extracted multi-scale features are combined by splicing or weighted fusion. The fusion weights can be adaptively adjusted according to different weather types. Simultaneously, eight directional convolutional kernels (east, south, west, north, or four diagonal kernels) are designed, with each kernel having a longer receptive field in that direction; This invention achieves high-precision, multi-dimensional forecasting of short-term precipitation (0-6 hours) through the above-mentioned technical solutions, and outputs precipitation amount, precipitation probability, precipitation type, and precipitation start time, providing decision support for disaster prevention and mitigation, hydrological scheduling, etc.; it solves the core problems of existing technologies such as insufficient spatial feature extraction, single spatiotemporal dependency modeling, fragmented multi-task prediction, difficulty in balancing task weights, and inaccurate prediction of heavy precipitation.

[0021] See appendix Figure 1 It can be seen that a multi-task precipitation prediction model integrating adaptive convolutional kernels and ConvLSTM-ConvGRU includes a sequence input layer, a multi-scale adaptive convolutional feature extraction layer, a ConvLSTM-ConvGRU parallel dual-branch layer, a multi-task decoding branch and cross-attention layer, a multi-task loss function and model training layer, and a model output and evaluation layer. The sequence input layer is used to receive GNSS water vapor data and meteorological data at multiple time steps; the data dimensions are defined as: batch size, number of time steps, number of channels, height, and width; The multi-scale adaptive convolutional feature extraction layer is used to extract features from each data input from the sequence input layer, and realize the concatenation and weighted fusion of the data features; the core components of the multi-scale adaptive convolutional feature extraction layer include adaptive anisotropic convolution kernel, deformable convolution kernel, multi-scale and orientation-sensitive convolution kernel; The ConvLSTM-ConvGRU parallel dual-branch layer is used to implement weight allocation between the current input and historical states, historical information filtering, output control, and long-term dependency storage. Multi-task decoding branches and cross-attention layers are used to construct multi-task decoding branches based on the output of the shared feature extraction layer, enabling multi-dimensional prediction of precipitation-related data. The multi-task loss function and model training layer are used to introduce an uncertainty-weighted dynamic loss adjustment strategy to achieve the training objective of balancing multiple tasks. The model output and evaluation layer is used to output multi-dimensional predictions related to precipitation and to evaluate the accuracy of the prediction results. This invention integrates an adaptive convolutional kernel system (including adaptive anisotropic convolutional kernels, deformable convolutions, multi-scale convolutional kernel groups, and orientation-sensitive convolutional kernels) with a ConvLSTM-ConvGRU dual-branch model to improve core forecast accuracy. This improvement is primarily achieved through: the system enhancing spatial distribution accuracy via adaptive anisotropic convolutional kernels; enhancing sparse region accuracy via deformable convolutional kernels; enhancing accuracy for precipitation of different intensities via multi-scale convolutional kernel groups; enhancing movement path accuracy via orientation-sensitive convolutional kernels; enhancing time-dependent accuracy via the ConvLSTM-ConvGRU dual-branch model; enhancing accuracy for complex processes via an adaptive fusion mechanism; enhancing multi-task consistency via inter-task cross-attention; and enhancing heavy precipitation accuracy via uncertainty-weighted loss. This results in: improved CSI / TS / HSS for precipitation presence / absence determination; reduced RMSE / MAE for precipitation values; improved SSIM for spatial precipitation distribution and reduced positioning errors; improved classification accuracy for precipitation type identification; reduced time error for precipitation start time; and improved CSI for heavy precipitation events.

[0022] Furthermore, the multi-scale adaptive convolutional feature extraction layer includes: Layers 1-2: 3×3 convolutional layers, extracting local features, with 64 convolutional kernels. After each of the two convolutional layers, ReLU activation layers and 2×2 max pooling layers are connected (the ReLU function expression is f(x)=max(0,x), and its mapping method is to change values ​​less than zero to zero while keeping other values ​​unchanged. It has the advantages of simple calculation and fast training convergence speed). Layers 3-4: 5×5 convolutional layers, extracting small-to-medium scale features, with 128 convolutional kernels, also connected to ReLU activation layers and 2×2 max pooling layers; Layers 5-6: 7×7 convolutional layers, extracting mid-scale features, with 256 convolutional kernels, connecting ReLU activation layers and 2×2 max pooling layers.

