Unified precipitation downscaling method based on multi-stream analysis diffusion model

Abstract

The invention discloses a unified rainfall downscaling method based on a multi-stream analysis diffusion model, and the method comprises the steps: collecting the meteorological data of a target region, including low-resolution rainfall data and preset auxiliary variable data, generating a low-resolution rainfall field, and calculating the rainfall deviation between a high-resolution real rainfall field and the low-resolution rainfall field; based on preset auxiliary variable data, combining the low-resolution rainfall data to form multi-modal condition input; constructing a precipitation downscaling model which takes the analysis diffusion model as a framework and is combined with a mixed attention U-shaped network, and predicting precipitation deviation; and training the rainfall downscaling model, performing reasoning by adopting the rainfall downscaling model, inputting meteorological data of the target area, outputting predicted rainfall deviation, and superposing the predicted rainfall deviation to the low-resolution rainfall field to obtain super-resolution rainfall. According to the method, multi-source meteorological and geographic information is effectively fused, high-fidelity and high-resolution rainfall data is generated, the generalization ability is high, and the calculation efficiency is high.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of meteorology and artificial intelligence, specifically to a unified precipitation downscaling method based on a multi-flow analytical diffusion model. Background Technology

[0002] Accurate precipitation forecasting is a core pillar of modern meteorology, playing a crucial role in key areas such as agricultural scheduling, water resource allocation, and disaster risk mitigation. However, the spatial resolution of precipitation data provided by global climate models and weather forecasting systems is typically low, making it difficult to meet the refined decision-making requirements of local areas. With the dramatic increase in demand for high-resolution precipitation data due to climate change, the necessity of advanced methods to bridge the gap between coarse-resolution model outputs and local-scale requirements is highlighted.

[0003] Precipitation downscaling is a key technology for addressing this core challenge. Existing downscaling methods are mainly divided into traditional methods (such as dynamical downscaling and statistical downscaling) and deep learning methods. Traditional methods often rely heavily on region-specific prior knowledge. Although these techniques can effectively improve the local characteristics of precipitation data, they greatly limit the model's generalization ability across different geographical landscapes. In contrast, deep learning techniques have significantly improved precipitation-related accuracy, but their adaptability is relatively limited; these models typically need to be retrained for each new geographical region. Another challenge is that the ever-increasing resource requirements may lead to excessively high training and deployment costs for models in global high-resolution application scenarios, thus limiting their practical scalability and accessibility. In addition, a key issue is the inability to efficiently utilize multivariate inputs, resulting in ambiguous output results or poor performance on extreme precipitation events.

[0004] Given the aforementioned limitations, there is an urgent need for a framework that is scalable and generalizable, capable of integrating multiple data sources and ensuring high-fidelity output results. Summary of the Invention

[0005] The purpose of this invention is to provide a unified precipitation downscaling method based on a multi-stream analytical diffusion model. This method can effectively integrate multi-source meteorological and geographic information, generate physically consistent, high-fidelity, and high-resolution precipitation data, and has strong generalization ability and high computational efficiency.

[0006] To achieve the above functions, this invention designs a unified precipitation downscaling method based on a multi-flow analytical diffusion model, executing the following steps S1-S5 to complete the precipitation downscaling processing of meteorological data in the target area:

[0007] Step S1: Collect meteorological data of the target area, including low-resolution precipitation data and preset auxiliary variable data, and generate a low-resolution precipitation field based on the low-resolution precipitation data; collect a high-resolution real precipitation field and calculate the precipitation deviation between the high-resolution real precipitation field and the low-resolution precipitation field.

[0008] Step S2: Based on the preset auxiliary variable data and combined with low-resolution precipitation data, construct multimodal conditions;

[0009] Step S3: Construct a precipitation downscaling model based on the analytical diffusion model and combined with a hybrid attention U-shaped network. The hybrid attention U-shaped network integrates channel attention modules and local importance attention modules. The precipitation downscaling model takes noisy precipitation bias, time step information, and multimodal conditions as inputs and the predicted precipitation bias as output.

