Deep learning based extreme precipitation fusion simulation parameter adaptive optimization method

Abstract

The application discloses an extreme precipitation fusion simulation parameter adaptive optimization method based on deep learning, comprising collecting multi-source heterogeneous data of extreme precipitation in a preset area, and preprocessing the multi-source heterogeneous data; adopting an attention mechanism and a convolution long short-term memory network to perform extreme precipitation fusion on the multi-source heterogeneous data to obtain fusion data, adopting meta-learning to perform shared prior learning on the fusion data to obtain a general parameter initialization strategy; adopting a spatiotemporal multi-scale self-supervision to pretrain the fusion data to obtain extended data, and constructing an extreme precipitation simulation model according to the extended data and the parameter initialization strategy; adopting a parameter sensitivity atlas and an adversarial network to perform regional customization adjustment on the extreme precipitation fusion simulation model to obtain regionalized parameters, adopting an adversarial parameter game to optimize the extreme precipitation fusion simulation model to obtain high-resolution simulation parameters, and outputting the regionalized parameters and the high-resolution simulation parameters as an optimization result.

Description

Technical Field

[0001] This invention relates to the field of adaptive optimization of parameters in extreme precipitation fusion simulations, and more particularly to an adaptive optimization method for extreme precipitation fusion simulation parameters based on deep learning. Background Technology

[0002] As a high-impact weather event, the accurate simulation of extreme precipitation is crucial for disaster prevention and mitigation, water resource allocation, and ecological protection. With the development of observation technology, multi-source heterogeneous data from radar, satellites, and ground stations provide abundant data sources for precipitation simulation. However, different data sources differ in spatiotemporal resolution, accuracy, and coverage, making it difficult for traditional fusion methods to effectively integrate multi-dimensional information, which can easily lead to data redundancy or loss of key features.

[0003] Existing extreme precipitation simulation models mostly rely on manually set parameters, lacking the ability to adapt to regional topography and climate characteristics. General parameters have large errors when applied to different regions. At the same time, extreme precipitation samples are scarce and their spatiotemporal evolution is complex, making it difficult for traditional models to capture their nonlinear dynamic characteristics, resulting in insufficient simulation resolution and accuracy.

[0004] Furthermore, existing parameter optimization methods often neglect the synergy between physical constraints and data-driven approaches, making it difficult to balance the statistical consistency and physical rationality of simulation results. Moreover, parameter adjustments lack dynamism and cannot respond in real-time to the spatiotemporal variations of precipitation processes. These problems make existing methods insufficient to meet the practical needs of refined simulation of extreme precipitation, necessitating a more efficient simulation method that combines data fusion capabilities, regional adaptability, and parameter self-optimization capabilities. Summary of the Invention

[0005] The purpose of this invention is to provide an adaptive optimization method for extreme precipitation fusion simulation parameters based on deep learning.

[0006] To achieve the above objectives, the present invention is implemented according to the following technical solution: This invention includes the following steps: Collect multi-source heterogeneous data on extreme precipitation in a preset area, and preprocess the multi-source heterogeneous data; the multi-source heterogeneous data includes radar data, satellite data, and ground observation data; An attention mechanism and a convolutional long short-term memory network are used to fuse the multi-source heterogeneous data to obtain fused data, and meta-learning is used to perform shared prior learning on the fused data to obtain a general parameter initialization strategy. The fused data is pre-trained using spatiotemporal multi-scale self-supervision to obtain extended data, and an extreme precipitation simulation model is constructed based on the extended data and the parameter initialization strategy. The extreme precipitation fusion simulation model is regionalized by using parameter sensitivity maps and adversarial networks to obtain regionalized parameters. The extreme precipitation fusion simulation model is then optimized by adversarial parameter game to obtain high-resolution simulation parameters. The regionalized parameters and the high-resolution simulation parameters are then output as optimization results.

[0007] Furthermore, a method for fusing extreme precipitation data from the multi-source heterogeneous data using an attention mechanism and a convolutional long short-term memory network includes: Radar reflectivity factor, satellite-retrieved precipitation, and ground station observations are obtained from multi-source heterogeneous data. The radar reflectivity factor, satellite-retrieved precipitation, and ground station observations are input into a convolutional long short-term memory network. The vertical reflectivity evolution features are extracted through a 3D convolutional layer to obtain radar features. A 2D convolutional layer is used to capture cloud image texture features to obtain satellite features. The station observations are interpolated into grid data through a fully connected layer, and then compressed to 32 channels through a 1×1 convolution to obtain station features. Calculate the spatial gradient of the feature maps from each data source, generate a saliency weight map to obtain the spatial attention weights, and express the following expression:

[0008] in For spatial location Attention weights at the location The spatial coordinates of the feature map For remote sensing features in Spatial gradient magnitude at that location For spectral characteristics in Spatial gradient magnitude at that location For geographical features in Spatial gradient magnitude at that location For feature map height, The width of the feature map; Global average pooling is performed on the data source features. The channel importance weights are learned through an MLP to obtain a three-dimensional weight matrix after broadcasting the channel attention weights in the spatial dimension. The expression is:

[0009] in This is the feature map of the c-th channel. For global average pooling, It is a multilayer perceptron. It is the sigmoid activation function. This is the three-dimensional weight matrix after broadcasting the attention weights of the c-th channel in the spatial dimension. Extreme precipitation data are fused based on the saliency weighting plot and channel importance weighting to obtain fused data, expressed as:

[0010] in For remote sensing features, For spectral characteristics, Geographical features To integrate data, For element-wise multiplication, This is a three-dimensional weight matrix after the channel attention weights are broadcast in the spatial dimension. For spatial attention weights.

[0011] Furthermore, the method for obtaining the general parameter initialization strategy includes: Obtain auxiliary variables for the preset regions, and divide the data of each source region into a meta-training task set and a meta-testing task set in a 4:1 ratio. Each task includes a support set and a query set. The auxiliary variables are terrain height and vegetation index. The support set consists of a small number of samples. The query set consists of validation samples. Input the meta-training task set and auxiliary variables into the meta-learning model. For each meta-training task, use the support set to update the model parameters and obtain the optimized initial parameters. The expression is:

[0012] in For meta-parameters, For the adapted parameters of task a, Step size, The gradient of the meta-parameters, Neural networks with meta-parameters For the support set of task a, Let be the loss function for task a; Use the updated parameters The meta-loss is calculated on the query set, and the initial parameters are updated via backpropagation. The expression is:

[0013] in Let be the query set for task a. For the number of tasks, The meta-learning rate; By visualizing attention weights, key features shared across regions are identified, and these key features are encoded as constraints for the initial parameters of the model. Use a small number of samples in the target region to optimize the initial parameters. Perform multiple rounds of fine-tuning to quickly converge to the regional optimum, and obtain the adaptation rule based on the optimum solution; then apply the optimized initial parameters. The adaptation rule output is a general parameter initialization strategy.