[0023] Furthermore, the ConvLSTM-ConvGRU parallel dual-branch layer includes: ConvLSTM branch: Set up 1 LSTM layer (the number of LSTM layers is generally 1-4 layers. Simply increasing the number of layers will not effectively improve the prediction accuracy and will lead to a slow model training speed). Its structure includes an input gate, a forget gate, an output gate, and a memory unit, which respectively realize the weight allocation of the current input and the historical state, the filtering of historical information, the control of output, and the long-term dependency storage. The branch output dimensions are: batch, time step, 512, H / 8, W / 8. ConvGRU branch: contains update gate, reset gate and candidate hidden state. The update gate determines how much historical information is retained, the reset gate controls the degree of influence of historical information on the current state, and the candidate hidden state combines the current input and the reset historical state. The branch output dimension is consistent with ConvLSTM. The adaptive fusion weight module takes as input the output features of the ConvLSTM and ConvGRU branches (both including batch, time step, 512, H / 8, W / 8), calculates the global average pooling features first, and then learns the fusion weights α and β through two fully connected layers, where α+β=1. The final fusion output is α×ConvLSTM output + β×ConvGRU output, and the fusion feature dimensions are: batch, time step, 512, H / 8, W / 8.

[0024] Furthermore, the multi-task decoding branch and cross-attention layer include task-specific branches (including precipitation occurrence probability prediction task, precipitation amount prediction task, and rainfall start time prediction task) and cross-attention mechanism between tasks; The inter-task cross-attention mechanism uses the intermediate layer features of each task branch as the query, key, and value; the attention weight of task i on task j = Where d represents the feature dimension, used for scaling to prevent gradient vanishing; Qi represents the intermediate layer features (query vector / matrix) of task i; used to calculate similarity by matching with the keys of all related tasks (including itself or other tasks j); K j T The intermediate layer features (key vectors / matrices) of task j are represented by a dot product with Qi to calculate the matching degree between the features of task i and task j; the enhanced features of task i = the original features. Where Vj represents the features of task j; the intermediate layer features (value vector / matrix) from task j are used to calculate attention weights through QK, and these weights are applied to Vj to perform a weighted summation of the features of task j, thus integrating valuable information from task j into task i; the present invention achieves information sharing and feature enhancement between tasks through the above technical solution, dynamically learns the correlation between tasks through the cross-attention mechanism, and improves the overall performance and consistency of multi-task prediction; Task-specific branches include main tasks and side tasks; The main task is to predict precipitation, including: Layer 1: 1×1 convolution from 512 channels to 256 channels to reduce dimensionality; Layer 2: 3×3 convolution upsampling to restore spatial resolution (256→128 channels); Layer 3: Continue upsampling to the original resolution (128→64 channels); Layer 4: 1×1 convolution outputs single-channel precipitation (regression value, in mm), using ReLU activation function to ensure non-negative output values; The auxiliary tasks include auxiliary task 1, auxiliary task 2 and auxiliary task 3; auxiliary task 1 is the precipitation probability prediction branch, auxiliary task 2 is the precipitation type identification branch, and auxiliary task 3 is the precipitation start time prediction branch. Among them, auxiliary task 1: precipitation probability prediction includes: Layer 1: 1×1 convolution from 512 channels to 128 channels; Layer 2: Global average pooling to obtain spatial feature vectors; Layer 3: Fully connected layer, outputting five probabilities: no rain, light rain, moderate rain, heavy rain, and torrential rain. The activation function is Softmax, but other activation functions can be used according to actual needs. Auxiliary Task 2: Precipitation Type Identification, including: Layer 1: 1×1 convolution from 512 channels to 128 channels; Layer 2: Extracts convection features (features from high convolutional layers) and stratus features (smoothing features from low layers); Layer 3: Fully connected layer, outputting probabilities of three types: convection, stratus, and hybrid, with the activation function being Softmax. Other activation functions can be used according to actual needs. Auxiliary Task 3: Prediction of precipitation start time, including: Layer 1: Attention mechanism in the time dimension, focusing on feature changes before precipitation begins; Layer 2: LSTM layer to process time series features; Layer 3: Fully connected layer, outputting the probability of precipitation starting every 15 minutes in the next 0-3 hours.

[0025] Furthermore, the multi-task loss function and model training layer include the multi-task loss function and model training; Multi-task loss functions include: Main task loss (precipitation): ,in Indicates sample weights, L1 represents the mean absolute error (i.e., the main task loss); N represents the number of samples; y p y represents the model's predicted value; t Represents the actual value; Losses in Auxiliary Task 1: Where L2 represents the loss of auxiliary task 1, corresponding to the precipitation probability prediction; Losses in Auxiliary Mission 2: Where L3 represents the loss of auxiliary task 2, corresponding to precipitation type identification; Losses in Auxiliary Mission 3: Where L4 represents the loss of auxiliary task 3, corresponding to the predicted start time of precipitation; Total loss function: Introduces an uncertainty parameter for the science system for each task. The total loss formula is: Among them, L total The total loss function is represented by σ1, σ2, σ3, and σ4, which represent the uncertainty parameters of the main task, auxiliary task 1, auxiliary task 2, auxiliary task 3, and auxiliary task 4, respectively. The total loss is used to calculate the loss during model training, calculated after each forward propagation, and serves as the source of gradients for backpropagation. At the same time, the total loss can also guide the direction of model parameter updates, automatically balance the weights of each task through the uncertainty parameters, and calculate gradients for backpropagation and parameter updates, thereby improving the model's prediction accuracy.