[0010] Step S4: Train the precipitation downscaling model to obtain a trained precipitation downscaling model;

[0011] Step S5: Use a precipitation downscaling model for inference, input meteorological data of the target area, output the predicted precipitation deviation, superimpose the predicted precipitation deviation onto the low-resolution precipitation field to obtain super-resolution precipitation, and perform post-processing. Use the post-processed super-resolution precipitation as the precipitation downscaling result of the meteorological data of the target area.

[0012] Beneficial effects: Compared with the prior art, the advantages of the present invention include:

[0013] 1. By constraining the analytical diffusion model with meteorological and geographical conditions and training it to learn precipitation bias, the model can generate high-resolution output results with high physical consistency, which fits the complex long-tail distribution characteristics of precipitation, thereby improving its generalization ability in various geographical regions.

[0014] 2. Enhance the multimodal data fusion effect and model transfer capability by using a hybrid attention mechanism (channel attention and local importance attention). This dual attention mechanism enables the model to adaptively focus on salient information and improve generalization performance without relying on region-specific prior knowledge.

[0015] 3. The model is designed to be efficient, enabling it to be trained on a small dataset and achieve good performance, thus improving computational efficiency and reducing computational overhead. Attached Figure Description

[0016] Figure 1 This is an architecture diagram of a unified precipitation downscaling method based on a multi-flow analytical diffusion model provided by an embodiment of the present invention;

[0017] Figure 2This is a schematic diagram of a hybrid attention U-shaped network provided according to an embodiment of the present invention;

[0018] Figure 3 This is a schematic diagram of a channel attention module provided according to an embodiment of the present invention;

[0019] Figure 4 This is a schematic diagram of a local importance attention module provided according to an embodiment of the present invention;

[0020] Figure 5 This is a comparison chart of critical success indices for extreme precipitation events provided by embodiments of the present invention;

[0021] Figure 6 This is a comparison chart of Hederick skill scores for extreme precipitation events provided according to an embodiment of the present invention;

[0022] Figure 7 This is a comparison chart of the radial average power spectral density of each model downscaling result and high-resolution reference data, provided according to an embodiment of the present invention. Detailed Implementation

[0023] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0024] This invention provides a unified precipitation downscaling leveraging method (Unified Precipitation Downscaling leveraging Physics-Informed Diffusion and Mixed-Attention UNet, UPD-Diff) based on a multi-flow analytical diffusion model, referencing... Figure 1 Perform the following steps S1-S5 to complete the precipitation downscaling of the meteorological data for the target area:

[0025] Step S1: Collect meteorological data of the target area, including low-resolution precipitation data and preset auxiliary variable data, and generate a low-resolution precipitation field based on the low-resolution precipitation data; collect a high-resolution real precipitation field and calculate the precipitation deviation between the high-resolution real precipitation field and the low-resolution precipitation field.

[0026] The specific steps of step S1 are as follows:

[0027] Step S1.1: Collect meteorological data for the target area, including low-resolution precipitation data and preset auxiliary variable data. Normalize the low-resolution precipitation data and preset auxiliary variable data using the Z-score method; the normalization formula is:

[0028] ;

[0029] In the formula, and These are the mean and standard deviation obtained from the training set, respectively. For meteorological data of the target area, The normalized meteorological data for the target area;

[0030] Step S1.2: Based on the normalized low-resolution precipitation data, perform bicubic interpolation to generate a low-resolution precipitation field;

[0031] Step S1.3: Collect high-resolution real precipitation fields and calculate the precipitation deviation between the high-resolution real precipitation fields and the low-resolution precipitation fields.

[0032] Step S2: Based on the preset auxiliary variable data and combined with low-resolution precipitation data, construct multimodal conditions;

[0033] In step S2, the normalized low-resolution precipitation data is concatenated with the normalized preset auxiliary variable data to form a multimodal condition.

[0034] The preset auxiliary variable data includes meteorological variables and geographical variables;

[0035] The meteorological variables include at least one of temperature, humidity, and wind field components;

[0036] The geographic variables include at least one of elevation and land-sea boundary data.