[0014] Furthermore, a method for pre-training the fused data using spatiotemporal multi-scale self-supervision to obtain extended data includes: The fused data is divided into temporal and spatial scales, and random transformations are performed on the fused data to generate enhanced samples; spatiotemporal prediction tasks, feature reconstruction tasks, and cross-scale consistency tasks are designed to learn spatiotemporal features from multiple dimensions. A self-supervised pre-trained model architecture is constructed using a deep learning model architecture, and multi-scale feature extraction and fusion are performed using the self-supervised pre-trained model architecture. The deep learning model architecture includes an encoder and a decoder, and the encoder includes a spatiotemporal convolutional model and a cross-scale attention module. Self-supervised pre-training is divided into single-task pre-training and multi-task joint training. For single-task pre-training: train the three main tasks separately and optimize the corresponding loss functions, as expressed in the following expression:

[0015]

[0016]

[0017] in For the predicted four future time steps, Four time steps for the real future. For mean square error loss, For the reconstructed mask area, For the original complete data, The mean absolute error, For similar structural indices, Reconstructed values ​​at 1km resolution. For true values ​​at 1km resolution, For the total variation loss in two-dimensional space, The total variation weight; For multi-task joint training: the losses from the three tasks are fused, and general spatiotemporal features are learned through parameter sharing, expressed as follows:

[0018] in The total pre-training loss, To predict losses, To rebuild the losses, For the excess score loss, To predict the weighting coefficients, To reconstruct the weighting coefficients, This refers to the super-resolution weighting coefficient; A recurrent learning rate strategy is adopted, using the AdamW optimizer. After pre-training, the encoder weights are frozen, and only the decoder output head is fine-tuned. Two types of extended data are generated using the pre-trained model: extreme scene synthesis and sample augmentation.

[0019] Furthermore, the method for constructing an extreme precipitation simulation model based on the extended data and the parameter initialization strategy includes: A convolutional long short-term memory network architecture with spatiotemporal attention is adopted as the main network of the extreme precipitation simulation model. The convolutional long short-term memory network architecture includes an encoder, a decoder, and an output layer. The encoder consists of 4 layers of ConvLSTM, which captures the time series dependence of precipitation and embeds a spatial attention module to enhance the features of the center of heavy precipitation. The decoder consists of 4 layers of upsampled convolutions, which combine skip connections to fuse the multi-scale features output by the encoder and introduce channel attention to dynamically adjust the feature weights of different data sources. The output layer consists of 1×1 convolution + sigmoid activation function, which outputs the precipitation intensity prediction of 0-100 mm / h. Atmospheric dynamics equations and topographic modulation factors are used as physical constraint modules in the extreme precipitation simulation model. Atmospheric dynamics equations are introduced as constraints: results conforming to atmospheric motion laws are simulated through an additional loss term, expressed as follows: ,in For additional losses, For the predicted water vapor flux, To reanalyze the water vapor flux of precipitation data, For divergence operators, For Euclidean norm; Terrain modulation factor: Encodes the digital elevation model into a terrain weight map through 1×1 convolution; The initial parameters optimized by the meta-learning output are loaded, and self-supervised pre-training is performed by inputting extended data. The pre-training is completed by using precipitation sequences. This includes the initial weights of convolutional layers and the attention mechanism. Given the core loss of an extreme precipitation simulation model, the expression is:

[0020] in To determine the core loss of the extreme precipitation simulation model, This represents the predicted precipitation within the effective area.

[0021] Furthermore, a method for obtaining regionalized parameters by using parameter sensitivity maps and adversarial networks to regionally customize the extreme precipitation fusion simulation model includes: Historical fused precipitation data and candidate driving factors were acquired, and an improved Peter Clark algorithm was used to construct a causal graph of a graph neural network through variable preprocessing, causal structure learning, and parameter-sensitive quantification. Based on the causal graph, the sensitivity coefficients of each driving factor are used as node weights. Using geospatial data as a carrier, the sensitivity coefficients of each driving factor in the sub-region are visualized to generate a parameter sensitivity spatial map. The value of each geographic grid in the parameter sensitivity spatial map is the adjustment weight of the corresponding driving factor on the model parameters, realizing the spatial expression of parameter sensitivity. A graph attention network employing a causal attention learning strategy is used to dynamically weight the model parameters based on driving factors. The input layer, causal attention layer, backdoor adjustment mechanism, and output are used as the model framework of the graph neural network. The backdoor adjustment mechanism borrows the concept of confusion factor control from causal inference, mitigating the impact of non-causal features by intervening in the loss function. The expression is:

[0022] in For the intervention loss function, To predict precipitation, To reanalyze precipitation data, As a shortcut feature, To force the shortcut feature to the value d, For intervention operations; Obtain the parameter adjustment weight matrix corresponding to each driving factor. Adjust the weight matrix based on parameters Regional adaptation is performed on the general parameters of meta-learning; The preset area is divided into sub-regions according to terrain complexity. Each sub-region is executed independently: input the driving factor data of the sub-region, the graph neural network dynamically calculates the parameters of the region, adjusts the weights, updates the model parameters, and verifies them with regional observation data. Generative Adversarial Networks (GANs) are introduced, consisting of a generator and a discriminator. The generator produces a simulated precipitation field based on regionalized parameters, while the discriminator distinguishes between simulated precipitation and real observations, with a focus on extreme values. The game objective is to minimize the generator loss and maximize the discriminator loss, expressed as:

[0023] in The probability of the discriminator classifying the input data. For the generator based on region parameters The output simulated precipitation field, The expected value is represented by the actual observed precipitation field data. After multiple rounds of adversarial training, the final regionalized high-resolution simulation parameters are output.

[0024] Furthermore, a method for obtaining high-resolution simulation parameters by optimizing the extreme precipitation fusion simulation model using adversarial parameter game theory includes: Construct two parametric games, one consisting of data-driven game and the other of physical constraint game. The main components of the parametric game are the generator, the discriminator, and the physical referee. The objective function is constructed by minimizing the Wasserstein distance between generated precipitation and real events and maximizing the discriminant's discriminant loss. The expression is as follows:

[0025] in For discriminator, For parameters Precipitation generated below, For actual precipitation, To calculate the Wasserstein distance between the generated precipitation and the actual precipitation, This is the distance weighting coefficient. For adversarial loss function; parameter The update rules are as follows: the discriminator is trained 5 times per round, and the parameters are updated using gradient ascent; the generator updates the discriminator once, and the parameters are optimized using gradient descent. ; We introduce three core physical regularization terms: mass conservation constraint, energy conservation constraint, and terrain sensitivity constraint, and construct a physical loss function. Based on the physical loss function, we construct the overall objective function. parameter The optimization strategy is as follows: The problem is decomposed using the alternating direction multiplier method: first, the physical constraint optimization parameters are fixed. Then fix the parameters Adjust the physical parameters and iterate until convergence; Optimal parameters are selected from the post-game parameter space using the Bayesian information criterion. ,satisfy ,in For the number of parameters, For sample size, For parameters The likelihood function, For parameters Bayesian information criterion; parameters Output high-resolution simulation parameters, parameters It includes the convolutional kernel weights of the convolutional long short-term memory network, the spatial weight matrix of the attention mechanism, and the threshold parameters of the physical constraints.