[0026] See appendix Figure 2 It can be seen that the method for predicting short-term precipitation using the multi-task precipitation prediction model that integrates adaptive convolutional kernels and ConvLSTM-ConvGRU includes the following steps: Step 1: First, acquire GNSS atmospheric water vapor data (PWV) and surface meteorological data; Step 2: Perform data preprocessing on the GNSS atmospheric water vapor data PWV and surface meteorological data obtained in Step 1 to obtain preprocessed data. Remove observation errors, outliers and invalid data, unify data format and dimensions, and retain the core correlation characteristics between water vapor and surface meteorological parameters. Step 3: Establish a multi-task neural network model integrating an adaptive convolutional kernel system and a ConvLSTM-ConvGRU dual-branch architecture. The adaptive convolutional kernel system includes adaptive anisotropic convolutional kernels, deformable convolutional kernels, multi-scale convolutional kernels, and direction-sensitive convolutional kernels. Specifically, the adaptive anisotropic convolutional kernel dynamically adjusts weights based on the directional characteristics of water vapor distribution; the deformable convolutional kernel adaptively adjusts the sampling position according to the irregular location of GNSS stations; the multi-scale convolutional kernel captures water vapor and meteorological features at different spatial scales; and the direction-sensitive convolutional kernel sets higher feature extraction weights along the prevailing wind direction. The ConvLSTM and ConvGRU branches are connected via... This invention achieves multi-task feature enhancement by learningable weight adaptive fusion and inter-task cross-attention mechanism. It designs a multi-type adaptive convolutional kernel combination system to dynamically adapt to the anisotropic, multi-scale, and directional characteristics of water vapor spatial distribution, solving the problem that traditional fixed convolutional kernels cannot adapt to complex water vapor fields. The feature maps extracted by the adaptive convolutional kernels are input into a dual-branch structure to capture long-term and short-term spatiotemporal dependencies, and then fused through learnable weights to balance the globality and dynamism of temporal prediction. Four parallel decoding branches are constructed, corresponding to four prediction tasks, and combined with an inter-task cross-attention mechanism to achieve feature interaction and enhancement between tasks, solving the problem of limited prediction dimensions in traditional single-task models. The multi-task precipitation prediction model of this invention, which integrates adaptive convolutional kernels and ConvLSTM-ConvGRU, accurately captures effective information on weather system characteristics at different scales, water vapor transport directionality, and irregular station distribution through the aforementioned adaptive convolutional kernel system. By leveraging the complementary advantages of ConvLSTM and ConvGRU, the model's adaptability to the complex evolution of short-term precipitation is improved. Simultaneously, it outputs grid-by-grid precipitation, five-level precipitation probability, three types of precipitation, and the precipitation start probability every 15 minutes for the next 0-3 hours, thereby improving the accuracy and precision of short-term precipitation forecasts. Step 4: Use the preprocessed data from Step 2 to train the multi-task neural network model that integrates the adaptive convolutional kernel system and the ConvLSTM-ConvGRU dual-branch model, and use the trained multi-task neural network model that integrates the adaptive convolutional kernel system and the ConvLSTM-ConvGRU dual-branch model to predict the short-term precipitation in the target area and verify the effectiveness of the model. Step 5: Evaluate the accuracy of the prediction results of the model described in Step 4 based on the recent actual precipitation results in the target area; Specifically, the multi-dimensional forecast results are compared with the recent actual precipitation results of the target area. Meteorological professional indicators such as TS score, POD (probability of detection), and CSI (critical success index) are used to comprehensively evaluate the accuracy of the forecast results from multiple dimensions such as forecast accuracy, success index, and error control.

[0027] Furthermore, in step one, the surface meteorological data includes 2m air temperature, 2m relative humidity, 10m wind speed, surface air pressure, and hourly precipitation; The method for acquiring GNSS atmospheric water vapor data and surface meteorological data includes the following steps: First, GNSS atmospheric water vapor data is acquired, specifically: precise single-point positioning calculations are performed on GNSS observation data to obtain the total tropospheric zenith delay (ZTD); the static zenith delay (ZHD) is calculated based on the Hopfield model using surface pressure, air temperature, and altitude; the wet zenith delay (ZWD) is obtained from ZWD = ZTD - ZHD; and the GNSS atmospheric water vapor data PWV is calculated from PWV = Π × ZWD, where the water vapor conversion coefficient Π is a dimensionless coefficient based on temperature calculations; the surface meteorological data (2m air temperature, 2m relative humidity, 10m wind speed, surface air pressure, and hourly precipitation) are obtained from observations at surface meteorological stations.