[0037] Step S3: Construct a precipitation downscaling model based on the elastic diffusion (EDM) model and combined with a mixed-attention UNet (MA-UNet), referring to... Figure 2 The hybrid attention U-shaped network integrates channel attention modules and local importance attention modules; the precipitation downscaling model takes noisy precipitation bias, time step information, and multimodal conditions as inputs and the predicted precipitation bias as output.

[0038] The analytical diffusion model described in step S3 includes a forward diffusion process and a reverse diffusion process, as detailed below:

[0039] Forward diffusion process: based on high-resolution precipitation baseline values Starting with the noise customization scheme preset in the analytical diffusion model, the following steps are taken: Gaussian noise is added incrementally at each time step, with this pre-defined noise scheme prioritizing the modeling of extreme precipitation events. At each time step... The precipitation deviation sample is generated as follows:

[0040] ;

[0041] In the formula, For time steps Precipitation deviation sample, For time steps Precipitation deviation sample, For noise scheduling parameters, Indicates time step noise, express Follows a standard normal distribution. Represents random noise; with time steps Advance to This process will It gradually transforms into isotropic noise.

[0042] Back diffusion process: at each time step It follows the Gaussian transition mechanism as follows, specifically as shown in the following formula:

[0043] ;

[0044] In the formula, ) is the noise estimate of the output of the hybrid attention U-shaped network. It is the time-varying variance obtained from the analytical diffusion model. This physical information constraint is embedded with meteorological restrictions to ensure the physical rationality of the reconstructed precipitation field. For multimodal conditions; Represents the transfer distribution of back diffusion. Indicates a normal distribution. For time steps Precipitation deviation sample, For time steps Precipitation deviation samples;

[0045] The analytical diffusion model predicts the noise introduced at each time step. The prediction loss function of the analytical diffusion model is constructed as follows:

[0046] ;

[0047] In the formula, To analyze the prediction loss function of the diffusion model, ) is the noise estimate of the output of the hybrid attention U-shaped network. Indicates noise. The expected value is expressed; the prediction loss function of the analytical diffusion model is aligned with the denoising target, and the prediction effect is optimized through multimodal input.

[0048] The hybrid attention U-Net is a U-Net network based on an encoder-decoder structure, which incorporates channel attention modules and local importance attention modules in the skip connection paths between the encoder and decoder;

[0049] Among them, reference Figure 3 The channel attention module assigns different weights to each channel of the input features, thereby reducing redundancy and prioritizing the retention of information features such as humidity and wind pattern to emphasize information-rich feature channels. The input layer of the channel attention module has a two-branch structure. One branch sequentially connects a max pooling layer, a linear layer, a ReLU activation function layer, and another linear layer. The other branch sequentially connects an average pooling layer, a linear layer, a ReLU activation function layer, and another linear layer. The outputs of the linear layers from both branches are concatenated, passed through a sigmoid activation function layer, and the input is weighted by calculating the tensor product between the output of the sigmoid activation function layer and the input of the channel attention module. The result is then output through the output layer.

[0050] Among them, reference Figure 4 The local importance attention module generates a spatial importance map to highlight areas of intense precipitation, thereby enhancing local details and improving the prediction accuracy of extreme weather events. After the input layer, the local importance attention module passes through a two-dimensional soft pooling layer and a convolutional kernel. The two-dimensional convolutional layer and convolutional kernel are The two-dimensional convolutional layer and convolutional kernel are The system consists of a 2D convolutional layer, a sigmoid activation function layer, and a bilinear interpolation function layer. The tensor product of the output of the bilinear interpolation function layer and the output of the local importance attention module is calculated and then output through the output layer.

[0051] In summary, the channel attention module and the local importance attention module enable the hybrid attention U-shaped network to efficiently integrate multimodal data, achieving a balance between global consistency and local accuracy.