[0026] The beneficial effects of this invention are: This invention presents an adaptive optimization method for extreme precipitation fusion simulation parameters based on deep learning. Compared with existing technologies, this invention has the following technical advantages: This invention efficiently fuses multi-source heterogeneous data through an attention mechanism and ConvLSTM, and combines meta-learning to obtain a general parameter initialization strategy, improving the initial adaptability of the model. Spatiotemporal multi-scale self-supervised pre-training expands extreme samples, and the physical constraint module ensures that the simulation results conform to the laws of atmospheric motion. Parameter regional customization is achieved through parameter sensitivity mapping and adversarial networks, and adversarial parameter game further optimizes high-resolution parameters, taking into account both statistical consistency and physical rationality. The entire process achieves adaptive parameter adjustment without manual setting, significantly improving the resolution, accuracy, and regional adaptability of extreme precipitation simulation, and providing precise data support for disaster prevention and mitigation. Attached Figure Description

[0027] Figure 1 This is a flowchart illustrating the steps of the adaptive optimization method for extreme precipitation fusion simulation parameters based on deep learning, as described in this invention. Detailed Implementation

[0028] The present invention will be further described below through specific embodiments. The illustrative embodiments and descriptions herein are used to explain the present invention, but are not intended to limit the present invention.

[0029] The adaptive optimization method for extreme precipitation fusion simulation parameters based on deep learning, as described in this invention, includes the following steps: like Figure 1 As shown, this embodiment includes the following steps: Collect multi-source heterogeneous data on extreme precipitation in a preset area, and preprocess the multi-source heterogeneous data; the multi-source heterogeneous data includes radar data, satellite data, and ground observation data; In the actual assessment, a certain watershed was selected as the study area. This watershed is a region with a high frequency of extreme heavy precipitation. The terrain includes plains, hills and mountains, and the climate and underlying surface characteristics are complex. The study time scale is the flood season (June to August) from 2018 to 2022. The temporal resolution is uniformly 15 minutes, and the spatial resolution is 1km×1km. Radar data: Reflectivity factor data from 7 S-band weather radars in the Yangtze-Huaihe River Basin, with a radial resolution of 1km and an elevation angle of 0.5°~10°; Satellite data: Precipitation data retrieved from cloud images of the FY-4A geostationary meteorological satellite, with a spatial resolution of 4km and a temporal resolution of 15 minutes; Ground observation data: 15-minute rainfall observation data from 862 national-level ground meteorological stations in the study area, as well as supplementary observation data from regional automatic stations; Auxiliary variables: Topographic height, slope, and aspect extracted from a 30m resolution digital elevation model, NDVI vegetation index from the MODIS satellite at a 1km resolution, and 500hPa geopotential height, 850hPa temperature advection, and whole-layer water vapor flux from reanalysis data. Preprocessing includes: unifying radar, satellite, and ground observation data to a spatial grid of 1km×1km and a temporal resolution of 15 minutes; removing outliers from ground station observations based on the 3σ criterion; performing Z-Score standardization on radar and satellite data; and performing minimum normalization (maximum - minimum) on ground station data. An attention mechanism and a convolutional long short-term memory network are used to fuse the multi-source heterogeneous data to obtain fused data, and meta-learning is used to perform shared prior learning on the fused data to obtain a general parameter initialization strategy. The fused data is pre-trained using spatiotemporal multi-scale self-supervision to obtain extended data, and an extreme precipitation simulation model is constructed based on the extended data and the parameter initialization strategy. The extreme precipitation fusion simulation model is regionalized by using parameter sensitivity maps and adversarial networks to obtain regionalized parameters. The extreme precipitation fusion simulation model is then optimized by adversarial parameter game to obtain high-resolution simulation parameters. The regionalized parameters and the high-resolution simulation parameters are then output as optimization results.

[0030] In this embodiment, the method for fusing extreme precipitation data from multiple heterogeneous sources using an attention mechanism and a convolutional long short-term memory network includes: Radar reflectivity factor, satellite-retrieved precipitation, and ground station observations are obtained from multi-source heterogeneous data. The radar reflectivity factor, satellite-retrieved precipitation, and ground station observations are input into a convolutional long short-term memory network. The vertical reflectivity evolution features are extracted through a 3D convolutional layer to obtain radar features. A 2D convolutional layer is used to capture cloud image texture features to obtain satellite features. The station observations are interpolated into grid data through a fully connected layer, and then compressed to 32 channels through a 1×1 convolution to obtain station features. Calculate the spatial gradient of the feature maps from each data source, generate a saliency weight map to obtain the spatial attention weights, and express the following expression:

[0031] in For spatial location Attention weights at the location The spatial coordinates of the feature map For remote sensing features in Spatial gradient magnitude at that location For spectral characteristics in Spatial gradient magnitude at that location For geographical features in Spatial gradient magnitude at that location For feature map height, The width of the feature map; Global average pooling is performed on the data source features. The channel importance weights are learned through an MLP to obtain a three-dimensional weight matrix after broadcasting the channel attention weights in the spatial dimension. The expression is:

[0032] in This is the feature map of the c-th channel. For global average pooling, It is a multilayer perceptron. It is the sigmoid activation function. This is the three-dimensional weight matrix after broadcasting the attention weights of the c-th channel in the spatial dimension. Extreme precipitation data are fused based on the saliency weighting plot and channel importance weighting to obtain fused data, expressed as:

[0033] in For remote sensing features, For spectral characteristics, Geographical features To integrate data, For element-wise multiplication, This is the three-dimensional weight matrix after broadcasting the channel attention weights in the spatial dimension. Spatial attention weights; The Convolutional Long Short-Term Memory (CLSTM) network consists of a network structure, a gating mechanism, and an output layer. The network structure uses a stack of three ConvLSTM layers. The first layer has 64 channels, a 3×3 convolutional kernel, a stride of 1, and padding of 1. The second layer has 128 channels, a 3×3 convolutional kernel, a stride of 1, and padding of 1. The third layer has 64 channels, a 3×3 convolutional kernel, a stride of 1, and padding of 1. Gating mechanisms include input gates, forget gates, output gates, cell states, and hidden states, expressed as:

[0034]

[0035]

[0036]

[0037]