[0028] Furthermore, in step two, preprocessing includes data gross error handling, data normalization, data pruning, and data denoising, which are used to reduce interference from irrelevant data, improve data quality, and reduce computational load. The gross error processing is based on the Laida criterion. If any sub-data to be tested satisfies formula (1), it is determined that the sub-data to be tested has a gross error, and the GNSS atmospheric water vapor data and ground meteorological data of the historical sub-time period corresponding to the sub-data to be tested are removed. ; in, This is the sub-data to be examined, which is a sub-data from GNSS atmospheric water vapor data and surface meteorological data. To and The average of all subdata belonging to the same data category. To and The standard deviation of all subdata belonging to the same data category and The total number of all sub-data belonging to the same data category; Specifically, the Raida criterion, also known as the 3σ criterion, can be used to determine whether gross errors exist for any given historical sub-period of GNSS atmospheric water vapor data or surface meteorological data. GNSS atmospheric water vapor data or surface meteorological data may consist of multiple sub-data sets, therefore, gross error judgment needs to be performed separately for each sub-data set. The average value of formula (1) and standard deviation All are calculated based on the same seed data; if any one of the sub-data to be tested... If the condition is satisfied with formula (1), then the sub-data to be tested is determined. Gross errors exist; data points containing gross errors, including GNSS atmospheric water vapor data and surface meteorological data, will also be judged to have gross errors and therefore need to be removed and no longer used for model training; gross error processing can ensure the validity of the data, thereby ensuring the correctness of model training.

[0029] Furthermore, in step three, the method for constructing the multi-task precipitation prediction model that integrates adaptive convolutional kernels and ConvLSTM-ConvGRU includes the following steps: First, through the sequence input layer [Batch size (Batch=32), Time step (Time=24, corresponding to 24 hours of continuous data), Number of channels (Channel=4), Height (Height=200), Width (Width=200)], the preprocessed GNSS atmospheric water vapor data and surface meteorological data are adjusted to a standard data structure of [Batch size, Time step, Number of channels, Height, Width] by dimensional rearrangement, ensuring the consistency of data temporal and spatial order, and providing standardized input for subsequent feature extraction; Subsequently, features are extracted sequentially through a multi-scale adaptive convolution module: first, 3×3, 5×5, and 7×7 multi-scale convolution kernel groups (paired with ReLU activation layers and 2×2 max pooling layers) capture water vapor and meteorological features at different spatial scales. Then, adaptive anisotropic convolution kernels and deformable convolution kernels are used to dynamically adjust weights and sampling positions, enhancing the extraction capabilities of features such as the directionality of water vapor distribution and the sparsity of stations. Multi-scale, multi-directional, and adaptive spatial features are extracted from the original meteorological data, providing high-quality input for subsequent spatiotemporal sequence modeling. This invention extracts features through an adaptive convolution kernel system, which includes adaptive anisotropic convolution kernels, deformable convolution kernels, multi-scale convolution kernel groups, and direction-sensitive convolution kernels. This system dynamically adapts to the anisotropic, multi-scale, and directional features of water vapor spatial distribution, solving the problem that traditional fixed convolution kernels cannot adapt to complex water vapor fields. Next, the extracted multi-scale spatial features (i.e., water vapor and meteorological features at different spatial scales, including local scale features, small-to-medium scale features, mesoscale features, directional features, and geometric adaptive features) are input into the ConvLSTM-ConvGRU parallel dual-branch module. The ConvLSTM branch (1 layer) is responsible for storing the long-term dependencies of precipitation features, while the ConvGRU branch is responsible for capturing short-term dynamic changes. The dual-branch features are fused through an adaptive fusion weight module to obtain the fused spatiotemporal features. A random deactivation layer is then added to alleviate overfitting, converting pure spatial features into spatiotemporal features, capturing the temporal evolution of the precipitation system, and outputting high-quality spatiotemporal features for use by the multi-task decoding branch. In this invention, after feature extraction, the ConvLSTM-ConvGRU parallel dual-branch module is input. The ConvLSTM branch is responsible for capturing long-term spatiotemporal dependencies, while the ConvGRU branch is responsible for mining short-term dynamic changes. The dual branches are adaptively fused through learnable weights, taking into account both the globality and dynamism of time series prediction. Furthermore, based on the fused spatiotemporal features, a multi-task decoding branch is constructed to realize precipitation regression, precipitation probability classification, precipitation type identification, and precipitation start time prediction, respectively. A cross-attention mechanism between tasks is used to enhance the interaction of multi-task features, achieving multi-task prediction outputs (precipitation, probability, type, and time). The cross-attention mechanism enhances feature interaction, and task collaboration and feature enhancement improve prediction accuracy and consistency, completing the initial construction of a multi-task precipitation prediction model that integrates adaptive convolutional kernels and ConvLSTM-ConvGRU. This invention constructs four parallel decoding branches: a main precipitation prediction task branch, an auxiliary precipitation probability prediction task branch, an auxiliary precipitation type identification task branch, and an auxiliary precipitation start time prediction task branch, corresponding to the four prediction tasks of precipitation, precipitation probability, precipitation type identification, and precipitation start time prediction, respectively. The cross-attention mechanism between tasks enhances features, simultaneously supporting precipitation, precipitation probability, precipitation type, and precipitation start time prediction, achieving feature interaction and enhancement between tasks, and solving the problem of limited prediction dimensions in traditional single-task models. Finally, the classified meteorological parameters, atmospheric water vapor, and long-term precipitation series data are input into the aforementioned multi-task precipitation prediction model that integrates adaptive convolutional kernels and ConvLSTM-ConvGRU. Iterative training is performed using an uncertainty-weighted multi-task loss function (including precipitation MSE loss, precipitation probability / type cross-entropy loss, and precipitation time binary cross-entropy loss), which automatically balances task weights without requiring manual parameter tuning. This adapts to differences in task difficulty and prevents any single task from dominating training. The multi-task collaborative learning effect improves convergence stability and generalization ability, thus completing the construction of the short-term nowcasting precipitation forecast model.