[0052] Step S4: Train the precipitation downscaling model to obtain a trained precipitation downscaling model;

[0053] In step S4, during the training of the precipitation downscaling model, the precipitation bias prediction loss function is constructed as follows:

[0054] ;

[0055] in, The precipitation bias prediction loss function is used in the precipitation downscaling model. Let be the predicted precipitation bias value for the i-th sample. Let be the actual precipitation deviation value for the i-th sample; The weights are determined by the intensity of precipitation, which increases with the intensity of precipitation to prioritize the accuracy of extreme event predictions. The total number of samples;

[0056] The training loss function for constructing the precipitation downscaling model is as follows:

[0057] ;

[0058] in, Let be the training loss function for the precipitation downscaling model. To analyze the prediction loss function of the diffusion model, The parameters are used to balance the diffusion mission and the precipitation deviation prediction mission.

[0059] In this embodiment, the AdamW optimizer is used, and the learning rate is set to The batch size is 16; the model is trained for 200 epochs. If the mean square error of the validation set does not improve for 20 consecutive epochs, the early stopping mechanism is triggered.

[0060] Performance evaluation uses pixel-level metrics to assess spatial fidelity, while using event-based metrics to measure the ability to detect extreme events, thereby achieving a comprehensive and integrated evaluation.

[0061] The core of the precipitation downscaling model lies in integrating the analytical diffusion model framework with a hybrid attention U-shaped network. This integration is specifically designed to perform precipitation downscaling tasks and is guided by multimodal inputs. Unlike traditional denoising diffusion probability models, the analytical diffusion model optimizes noise scheduling and sampling techniques to better adapt to the long-tailed distribution unique to precipitation data. This optimization significantly enhances the correlation between the generated super-resolution precipitation field and its high-resolution ground truth values. The complete training process is as follows:

[0062] Training dataset ;

[0063] In the formula, Let j be the j-th high-resolution precipitation baseline value. Let j be the j-th multimodal condition; N be the total number of samples;

[0064] Set diffusion time step Balance parameters ;

[0065] Initialize the parameters of the hybrid attention U-shaped network ;

[0066] Initialize the optimizer;

[0067] Repeat execution:

[0068] from Sampling a mini batch ; This represents the high-resolution precipitation baseline value in the mini-batch. This indicates the multimodal conditions in a mini-batch;

[0069] Sampling time step , This indicates uniform sampling of the set;

[0070] Sampling noise ; Indicates noise. express It follows a standard normal distribution;

[0071] Using analytical diffusion model noise scheduling, based on Constructing time steps Precipitation deviation sample ;

[0072] Through a hybrid attention U-shaped network ( Predicting noise ;

[0073] Calculate the prediction loss function of the analytical diffusion model. ;

[0074] Optionally, if the final precipitation deviation value is predicted in the same step... , then Prepare the precipitation deviation value; based on the predicted precipitation deviation value Deviation from actual precipitation Calculate the precipitation bias prediction loss function of the precipitation downscaling model. , where M represents the number of randomly selected small batch samples; Let be the predicted precipitation bias value for the i-th sample. This represents the true precipitation deviation value for the i-th sample. As weight;

[0075] Calculate the training loss function of the precipitation downscaling model. ;

[0076] Using optimizers and loss Update parameters of the hybrid attention U-shaped network ;

[0077] Until the convergence condition is met.

[0078] Step S5: Use a precipitation downscaling model for inference, input meteorological data of the target area, output the predicted precipitation deviation, superimpose the predicted precipitation deviation onto the low-resolution precipitation field to obtain super-resolution precipitation, and perform post-processing. Use the post-processed super-resolution precipitation as the precipitation downscaling result of the meteorological data of the target area.

[0079] The specific steps of step S5 are as follows:

[0080] Step S5.1: Collect meteorological data for the target area. To maintain consistency with training, all input meteorological variables are normalized using the Z-score method, as shown in the following formula:

[0081] ;

[0082] In the formula, and These are the mean and standard deviation obtained from the training set, respectively. For meteorological data of the target area, The target area meteorological data is normalized; this normalization process standardizes the input, preserves the relative contribution of variables, and improves the stability of model training.