[0038] in For the current input, This is the hidden state from the previous moment. This represents the cell state at the previous moment. Currently in a hidden state. The current cell state, For input gate, For the Gate of Oblivion For output gate, For matrix multiplication, It is the hyperbolic tangent function. The weights are inputs to the input gate. The weights for the hidden state to the input gate, For input gate bias, The weights input to the forget gate, To hide the state and the weight of the forget gate, Forget Gate Offset The weights of the input and output gates, The weights of the hidden state and the output gate, For output gate bias, The weights are input to the candidate states. The weights for transitioning from the hidden state to the candidate state. For candidate states, a bias is applied. The output layer maps the hidden state to a single-channel precipitation intensity through a 1×1 convolution, with ReLU as the activation function and a truncation threshold of 0-100 mm / h. In the actual evaluation, the input multi-source heterogeneous data were preprocessed as follows: unified to a 1km grid and a 15-minute time resolution; station observation error values ​​were removed based on the 3σ criterion; Z-Score normalization was performed on R and S, and maximum-minimum normalization was used for G. The 3D convolutional layer has a 3×3×3 kernel and 64 channels; the spatial gradient is calculated using the Sobel operator to calculate the spatial gradient magnitude of the feature maps of each data source. First, the spatial gradients of the radar, satellite, and site feature maps are normalized to [0,1] by the maximum-minimum method, and then they are added together to obtain the result. Radar features: Radar reflectivity factors are input into a 3D convolutional layer with 3×3×3 kernels and 64 channels to extract vertical reflectivity evolution features; Satellite features: Texture features and spatial distribution features of precipitation in FY-4A cloud images are captured through a 2D convolutional layer (3×3 kernels and 64 channels); Station features: Ground station observations are interpolated into grid data through a fully connected layer, and then compressed to 32 channels through a 1×1 convolution to obtain the grid features of station precipitation; Spatial attention weights: The Sobel operator is used to calculate the spatial gradient magnitude of the feature maps of each data source. The gradient values ​​are first normalized to [0,1], and then calculated using the spatial attention weight formula in the document. The feature map height is 256 and the width is 256. Channel attention weights: Global average pooling is performed on the data source features, and channel importance weights are learned through a 2-layer MLP (64 hidden neurons). After Sigmoid activation, a three-dimensional weight matrix is ​​generated and broadcast in the spatial dimension. The spatial attention weights and channel attention weights are multiplied element-wise with the radar, satellite, and station features and then summed to obtain the fused extreme precipitation feature data. The fused data dimension is 256×256×64.

[0039] In this embodiment, the method for obtaining the general parameter initialization strategy includes: Obtain auxiliary variables for the preset region, and divide the data of each source region into a meta-training task set and a meta-test task set in a 4:1 ratio. In each task, the ratio of support set to query set is 1:4. The support set consists of a small number of labeled samples, and the query set consists of validation samples. The auxiliary variables are terrain height and vegetation index. The support set consists of a small number of samples, and the query set consists of validation samples. Input the meta-training task set and auxiliary variables into the meta-learning model. For each meta-training task, use the support set to update the model parameters and obtain the optimized initial parameters. The expression is:

[0040] in For meta-parameters, For the adapted parameters of task a, Step size, The gradient of the meta-parameters, Neural networks with meta-parameters For the support set of task a, Let be the loss function for task a; Use the updated parameters The meta-loss is calculated on the query set, and the initial parameters are updated via backpropagation. The expression is:

[0041] in Let be the query set for task a. For the number of tasks, The meta-learning rate; By visualizing attention weights, key features shared across regions are identified, and these key features are encoded as constraints for the initial parameters of the model. Use a small number of samples in the target region to optimize the initial parameters. Perform multiple rounds of fine-tuning to quickly converge to the regional optimum, and obtain the adaptation rule based on the optimum solution; then apply the optimized initial parameters. The adaptation rule output is a general parameter initialization strategy; In actual assessment, The value is 0.001, and its range is (0.0001, 0.1). The value is 0.0001, and its range is (1e -5 ,0.001); A watershed is divided into 10 source regions, with 8 meta-training task sets and 2 meta-testing task sets. The key features shared across regions are topographic height and water vapor flux, which are encoded as constraints for the initial parameters of the model. In the study area, 50 extreme precipitation samples are used to fine-tune the optimized initial parameters in 10 rounds, which quickly converge to the regional optimum. Finally, the optimized initial parameters plus the regional adaptation rules are output as a general parameter initialization strategy.

[0042] In this embodiment, the method for pre-training the fused data using spatiotemporal multi-scale self-supervision to obtain extended data includes: The fused data is divided into temporal and spatial scales, and random transformations are performed on the fused data to generate enhanced samples. The temporal scales are fine-scale, medium-scale, and coarse-scale; the spatial scales are high-resolution, medium-resolution, and low-resolution; the random transformations include temporal transformations and spatial transformations; the temporal transformations are time series reversal and random cropping; the spatial transformations are random rotation and local occlusion. The design includes a spatiotemporal prediction task, a feature reconstruction task, and a cross-scale consistency task, which learn spatiotemporal features from multiple dimensions. The spatiotemporal prediction task is for predicting future frames; the feature reconstruction task is for reconstructing occluded regions; and the cross-scale consistency task is for scale transformation prediction. A self-supervised pre-trained model architecture is constructed using a deep learning model architecture for multi-scale feature extraction and fusion. The deep learning model architecture includes an encoder and a decoder. The encoder consists of a spatiotemporal convolutional model and a cross-scale attention module. The spatiotemporal convolutional module is a 3-layer ConvLSTM that captures fine-scale spatiotemporal dependencies. The cross-scale attention module performs self-attention on medium / low-resolution data to extract large-scale circulation features. The decoder includes an upsampling module and a multi-task output module. The upsampling module restores low-resolution features to high resolution through transposed convolution. The multi-task output head corresponds to three self-supervised tasks. Self-supervised pre-training is divided into single-task pre-training and multi-task joint training. For single-task pre-training: train the three main tasks separately and optimize the corresponding loss functions, as expressed in the following expression:

[0043]

[0044]

[0045] in For the predicted four future time steps, Four time steps for the real future. For mean square error loss, For the reconstructed mask area, For the original complete data, The mean absolute error, For similar structural indices, Reconstructed values ​​at 1km resolution. For true values ​​at 1km resolution, For the total variation loss in two-dimensional space, The total variation weight; For multi-task joint training: the losses from the three tasks are fused, and general spatiotemporal features are learned through parameter sharing, expressed as follows:

[0046] in The total pre-training loss, To predict losses, To rebuild the losses, For the excess score loss, To predict the weighting coefficients, To reconstruct the weighting coefficients, This refers to the super-resolution weighting coefficient; A recurrent learning rate strategy is adopted, using the AdamW optimizer. After pre-training, the encoder weights are frozen, and only the decoder output head is fine-tuned. Two types of expanded data are generated using the pre-trained model: extreme scene synthesis and sample augmentation. Extreme scene synthesis: Input the precursor features of historical extreme precipitation events and generate new extreme precipitation sequences through the decoder. Sample augmentation: Perform spatiotemporal interpolation expansion on the existing data. In actual assessment, The value is 0.3, and its range is (0.2, 0.4). The value is 0.4, and its range is (0.3, 0.5). The value is 0.3, and the range is (0.2, 0.4); fine scale is 15 minutes, medium scale is 1 hour, and coarse scale is 3 hours. According to the spatial scale, it is divided into high resolution 1km, medium resolution 4km, and low resolution 16km. Spatiotemporal interpolation: Linear interpolation is used in the time dimension to supplement the time gaps in the precipitation sequence; Kriging interpolation is used in the spatial dimension to optimize the spatial interpolation effect of ground station observations and generate continuous grid data. The design includes a spatiotemporal prediction task to predict precipitation in the next four time steps, a feature reconstruction task to reconstruct precipitation characteristics in the occluded area, and a cross-scale consistency task to upsample low-resolution precipitation to a 1km high resolution. The encoder-decoder architecture is adopted. The encoder is a 3-layer ConvLSTM with a cross-scale attention module, and the decoder is a transposed convolutional upsampling module plus a multi-task output head. The 3-layer ConvLSTM has 64 / 128 / 64 channels and 3×3 cores. A recurrent learning rate strategy and the AdamW optimizer are employed. Single-task pre-training is performed for 100 epochs each, followed by multi-task joint training with weighted coefficients for 200 epochs. The learning rate range of the recurrent learning rate strategy is 1e. -5 ~1e -3 The AdamW optimizer's weight decay is 0.01; Extreme scene synthesis: Inputting precursor features of historical extreme precipitation events, a decoder generates 15,000 new extreme precipitation sequences; Sample augmentation: Spatiotemporal interpolation is performed on the existing fused data. Linear interpolation is used in the time dimension to fill the gaps between precipitation sequences, and Kriging interpolation is used in the spatial dimension to optimize the spatial continuity of station observations, generating 38,000 augmented samples; Finally, the original data and the extended data are merged to obtain a training dataset with a total sample size of 156,000 sets.

[0047] In this embodiment, the method for constructing an extreme precipitation simulation model based on the extended data and the parameter initialization strategy includes: A convolutional long short-term memory network architecture with spatiotemporal attention is adopted as the main network of the extreme precipitation simulation model. The convolutional long short-term memory network architecture includes an encoder, a decoder, and an output layer. The encoder consists of 4 layers of ConvLSTM, which captures the time series dependence of precipitation and embeds a spatial attention module to enhance the features of the center of heavy precipitation. The decoder consists of 4 layers of upsampled convolutions, which combine skip connections to fuse the multi-scale features output by the encoder and introduce channel attention to dynamically adjust the feature weights of different data sources. The output layer consists of 1×1 convolution + sigmoid activation function, which outputs the precipitation intensity prediction of 0-100 mm / h. Atmospheric dynamics equations and topographic modulation factors are used as physical constraint modules in the extreme precipitation simulation model. Atmospheric dynamics equations are introduced as constraints: results conforming to atmospheric motion laws are simulated through an additional loss term, expressed as follows: ,in For additional losses, For the predicted water vapor flux, To reanalyze the water vapor flux of precipitation data, For divergence operators, For Euclidean norm; Terrain modulation factor: Encodes the digital elevation model into a terrain weight map through 1×1 convolution; The initial parameters optimized by the meta-learning output are loaded, and self-supervised pre-training is performed by inputting extended data. The pre-training is completed by using precipitation sequences. This includes the initial weights of convolutional layers and the attention mechanism. Given the core loss of an extreme precipitation simulation model, the expression is:

[0048] in To determine the core loss of the extreme precipitation simulation model, Forecast precipitation within the effective area; In the actual evaluation, the encoder consists of 4 layers of ConvLSTM with 64 / 128 / 256 / 512 channels per layer, all using 3×3 convolutional kernels with a stride of 1 and padding of 1. A spatial attention module is embedded to enhance the features of the center of heavy precipitation. Decoder: 4 layers of upsampled convolutions, combined with skip connections to fuse multi-scale features from the encoder output, and channel attention is introduced to dynamically adjust the weights of radar, satellite, and site features; the upsampled convolutions are transposed convolutions with 3×3 kernels; Output layer: 1×1 convolution + Sigmoid activation function, outputting hourly precipitation intensity prediction of 0-100mm / h, matching the intensity range of extreme precipitation; Atmospheric dynamics equation constraints: The results conforming to the laws of atmospheric motion are simulated through additional loss terms, and the divergence error between the predicted water vapor flux and the water vapor flux in the reanalysis data is calculated. The Euclidean norm loss is used as the physical constraint loss. Topographic modulation factor: The topographic height extracted from the 30m DEM is encoded into a 1km×1km topographic weight map through 1×1 convolution, which quantifies the lifting and blocking effects of topography on precipitation and serves as one of the model input features. Parameter loading: Load the general parameter initialization strategy of the meta-learning output, including the ConvLSTM convolutional layer weights and the initial weights of the attention mechanism; Pre-training fine-tuning: Perform pre-training with precipitation sequence completion on the input extended dataset, with 150 training epochs and a learning rate of 1e. -4 Core loss setting: The core loss formula of the model in the document is adopted, combined with MSE loss and SSIM loss, with a weighting coefficient of 1:0.5, which effectively balances the numerical accuracy of precipitation intensity with the structural similarity of spatial distribution.

[0049] In this embodiment, the method for obtaining regionalized parameters by using parameter sensitivity maps and adversarial networks to perform regionalized adjustments on the extreme precipitation fusion simulation model includes: Historical fused precipitation data and candidate driving factors were acquired. An improved Peter Clark algorithm was used to construct a causal graph using a graph neural network through variable preprocessing, causal structure learning, and parameter sensitivity quantification. Candidate driving factors included 500 hPa geopotential height, 850 hPa temperature advection, whole-layer water vapor flux, altitude, slope, aspect, NDVI, soil moisture, and land use type. Variable preprocessing standardized the driving factors, and Kendall's rank correlation coefficient test was used to screen variables significantly associated with extreme precipitation. Causal structure learning identified causal relationships between factors based on conditional independence tests. Parameter sensitivity quantification calculated the influence weight of each factor on precipitation based on the causal effect strength, generating a parameter sensitivity causal graph, where nodes are driving factors and edge weights are sensitivity coefficients. Based on the causal graph, the sensitivity coefficients of each driving factor are used as node weights. Using geospatial data as a carrier, the sensitivity coefficients of each driving factor in the sub-region are visualized to generate a parameter sensitivity spatial map. The value of each geographic grid in the parameter sensitivity spatial map is the adjustment weight of the corresponding driving factor on the model parameters, realizing the spatial expression of parameter sensitivity. A graph attention network employing a causal attention learning strategy dynamically weights model parameters based on driving factors. The model framework comprises an input layer, a causal attention layer, a backdoor adjustment mechanism, and an output. The input layer transforms the causal graph into graph-structured data, with node features representing the sensitivity coefficients of driving factors and edge features representing the strength of causal relationships between factors. The causal attention layer learns the importance weights of nodes through a multi-head attention mechanism, expressed as:

[0050] in Let r be the feature vector of the r-th node. For attention parameter vectors, Let y be the feature vector of the y-th node. Let be the feature vector of the z-th node. Let r be the importance weight between the r-th node and the y-th node. For splicing operations, For transpose; The backdoor adjustment mechanism borrows the concept of confusion factor control from causal inference, mitigating the impact of non-causal features by intervening in the loss function. The expression is:

[0051] in For the intervention loss function, To predict precipitation, To reanalyze precipitation data, As a shortcut feature, To force the shortcut feature to the value d, For intervention operations; Obtain the parameter adjustment weight matrix corresponding to each driving factor. Adjust the weight matrix based on parameters The general parameters of meta-learning are adapted to different regions, and the expression is:

[0052] in For the number of driving factors, The number of parameters to be adjusted in the extreme precipitation fusion simulation model. For the actual observed value of the driving factor in the v-th region, It is a nonlinear mapping function. The parameter adjustment weight matrix for the v-th region For regionalization parameters, These are the optimized initial parameters; The preset area is divided into sub-regions according to terrain complexity. Each sub-region is executed independently: input the driving factor data of the sub-region, the graph neural network dynamically calculates the parameters of the region, adjusts the weights, updates the model parameters, and verifies them with regional observation data. Generative Adversarial Networks (GANs) are introduced, consisting of a generator and a discriminator. The generator produces a simulated precipitation field based on regionalized parameters, while the discriminator distinguishes between simulated precipitation and real observations, with a focus on extreme values. The game objective is to minimize the generator loss and maximize the discriminator loss, expressed as:

[0053] in The probability of the discriminator classifying the input data. For the generator based on region parameters The output simulated precipitation field, The expected value is represented by the actual observed precipitation field data. After multiple rounds of adversarial training, the final regionalized high-resolution simulation parameters are output. In actual evaluation, the discriminator focuses on extreme values: the loss function of the discriminator is weighted, and the loss weight of extreme precipitation values ​​is increased by 5 times, so that the discriminator focuses on learning the characteristic differences of extreme precipitation during training, and achieves accurate differentiation of extreme values. Parameter sensitivity map construction: Nine candidate driving factors were selected, including 500 hPa geopotential height, 850 hPa temperature advection, whole-layer water vapor flux, altitude, slope, aspect, NDVI, soil moisture, and land use type. Seven factors significantly correlated with extreme precipitation (P < 0.01) were identified using Kendall's rank correlation coefficient test: whole-layer water vapor flux, altitude, slope, NDVI, 500 hPa geopotential height, 850 hPa temperature advection, and soil moisture. An improved Peter Clark algorithm was used to construct a causal graph of a graph neural network through variable preprocessing, causal structure learning, and parameter sensitivity quantification. Nodes represented the selected driving factors, and edge weights were the sensitivity coefficients of each factor to precipitation. Using a 1 km × 1 km geographic grid, the sensitivity coefficients of the driving factors within each sub-region were visualized, generating a parameter sensitivity spatial map. The value of each grid cell represented the adjustment weight of the corresponding driving factor on the model parameters, achieving a spatialized expression of parameter sensitivity. Parameter regional adaptation based on causal attention GAT: The GAT employs a causal attention learning strategy. The input layer transforms the causal graph into graph-structured data, with node features representing the sensitivity coefficients of driving factors and edge features representing the strength of causal associations between factors. Node importance weights are learned through a 3-head multi-head attention mechanism, with the activation function being LeakyReLU (slope 0.2). An intervention loss function from the document is introduced to control the influence of non-causal features such as soil moisture and land use type, and shortcut features are forced to be set to the mean to reduce the interference of confounding factors on the model. Parameter adjustment weight matrices are obtained for 7 types of driving factors (7 being the number of driving factors, 1280 being the number of parameters to be adjusted in the model). General parameters are adapted using the regionalization parameter formula from the document to obtain the initial regionalization parameters for each sub-region. 1. GAN Model Setup: A generative adversarial network (GAN) is constructed. The generator is a pre-adapted extreme precipitation simulation model (inputting regional parameters and outputting simulated precipitation fields). The discriminator is a CNN classifier, which weights extreme values ​​(increasing the loss weight of hourly rainfall intensity ≥ 50 mm to 5 times) to distinguish between simulated extreme precipitation and real observations. The game objective is to minimize the generator loss and maximize the discriminator loss. Training is performed according to the GAN loss formula in the document, with a total of 200 training rounds. Convergence is defined as the rate of change of the generator loss function being less than 1e-4 for 10 consecutive rounds. After multiple rounds of adversarial training, the regionalized high-resolution simulation parameters of each sub-region in the Yangtze-Huaihe River Basin are output, covering regionally customized parameters such as ConvLSTM channel attention weights and terrain modulation factor coefficients.

[0054] In this embodiment, the method for obtaining high-resolution simulation parameters by optimizing the extreme precipitation fusion simulation model using adversarial parameter game theory includes: Construct a parametric game consisting of two parts: a data-driven game and a physical-constrained game. The main components of the parametric game are a generator, a discriminator, and a physical referee. The generator is an extreme precipitation simulation model with the following parameters: Outputs a high-resolution precipitation field; Discriminator: Graph neural network classifier to distinguish generated precipitation from actual extreme precipitation events; Physical judge: Simplifies the physical model and verifies the physical consistency of generated precipitation; Data-driven game theory: Optimizes parameters. To make the output high-resolution precipitation field statistically approximate real extreme precipitation events; Physically constrained game theory: Introducing a physical regularization term through a physical referee model to correct parameters. Satisfy atmospheric dynamic constraints; The objective function is constructed by minimizing the Wasserstein distance between generated precipitation and real events and maximizing the discriminant's discriminant loss. The expression is as follows:

[0055] in For discriminator, For parameters Precipitation generated below, For actual precipitation, To calculate the Wasserstein distance between the generated precipitation and the actual precipitation, This is the distance weighting coefficient. For adversarial loss function; parameter The update rules are as follows: the discriminator is trained 5 times per round, and the parameters are updated using gradient ascent; the generator updates the discriminator once, and the parameters are optimized using gradient descent. ; We introduce three core physical regularization terms: mass conservation constraint, energy conservation constraint, and terrain sensitivity constraint. We then construct a physical loss function, and based on this function, we construct the overall objective function, expressed as:

[0056]

[0057]

[0058]