[0030] Furthermore, in step four, the training method for the multi-task neural network model that integrates the adaptive convolutional kernel system and the ConvLSTM-ConvGRU dual-branch model includes: (1) Divide the preprocessed data into training set and test set; (2) Determine the kernel size and pooling method for the convolutional network; (3) Initialize all weight parameters of the ConvLSTM and ConvGRU models; (4) Calculate the output values ​​of neurons in the network layer by forward propagation; (5) Backpropagate to calculate the error terms of neurons in the network layers; (6) Calculate the gradient of each parameter based on the training error; (7) Use the optimization algorithm to update the weight parameters and continue iterating until the total error converges, then stop training; (8) Output the prediction results of ConvLSTM and ConvGRU models, that is, output the multi-dimensional prediction results including precipitation amount, precipitation probability, precipitation type and precipitation start time in the next 0-3 hours.

[0031] Furthermore, in step four, the trained fusion adaptive convolutional kernel system and the ConvLSTM-ConvGRU dual-branch multi-task neural network model are used to predict short-term precipitation. The specific method is as follows: The GNSS atmospheric water vapor data and surface meteorological data preprocessed in step two are input into the model, and multi-dimensional prediction results are output simultaneously, including: grid-by-grid precipitation (accurate to 0.1 mm), unit mm, 5 levels of precipitation probability (the sum of the probabilities of each level is 1), 3 types of precipitation (predicted precipitation categories and confidence levels), and the probability of precipitation starting every 15 minutes in the next 0-3 hours (probability values ​​for a total of 12 time nodes), forming a complete short-term precipitation prediction report; The five levels of precipitation include no rain, light rain, moderate rain, heavy rain, and rainstorm; the three types of precipitation include convective precipitation, stratiform precipitation, and mixed precipitation.

[0032] Furthermore, in step five, the method for evaluating the accuracy of the prediction results obtained in step four includes: The forecast and observation results are categorized into two cases: with rain and without rain. If the observation data shows precipitation but the forecast data does not, the forecast is considered missing. If the observation data shows precipitation but the forecast data does not, the forecast is incorrect. If neither the observation data nor the forecast shows precipitation, then it is considered that there is no precipitation. A threshold is set to determine if precipitation is present. If the precipitation intensity is greater than the set threshold, precipitation is confirmed. If the intensity is less than the set threshold, there is no precipitation, as shown in Table 1. The TS score is typically used to evaluate the accuracy of short-term and upcoming precipitation forecasts, and its formula is as follows: Table 1. Evaluation of Precipitation Forecast ; ; ; ; ; in, To predict the correct number of times, To predict the number of errors, ts' represents the number of missing observations, and ts' represents the ts score of the numerical forecast or the optimized forecast. In addition, the probability of detection (POD), false alarm rate (FAr), and critical success index (CSI), commonly used in hydrometeorological research, are used to evaluate the short-term and critical forecasting capabilities of the forecasting model. The calculation formulas are as follows: ; ; ; Among them, the number of hits (H) is the number of times that rain was predicted and rain was actually observed; the number of misses (M) is the number of times that rain was predicted but rain was actually observed; and the number of false alarms (F) is the number of times that rain was predicted but rain was actually observed.

[0033] Example: The present invention will be described in detail here by taking the application of the present invention to a target area for short-term precipitation prediction as an example. The present invention will also have guiding significance for the application of the present invention to other target areas for short-term precipitation prediction.

[0034] In this embodiment, the target area is a lake station in a certain city, located in the center of a certain district of the city. It is surrounded by a high-density built-up area and a lake, and is significantly affected by the urban heat island effect and lake-land breeze circulation. It is a complex and typical area for short-term precipitation forecasting in the city. Location of the lake station in the city: 114.32°E, 30.57°N; elevation: 23m.

[0035] This embodiment uses existing technology to predict short-term precipitation in a target area. However, due to limitations such as the inability of fixed convolution kernels to adapt to the irregular distribution of urban land-water boundaries, the lack of a direction-sensitive mechanism leading to deviations in the prediction of lake-land wind and water vapor paths, the difficulty of a single ConvLSTM time series model to take into account both long-term and short-term dependencies, and the inability of a single task output to meet the multidimensional needs of urban disaster prevention and mitigation, the prediction accuracy in this area is limited.