[0083] Step S5.2: The inference process utilizes the backdiffusion mechanism of the analytical diffusion model. Under the constraint of normalized multimodal input, the model iteratively denoises noisy samples of the precipitation baseline. The denoising process is configured to 100 steps to balance computational efficiency and output fidelity. Using the backdiffusion process of the analytical diffusion model, noise samples randomly generated and following a standard normal distribution are denoised. Initially, the analytical diffusion model is used at each time step. Perform reverse diffusion transformation:

[0084] ;

[0085] In the formula, For time steps Precipitation deviation sample, For time steps Precipitation deviation sample, It is the noise estimate of the output of the hybrid attention U-shaped network. It is a multimodal condition. To analyze the time-varying variance obtained from the diffusion model, after a preset number of iterations, the analytical diffusion model outputs the predicted normalized precipitation bias value. ; For random noise, Represents random noise term It follows a standard normal distribution;

[0086] Step S5.3: Since the diffusion process is based on normalized data, the predicted normalized precipitation bias is adjusted by inverse Z-fractional normalization. To convert back to the original scale, the formula is:

[0087] ;

[0088] In the formula, This represents the precipitation deviation value at the original scale. and These are the mean and standard deviation of precipitation deviations, calculated from the training set.

[0089] Super-resolution precipitation is obtained by superimposing the original-scale precipitation bias values ​​onto the bicubic interpolated low-resolution precipitation field, as shown in the formula:

[0090] ;

[0091] in, This is a low-resolution precipitation field after bicubic interpolation. This represents the precipitation deviation value at the original scale. For super-resolution precipitation; this step ensures that the model output, based on coarse-scale information, supplements the fine-scale details captured by the baseline.

[0092] Step S5.4: To ensure physical validity, a threshold correction is performed on the super-resolution precipitation data, as shown in the following formula:

[0093] ;

[0094] in, Threshold-corrected super-resolution precipitation This step, which uses a maximum value function, eliminates physically unreasonable negative precipitation values, improving the applicability of the output in practical applications such as weather forecasting and climate simulation. The resulting super-resolution precipitation field possesses both high spatial fidelity and conforms to meteorological constraints.

[0095] During the inference phase, the precipitation downscaling model generates a high-resolution precipitation field by predicting a precipitation baseline value.

[0096] By adopting this baseline centering strategy, the model can focus on capturing fine-scale details rather than directly reconstructing the entire high-resolution precipitation field. The complete inference process is as follows:

[0097] Normalized multimodal conditions ;

[0098] Low-resolution interpolated precipitation data ;

[0099] Parameters of the trained hybrid attention U-shaped network ;

[0100] diffusion steps ;

[0101] Mean of precipitation bias in training set and standard deviation ;

[0102] Initial sampling noise The initial noise follows a standard normal distribution.

[0103] for ,implement:

[0104] Based on noise prediction Using a hybrid attention U-shaped network to predict noise estimates ;

[0105] like ,sampling ;otherwise ; For random noise;

[0106] calculate ;

[0107] End the loop;

[0108] The predicted normalized precipitation bias value was obtained. ;

[0109] Inverse normalization is performed on the predicted normalized precipitation bias: ;

[0110] Reconstructing high-resolution precipitation: ;

[0111] Post-processing: ;

[0112] Return threshold-corrected super-resolution precipitation ;

[0113] The inference process seamlessly integrates several key steps: data normalization, diffusion-based denoising iteration, inverse normalization, benchmark superposition, and post-processing.

[0114] The following are application examples of the present invention:

[0115] 1. Experimental setup:

[0116] The performance of the proposed model was evaluated through ablation experiments and comparisons with current state-of-the-art benchmark models. Evaluation metrics included pixel-level fidelity (such as PSNR, SSIM) and meteorological efficacy (such as CSI, HSS), with a focus on the accurate capture of extreme precipitation events.

[0117] All experiments were conducted on a global test set sampled from the CMA-GFS and ERA5 datasets.

[0118] In addition, all experiments were run on a single Nvidia RTX 4090 GPU: the model was trained for 200 epochs, with early stopping implemented based on a patience threshold of 20 epochs.

[0119] Regarding runtime characteristics: On an RTX 4090, with a configuration of 100 diffusion steps and a batch size of 8, it took 34 minutes to process 400 blocks.