[0059] in The weight of the physical regularization term. For water vapor flux divergence error, For wind field, For terrain height, For parameters The physical loss function under, For parameters The overall objective function is as follows: For gradient boosting tree model, The x-axis is... The vertical axis is , It is a divergence operator; parameter The optimization strategy is as follows: The problem is decomposed using the alternating direction multiplier method: first, the physical constraint optimization parameters are fixed. Then fix the parameters Adjust the physical parameters and iterate until convergence; Optimal parameters are selected from the post-game parameter space using the Bayesian information criterion. ,satisfy ,in For the number of parameters, For sample size, For parameters Bayesian information criterion, For parameters The likelihood function, the predicted likelihood function value of the extreme precipitation fusion simulation model for the real observed precipitation field in the preset area, is obtained by taking the negative exponent of the core loss of the extreme precipitation fusion simulation model. parameters Output high-resolution simulation parameters, parameters It includes the convolutional kernel weights of the convolutional long short-term memory network, the spatial weight matrix of the attention mechanism, and the threshold parameters of the physical constraints; In actual assessment, The value is 0.1, and the range is (0.05, 0.2); the physical parameters are the mass conservation coefficient, energy conservation coefficient, and terrain sensitivity coefficient. Generator: A regionally customized extreme precipitation simulation model with the following parameters. Outputs a precipitation field with a resolution of 1km; Discriminator: Graph neural network classifier, inputting the spatiotemporal characteristics of the precipitation field to distinguish generated precipitation from real extreme precipitation events; Physical judge: Simplified atmospheric dynamics model to verify the physical consistency of generated precipitation, with core verification of the three constraints of mass conservation, energy conservation, and topographic sensitivity; Parameter optimization strategy: The optimization problem is decomposed using the alternating direction multiplier method. First, the physical constraints are fixed to optimize the parameters. Gradient descent is used to correct physical parameters, with the physical loss function as the objective. The maximum number of iterations is set to 100, and the learning rate is set to 0.0005. Iteration continues until the rate of change of the physical loss function is less than 1e. -5 The time limit is reached; the physical constraints are the mass conservation coefficient, energy conservation coefficient, and terrain sensitivity coefficient. From the parameter space after the game, the optimal parameters were selected by the Bayesian information criterion. The number of parameters was 1280 and the sample size was 156,000. The final output high-resolution simulation parameters are: the kernel weights and biases of ConvLSTM (896 sets in total); the spatial weight matrix and channel weight matrix of the attention mechanism (320 sets in total); and the threshold parameters of physical constraints, including mass conservation coefficient, energy conservation coefficient, and terrain sensitivity coefficient (64 sets in total).

[0060] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A deep learning-based adaptive optimization method for extreme precipitation fusion simulation parameters, characterized in that, Includes the following steps: Collect multi-source heterogeneous data on extreme precipitation in a preset area, and preprocess the multi-source heterogeneous data; the multi-source heterogeneous data includes radar data, satellite data, and ground observation data; An attention mechanism and a convolutional long short-term memory network are used to fuse the multi-source heterogeneous data to obtain fused data, and meta-learning is used to perform shared prior learning on the fused data to obtain a general parameter initialization strategy. The fused data is pre-trained using spatiotemporal multi-scale self-supervision to obtain extended data, and an extreme precipitation simulation model is constructed based on the extended data and the parameter initialization strategy. The extreme precipitation fusion simulation model is regionalized by using parameter sensitivity maps and adversarial networks to obtain regionalized parameters. The extreme precipitation fusion simulation model is then optimized by adversarial parameter game to obtain high-resolution simulation parameters. The regionalized parameters and the high-resolution simulation parameters are then output as optimization results.

2. The adaptive optimization method for extreme precipitation fusion simulation parameters based on deep learning according to claim 1, characterized in that, A method for fusing extreme precipitation data from multiple heterogeneous sources using an attention mechanism and a convolutional long short-term memory network includes: Radar reflectivity factor, satellite-retrieved precipitation, and ground station observations are obtained from multi-source heterogeneous data. The radar reflectivity factor, satellite-retrieved precipitation, and ground station observations are input into a convolutional long short-term memory network. The vertical reflectivity evolution features are extracted through a 3D convolutional layer to obtain radar features. A 2D convolutional layer is used to capture cloud image texture features to obtain satellite features. The station observations are interpolated into grid data through a fully connected layer, and then compressed to 32 channels through a 1×1 convolution to obtain station features. Calculate the spatial gradient of the feature maps from each data source, generate a saliency weight map to obtain the spatial attention weights, and express the following expression: ； in For spatial location Attention weights at the location The spatial coordinates of the feature map For remote sensing features in Spatial gradient magnitude at that location For spectral characteristics in Spatial gradient magnitude at that location For geographical features in Spatial gradient magnitude at that location For feature map height, The width of the feature map; Global average pooling is performed on the data source features. The channel importance weights are learned through an MLP to obtain a three-dimensional weight matrix after broadcasting the channel attention weights in the spatial dimension. The expression is: ； in This is the feature map of the c-th channel. For global average pooling, It is a multilayer perceptron. It is the sigmoid activation function. This is the three-dimensional weight matrix after broadcasting the attention weights of the c-th channel in the spatial dimension. Extreme precipitation data are fused based on the saliency weighting plot and channel importance weighting to obtain fused data, expressed as: ； in For remote sensing features, For spectral characteristics, Geographical features To integrate data, For element-wise multiplication, For spatial attention weights.

3. The adaptive optimization method for extreme precipitation fusion simulation parameters based on deep learning according to claim 1, characterized in that, The method for obtaining the general parameter initialization strategy includes: Obtain auxiliary variables for the preset regions, and divide the data of each source region into a meta-training task set and a meta-testing task set in a 4:1 ratio. Each task includes a support set and a query set. The auxiliary variables are terrain height and vegetation index. The support set consists of a small number of samples. The query set consists of validation samples. Input the meta-training task set and auxiliary variables into the meta-learning model. For each meta-training task, use the support set to update the model parameters and obtain the optimized initial parameters. The expression is: ； in For meta-parameters, For the adapted parameters of task a, Step size, The gradient of the meta-parameters, Neural networks with meta-parameters For the support set of task a, Let be the loss function for task a; Use the updated parameters The meta-loss is calculated on the query set, and the initial parameters are updated via backpropagation. The expression is: ； in Let be the query set for task a. For the number of tasks, The meta-learning rate; By visualizing attention weights, key features shared across regions are identified, and these key features are encoded as constraints for the initial parameters of the model. Use a small number of samples in the target region to optimize the initial parameters. Perform multiple rounds of fine-tuning to quickly converge to the regional optimum, and obtain the adaptation rule based on the optimum solution; then apply the optimized initial parameters. The adaptation rule output is a general parameter initialization strategy.