[0036] This embodiment utilizes the advantages of this invention to predict short-term precipitation in a target area. It integrates core technologies such as an adaptive convolutional kernel system (anisotropic + deformable + multi-scale + orientation sensitive) to dynamically adjust the receptive field and sampling position, ConvLSTM-ConvGRU parallel dual-branch to balance long-term dependencies and short-term dynamic capture, inter-task cross-attention mechanism to enhance multi-task feature interaction, uncertainty-weighted loss to strengthen focus on heavy precipitation, and GNSS water vapor high spatiotemporal resolution data fusion. This achieves high-precision and high-accuracy multi-task short-term precipitation prediction for a target area. The test results of this embodiment using this invention for a target area are as follows: Figure 3 , Figure 4 As shown; exist Figure 3 In the diagram, the horizontal axis represents time, and the vertical axis represents precipitation. Figure 3 The red dashed line represents the prediction results of the fusion prediction model described in this invention, and the blue solid line represents the measured precipitation results under the same conditions. Figure 3 It can be seen that: (1) The present invention has extremely high numerical forecast accuracy: There is a very strong linear positive correlation between the predicted precipitation and the measured value of the model described in the present invention; It can be seen that the model described in the present invention can not only accurately determine "whether it rains", but also quantify "how much rain falls" very accurately, which greatly reduces the systematic bias (such as generally underestimating or overestimating) common in traditional models; (2) The present invention has excellent ability to capture heavy precipitation events: Figure 3 The predicted curve of the model described in this invention can closely follow the peak time and peak intensity of the measured curve, proving that the model described in this invention does not suffer from serious underreporting or peak underestimation when dealing with sudden, short-term heavy rainfall, thus overcoming the shortcomings of existing technologies such as "slow response and low estimation". exist Figure 4 In the diagram, the horizontal axis (X-axis) represents the measured precipitation; the vertical axis (Y-axis) represents the model-predicted precipitation. Figure 4The blue dots represent the "predicted value vs. measured value" paired data for each sample point (i.e., each prediction time). The denser the blue dots are distributed near the diagonal, the more accurate the prediction. If the blue dots are scattered or clustered below the diagonal, it indicates a large error or underestimation. The red dashed line (y=x line, also the diagonal): represents the perfect prediction line (1:1 line). If all the blue dots fall on this line, it means that the predicted value is exactly equal to the measured value; this is the benchmark reference line for evaluation, representing the "ideal state". exist Figure 4 In this study, the correlation coefficient (Pearson correlation coefficient) ranges from -1 to 1. The correlation coefficient of this invention, 0.8754, is very close to 1, indicating a strong positive correlation between the predicted precipitation and the actual observed precipitation. This demonstrates that the model described in this invention has high accuracy in predicting short-term precipitation. This overcomes the problem of weak positive correlation between predicted precipitation and actual observed precipitation in existing technologies (approximately between 0.6 and 0.7), resulting in low prediction accuracy.

[0037] The model described in this invention achieves an average TS score of 0.4805 in short-term precipitation prediction (TS=0.5419 for light rain, TS=0.4430 for moderate rain, and TS=0.4565 for heavy rain), which is 34.6% higher than the average TS (0.357) of existing technologies; the CSI for heavy precipitation is improved from 0.26 to 0.39 (+50.0%), and a new precipitation type identification function (accuracy of 76.8%) and multi-task consistent output (accuracy of 87.3%) are added. It is evident that this invention can effectively overcome the bottleneck of short-term precipitation prediction in complex urban underlying surface areas, providing reliable technical support for urban flooding early warning, traffic scheduling, and major event security, and has significant novelty and inventiveness.

[0038] All other unspecified parts belong to the prior art.

Claims

1. A multi-task precipitation prediction model integrating adaptive convolutional kernels and ConvLSTM-ConvGRU, characterized in that: It includes a sequence input layer, a multi-scale adaptive convolutional feature extraction layer, a ConvLSTM-ConvGRU parallel dual-branch layer, a multi-task decoding branch and cross-attention layer, a multi-task loss function and model training layer, and a model output and evaluation layer; The sequence input layer is used to receive GNSS water vapor data and meteorological data at multiple time steps; The multi-scale adaptive convolutional feature extraction layer is used to extract features from each data input to the sequence input layer, and to achieve concatenation and weighted fusion of the data features. The ConvLSTM-ConvGRU parallel dual-branch layer is used to implement weight allocation between the current input and historical states, historical information filtering, output control, and long-term dependency storage. Multi-task decoding branches and cross-attention layers are used to construct multi-task decoding branches based on the output of the shared feature extraction layer, enabling multi-dimensional prediction of precipitation-related data. The multi-task loss function and model training layer are used to introduce an uncertainty-weighted dynamic loss adjustment strategy to achieve the training objective of balancing multiple tasks. The model output and evaluation layer is used to output multi-dimensional predictions related to precipitation and to evaluate the accuracy of the prediction results.