[0120] 2. Dataset and Evaluation Metrics:

[0121] The dataset used in this embodiment integrates high-resolution (0.125°) meteorological variables from CMA-GFS with elevation and land / sea data from ERA5. This global multimodal dataset captures both dynamic weather patterns and static geographic features.

[0122] The preprocessing steps included Z-score normalization and extraction of 224×224 blocks. The model learned the bias between predicting high-resolution precipitation and predicting low-resolution precipitation using bicubic interpolation.

[0123] The training set is balanced with 4000 images, and the validation and test sets each contain 400 images.

[0124] 3. Quantitative and Qualitative Results:

[0125] (1) Ablation experiments: To evaluate the contributions of different components in the framework, experiments were conducted based on the following configurations, each corresponding to a specific dimension of the framework:

[0126] EDM-SingleInput: Based on the baseline model, it integrates only low-resolution precipitation data with auxiliary meteorological factors and geographical features.

[0127] EDM-MultiInput: Incorporates low-resolution precipitation data while overlaying auxiliary meteorological and geographical variables.

[0128] EDM-ChannelAttn (EDM-Channel Attention): Introduces a channel attention mechanism into the U-Net network architecture, prioritizing the most relevant feature channels to improve performance.

[0129] EDM-LocalAttn (EDM-LocalAttention): Integrates local importance attention into the U-Net network architecture, directing the model's focus to key spatial regions in the data.

[0130] UPD-Diff (UPD-Diffusion Model): Represents a complete precipitation downscaling model architecture. It combines channel attention and local importance attention in the U-Net network architecture to provide a comprehensive solution.

[0131] EDM-ChannelSpatialAttn (EDM-Channel Spatial Attention): Simultaneously utilizes channel and spatial attention to focus on features more broadly, enhancing the model's ability to capture complex patterns.

[0132] EDM-CrossAttn (EDM-CrossAttn): Employs a cross-attention mechanism to model the complex relationship between precipitation and auxiliary variables.

[0133] From relying solely on low-resolution precipitation data to integrating auxiliary meteorological and geographical variables, the SSIM improved from 0.6672 to 0.7766, highlighting the importance of multimodal information in capturing complex spatial patterns in precipitation fields. However, the PSNR only improved slightly, while the RMSE increased slightly, indicating that although multi-input integration helps improve structure fidelity, further optimization is needed to achieve optimal pixel-level accuracy.

[0134] Analysis of the impact of different attention mechanisms on multi-input baseline models revealed differential results:

[0135] Compared to EDM-MultiInput, EDM-ChannelAttn shows improvements in both PSNR and Corr, demonstrating that prioritizing key feature channels enhances pixel-level prediction and linear correlation. However, its SSIM is slightly lower.

[0136] Conversely, EDM-LocalAttn leads to a significant increase in RMSE, and SSIM and Corr are significantly lower than EDM-MultiInput.

[0137] This suggests that focusing solely on local saliency while ignoring the global context or channel-specific feature weighting may disrupt overall structural consistency and amplify pixel errors.

[0138] EDM-CrossAttn shows a slight improvement in PSNR and Corr, but a significant decrease in SSIM. This is consistent with expectations for heterogeneous data distribution: without proper distribution alignment, the complexity of cross-attention may not be fully utilized and could instead interfere with structural information.

[0139] The performance of EDM-ChannelSpatialAttn did not surpass that of a simpler configuration. This indicates that simple combinations of attention mechanisms do not necessarily improve performance.

[0140] Ultimately, the precipitation downscaling model framework proposed in this invention demonstrates robust performance: among the benchmark models compared, it achieves the highest SSIM and Corr, and is also competitive in PSNR and RMSE. This reflects the complementary advantages of the hybrid attention U-shaped network design: channel attention effectively balances the contributions of different input features, while local importance attention refines the focus on spatially significant precipitation patterns. This combination enables the precipitation downscaling model of this invention to maintain competitive pixel-level fidelity, preserve structural information, and achieve high overall relevance, while also verifying the effectiveness of the hybrid attention strategy in capturing fine-grained details and broader global contextual features of precipitation. Notably, this robust performance, combined with the excellent pixel-level performance of the precipitation downscaling model, demonstrates its superior ability to detect extreme precipitation events at different intensity thresholds.