4. The adaptive optimization method for extreme precipitation fusion simulation parameters based on deep learning according to claim 1, characterized in that, A method for pre-training the fused data using spatiotemporal multi-scale self-supervision to obtain extended data includes: The fused data is divided into temporal and spatial scales, and random transformations are performed on the fused data to generate enhanced samples; spatiotemporal prediction tasks, feature reconstruction tasks, and cross-scale consistency tasks are designed to learn spatiotemporal features from multiple dimensions. A self-supervised pre-trained model architecture is constructed using a deep learning model architecture, and multi-scale feature extraction and fusion are performed using the self-supervised pre-trained model architecture. The deep learning model architecture includes an encoder and a decoder, and the encoder includes a spatiotemporal convolutional model and a cross-scale attention module. Self-supervised pre-training is divided into single-task pre-training and multi-task joint training. For single-task pre-training: train the three main tasks separately and optimize the corresponding loss functions, as expressed in the following expression: ； ； ； in For the predicted four future time steps, Four time steps for the real future. For mean square error loss, For the reconstructed mask area, For the original complete data, The mean absolute error, For similar structural indices, Reconstructed values ​​at 1km resolution. For true values ​​at 1km resolution, For the total variation loss in two-dimensional space, The total variation weight; For multi-task joint training: the losses from the three tasks are fused, and general spatiotemporal features are learned through parameter sharing, expressed as follows: ； in The total pre-training loss, To predict losses, To rebuild the losses, For the excess score loss, To predict the weighting coefficients, To reconstruct the weighting coefficients, This refers to the super-resolution weighting coefficient; A recurrent learning rate strategy is adopted, using the AdamW optimizer. After pre-training, the encoder weights are frozen, and only the decoder output head is fine-tuned. Two types of extended data are generated using the pre-trained model: extreme scene synthesis and sample augmentation.

5. The adaptive optimization method for extreme precipitation fusion simulation parameters based on deep learning according to claim 1, characterized in that, A method for constructing an extreme precipitation simulation model based on the extended data and the parameter initialization strategy includes: A convolutional long short-term memory network architecture with spatiotemporal attention is adopted as the main network of the extreme precipitation simulation model. The convolutional long short-term memory network architecture includes an encoder, a decoder, and an output layer. The encoder consists of 4 layers of ConvLSTM, which captures the time series dependence of precipitation and embeds a spatial attention module to enhance the features of the center of heavy precipitation. The decoder consists of 4 layers of upsampled convolutions, which combine skip connections to fuse the multi-scale features output by the encoder and introduce channel attention to dynamically adjust the feature weights of different data sources. The output layer consists of 1×1 convolution + sigmoid activation function, which outputs the precipitation intensity prediction of 0-100 mm / h. Atmospheric dynamics equations and topographic modulation factors are used as physical constraint modules in the extreme precipitation simulation model. Atmospheric dynamics equations are introduced as constraints: results conforming to atmospheric motion laws are simulated through an additional loss term, expressed as follows: ,in For additional losses, 1. To predict precipitation, To reanalyze precipitation data, For divergence operators, For Euclidean norm; Terrain modulation factor: Encodes the digital elevation model into a terrain weight map through 1×1 convolution; The initial parameters optimized by the meta-learning output are loaded, and self-supervised pre-training is performed by inputting extended data. The pre-training is completed by using precipitation sequences. This includes the initial weights of convolutional layers and the attention mechanism. Given the core loss of an extreme precipitation simulation model, the expression is: ； in To determine the core loss of the extreme precipitation simulation model, This represents the predicted precipitation within the effective area.

6. The adaptive optimization method for extreme precipitation fusion simulation parameters based on deep learning according to claim 1, characterized in that, A method for obtaining regionalized parameters by using parameter sensitivity maps and adversarial networks to regionally customize the extreme precipitation fusion simulation model includes: Historical fused precipitation data and candidate driving factors were acquired, and an improved Peter Clark algorithm was used to construct a causal graph of a graph neural network through variable preprocessing, causal structure learning, and parameter-sensitive quantification. A graph attention network employing a causal attention learning strategy is used to dynamically weight the model parameters based on driving factors. The input layer, causal attention layer, backdoor adjustment mechanism, and output are used as the model framework of the graph neural network. The backdoor adjustment mechanism borrows the concept of confusion factor control from causal inference, mitigating the impact of non-causal features by intervening in the loss function. The expression is: ； in For the intervention loss function, To predict precipitation, To reanalyze precipitation data, As a shortcut feature, To force the shortcut feature to the value d, For intervention operations; Obtain the parameter adjustment weight matrix corresponding to each driving factor. Adjust the weight matrix based on parameters Regional adaptation is performed on the general parameters of meta-learning; The preset area is divided into sub-regions according to terrain complexity. Each sub-region is executed independently: input the driving factor data of the sub-region, the graph neural network dynamically calculates the parameters of the region, adjusts the weights, updates the model parameters, and verifies them with regional observation data. Generative Adversarial Networks (GANs) are introduced, consisting of a generator and a discriminator. The generator produces a simulated precipitation field based on regionalized parameters, while the discriminator distinguishes between simulated precipitation and real observations, with a focus on extreme values. The game objective is to minimize the generator loss and maximize the discriminator loss, expressed as: ； in The probability of the discriminator classifying the input data. For the generator based on region parameters The output simulated precipitation field, The expected value is represented by the actual observed precipitation field data. After multiple rounds of adversarial training, the final regionalized high-resolution simulation parameters are output.

7. The adaptive optimization method for extreme precipitation fusion simulation parameters based on deep learning according to claim 1, characterized in that, A method for obtaining high-resolution simulation parameters by optimizing the extreme precipitation fusion simulation model using adversarial parameter game theory includes: Construct two parametric games, one consisting of data-driven game and the other of physical constraint game. The main components of the parametric game are the generator, the discriminator, and the physical referee. The objective function is constructed by minimizing the Wasserstein distance between generated precipitation and real events and maximizing the discriminant's discriminant loss. The expression is as follows: ； in For discriminator, For parameters Precipitation generated below, For actual precipitation, To calculate the Wasserstein distance between the generated precipitation and the actual precipitation, This is the distance weighting coefficient. For adversarial loss function; parameter The update rules are as follows: the discriminator is trained 5 times per round, and the parameters are updated using gradient ascent; the generator updates the discriminator once, and the parameters are optimized using gradient descent. ; We introduce three core physical regularization terms: mass conservation constraint, energy conservation constraint, and terrain sensitivity constraint, and construct a physical loss function. Based on the physical loss function, we construct the overall objective function. parameter The optimization strategy is as follows: The problem is decomposed using the alternating direction multiplier method: first, the physical constraint optimization parameters are fixed. Then fix the parameters Adjust the physical parameters and iterate until convergence; Optimal parameters are selected from the post-game parameter space using the Bayesian information criterion. ,satisfy ,in For the number of parameters, For sample size, For parameters The likelihood function, For parameters Bayesian information criterion; parameters Output high-resolution simulation parameters, parameters It includes the convolutional kernel weights of the convolutional long short-term memory network, the spatial weight matrix of the attention mechanism, and the threshold parameters of the physical constraints.

Images