2. The multi-task precipitation prediction model fusing adaptive convolutional kernels and ConvLSTM-ConvGRU as described in claim 1, characterized in that: Multi-scale adaptive convolutional feature extraction layer, including: Layers 1-2: 3×3 convolutional layers to extract local features, with 64 convolutional kernels. ReLU activation layers and 2×2 max pooling layers are connected after each of the two convolutional layers. Layers 3-4: 5×5 convolutional layers, extracting small-to-medium scale features, with 128 convolutional kernels, also connected to ReLU activation layers and 2×2 max pooling layers; Layers 5-6: 7×7 convolutional layers, extracting mid-scale features, with 256 convolutional kernels, connecting ReLU activation layers and 2×2 max pooling layers.

3. The multi-task precipitation prediction model fusing adaptive convolutional kernels and ConvLSTM-ConvGRU as described in claim 1 or 2, characterized in that: The ConvLSTM-ConvGRU parallel dual-branch layer includes: ConvLSTM branch: Set up 1 LSTM layer, whose structure includes input gate, forget gate, output gate and memory unit, respectively to implement the weight allocation of current input and historical state, historical information filtering, output control and long-term dependency storage. The branch output dimensions are: batch, time step, 512, H / 8, W / 8. ConvGRU branch: contains update gate, reset gate and candidate hidden state. The update gate determines how much historical information is retained, the reset gate controls the degree of influence of historical information on the current state, and the candidate hidden state combines the current input and the reset historical state. The branch output dimension is consistent with ConvLSTM. Adaptive fusion weight module: The input is the output features of ConvLSTM and ConvGRU. First, the global average pooling features are calculated, and then the fusion weights α and β are learned through two fully connected layers, where α+β=1. The final fusion output is α×ConvLSTM output+β×ConvGRU output. The fusion feature dimensions are: batch, time step, 512, H / 8, W / 8.

4. The multi-task precipitation prediction model fusing adaptive convolutional kernels and ConvLSTM-ConvGRU as described in claim 3, characterized in that: Multi-task decoding branches and cross-attention layers include task-specific branches and inter-task cross-attention mechanisms; the inter-task cross-attention mechanism uses the intermediate layer features of each task branch as queries, keys, and values; the attention weight of task i on task j = , where d represents the feature dimension; Q i K represents the intermediate layer features of task i; j T The intermediate layer features represent the features of task j; the enhanced features of task i equal the original features. ; Among them, V j Represent the characteristics of task j; Task-specific branches include main tasks and side tasks; The main task is to predict precipitation, including: Layer 1: 1×1 convolution from 512 channels to 256 channels to reduce dimensionality; Layer 2: 3×3 convolution upsampling to restore spatial resolution; Layer 3: Continue upsampling to the original resolution; Layer 4: 1×1 convolution to output single-channel precipitation data. The auxiliary tasks include auxiliary task 1, auxiliary task 2 and auxiliary task 3; auxiliary task 1 is the precipitation probability prediction branch, auxiliary task 2 is the precipitation type identification branch, and auxiliary task 3 is the precipitation start time prediction branch. Among them, auxiliary task 1: precipitation probability prediction includes: Layer 1: 1×1 convolution from 512 channels to 128 channels; Layer 2: Global average pooling to obtain spatial feature vectors; Layer 3: Fully connected layer, outputting five probabilities: no rain, light rain, moderate rain, heavy rain, and torrential rain, with the activation function being Softmax; Auxiliary Task 2: Precipitation Type Identification, including: Layer 1: 1×1 convolution from 512 channels to 128 channels; Layer 2: Extracts convection features and stratus features; Layer 3: Fully connected layer, outputs three probabilities: convection, stratus, and mixed type, with Softmax as the activation function; Auxiliary Task 3: Prediction of precipitation start time, including: Layer 1: Attention mechanism in the time dimension, focusing on feature changes before precipitation begins; Layer 2: LSTM layer to process time series features; Layer 3: Fully connected layer, outputting the probability of precipitation starting every 15 minutes in the next 0-3 hours.

5. The multi-task precipitation prediction model fusing adaptive convolutional kernels and ConvLSTM-ConvGRU as described in claim 4, characterized in that: The multi-task loss function and model training layer include the multi-task loss function and model training; Multi-task loss functions include: Main task losses: ,in Indicates sample weights, N represents the sample size; y p y represents the model's predicted value; t Represents the actual value; Losses in Auxiliary Task 1: ; Losses in Auxiliary Mission 2: ; Losses in Auxiliary Mission 3: ; Total loss function: Introduces an uncertainty parameter for the science system for each task. The total loss formula is: 。 6. A method for predicting short-term precipitation using a multi-task precipitation prediction model that fuses adaptive convolutional kernels and ConvLSTM-ConvGRU as described in any one of claims 1-5, characterized in that: Includes the following steps, Step 1: First, acquire GNSS atmospheric water vapor data (PWV) and surface meteorological data; Step 2: Perform data preprocessing on GNSS atmospheric water vapor data (PWV) and surface meteorological data to obtain preprocessed data. Remove observation errors, outliers, and invalid data, unify data format and dimensions, and retain the core correlation characteristics between water vapor and surface meteorological parameters. Step 3: Establish a multi-task neural network model that integrates an adaptive convolutional kernel system and a ConvLSTM-ConvGRU dual-branch system. The adaptive convolutional kernel system includes adaptive anisotropic convolutional kernels, deformable convolutional kernels, multi-scale convolutional kernels, and orientation-sensitive convolutional kernels. The ConvLSTM branch and the ConvGRU branch are adaptively fused through learnable weights and combined with a cross-attention mechanism between tasks to achieve multi-task feature enhancement. Step 4: Train the multi-task neural network model with the fusion of adaptive convolutional kernel system and ConvLSTM-ConvGRU dual branch using the preprocessed data, and use the trained multi-task neural network model with the fusion of adaptive convolutional kernel system and ConvLSTM-ConvGRU dual branch to predict the short-term precipitation in the target area to verify the effectiveness of the model. Step 5: Evaluate the accuracy of the prediction results of the model described in Step 4 based on the recent actual precipitation results in the target area; Specifically, the multi-dimensional forecast results are compared with the recent actual precipitation results of the target area. Meteorological professional indicators such as TS score, POD, and CSI are used to comprehensively evaluate the accuracy of the forecast results from multiple dimensions, including forecast accuracy, success index, and error control.

7. The method for predicting short-term precipitation using a multi-task precipitation prediction model that integrates adaptive convolutional kernels and ConvLSTM-ConvGRU according to claim 6, characterized in that: In step one, the ground meteorological data includes 2m air temperature, 2m relative humidity, 10m wind speed, ground air pressure, and hourly precipitation.

8. The method for predicting short-term precipitation using a multi-task precipitation prediction model that integrates adaptive convolutional kernels and ConvLSTM-ConvGRU according to claim 7, characterized in that: In step two, preprocessing includes handling gross errors in the data, data normalization, data pruning, and data denoising.

9. The method for predicting short-term precipitation using a multi-task precipitation prediction model that integrates adaptive convolutional kernels and ConvLSTM-ConvGRU according to claim 8, characterized in that: In step three, the method for constructing the multi-task precipitation prediction model that integrates adaptive convolutional kernels and ConvLSTM-ConvGRU includes the following steps: First, by rearranging the dimensions of the sequence input layer, the preprocessed GNSS atmospheric water vapor data and surface meteorological data are adjusted into a standard data structure; Subsequently, features are extracted sequentially through a multi-scale adaptive convolution module: first, water vapor and meteorological features at different spatial scales are captured by 3×3, 5×5, and 7×7 multi-scale convolution kernel groups; then, the weights and sampling positions are dynamically adjusted by adaptive anisotropic convolution kernels and deformable convolution kernels to enhance the extraction capabilities of water vapor distribution directionality and station sparsity features. Next, the extracted multi-scale spatial features are input into the ConvLSTM-ConvGRU parallel dual-branch module, where the ConvLSTM branch is used to store the long-term dependence of precipitation features and the ConvGRU branch is used to capture short-term dynamic changes. The dual-branch feature fusion is completed through the adaptive fusion weight module, and a random deactivation layer is connected to alleviate overfitting. Furthermore, a multi-task decoding branch is constructed based on the fused spatiotemporal features to realize precipitation regression, precipitation probability classification, precipitation type identification, and precipitation start time prediction, respectively. The multi-task feature interaction is enhanced through a cross-attention mechanism between tasks. Finally, GNSS atmospheric water vapor data and surface meteorological data were input into the model, and iterative training was performed using an uncertainty-weighted multi-task loss function to complete the construction of the short-term precipitation forecast model.

10. The method for predicting short-term precipitation using a multi-task precipitation prediction model that integrates adaptive convolutional kernels and ConvLSTM-ConvGRU according to claim 9, characterized in that: In step four, the trained fusion adaptive convolutional kernel system and the ConvLSTM-ConvGRU dual-branch multi-task neural network model are used to predict short-term precipitation. The specific method is as follows: The preprocessed GNSS atmospheric water vapor data and surface meteorological data are input into the model, and multi-dimensional prediction results are output simultaneously, including: grid-by-grid precipitation (in mm), probability of precipitation at level 5, three types of precipitation, and probability of precipitation starting every 15 minutes in the next 0-3 hours, forming a complete short-term precipitation prediction report. The five levels of precipitation include no rain, light rain, moderate rain, heavy rain, and rainstorm; the three types of precipitation include convective precipitation, stratiform precipitation, and mixed precipitation.

Images