[0141] This invention provides a comparison chart of the Critical Success Index (CSI) for extreme precipitation events. Figure 5 The comparison chart of Heidegger Skill Score (HSS) for extreme precipitation events is shown below. Figure 6 ;

[0142] (2) Comparison with current best baseline models: The performance of the precipitation downscaling model of the present invention is evaluated by comparing it with a variety of current best baseline model methods, including CNN-based methods, GAN-based methods, and standard U-shaped network baseline models (whose architecture is consistent with the model of the present invention, but uses pixel-level L1 loss for training). This choice provides a reliable basis for a comprehensive comparison of the analytical diffusion model with existing frameworks.

[0143] Reference Figure 7 To further investigate the model's ability to reproduce real spatial structures at different scales, the radial average power spectral density was analyzed in this embodiment.

[0144] The spectrum of the precipitation downscaling model of this invention closely matches that of high-resolution reference data, especially in the high-frequency regions corresponding to fine-scale details. Many benchmark models often struggle to accurately reproduce small-scale features, indicating that the precipitation downscaling model of this invention not only preserves large-scale patterns but also generates more realistic and detailed textures than other deep learning models. This spectral fidelity is crucial for capturing the hierarchical characteristics of precipitation systems.

[0145] Finally, the meteorological skill of the precipitation downscaling model of this invention was compared with that of the current best baseline model. The baseline model, which is partially optimized based on L1 loss, performs better in light rain scenarios but poorly in high-intensity precipitation scenarios.

[0146] In contrast, the method of this invention demonstrates a significant advantage in heavy and extreme precipitation scenarios by learning the overall data distribution.

[0147] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A unified precipitation downscaling method based on a multi-flow analytical diffusion model, characterized in that, Perform the following steps S1-S5 to complete the precipitation downscaling of the meteorological data for the target area: Step S1: Collect meteorological data of the target area, including low-resolution precipitation data and preset auxiliary variable data, and generate a low-resolution precipitation field based on the low-resolution precipitation data; collect a high-resolution real precipitation field and calculate the precipitation deviation between the high-resolution real precipitation field and the low-resolution precipitation field. Step S2: Based on the preset auxiliary variable data and combined with low-resolution precipitation data, construct multimodal conditions; Step S3: Construct a precipitation downscaling model based on the analytical diffusion model and combined with a hybrid attention U-shaped network, wherein the hybrid attention U-shaped network integrates channel attention module and local importance attention module; The precipitation downscaling model takes noisy precipitation bias, time step information, and multimodal conditions as inputs and the predicted precipitation bias as output. Step S4: Train the precipitation downscaling model to obtain a trained precipitation downscaling model; Step S5: Use a precipitation downscaling model for inference, input meteorological data of the target area, output the predicted precipitation deviation, superimpose the predicted precipitation deviation onto the low-resolution precipitation field to obtain super-resolution precipitation, and perform post-processing. Use the post-processed super-resolution precipitation as the precipitation downscaling result of the meteorological data of the target area.

2. The unified precipitation downscaling method based on a multi-flow analytical diffusion model according to claim 1, characterized in that, The specific steps of step S1 are as follows: Step S1.1: Collect meteorological data of the target area, including low-resolution precipitation data and preset auxiliary variable data. Normalize the low-resolution precipitation data and preset auxiliary variable data using the Z-score method. Step S1.2: Based on the normalized low-resolution precipitation data, perform bicubic interpolation to generate a low-resolution precipitation field; Step S1.3: Collect high-resolution real precipitation fields and calculate the precipitation deviation between the high-resolution real precipitation fields and the low-resolution precipitation fields.

3. The unified precipitation downscaling method based on a multi-flow analytical diffusion model according to claim 2, characterized in that, In step S2, the normalized low-resolution precipitation data is concatenated with the normalized preset auxiliary variable data to form a multimodal condition. The preset auxiliary variable data includes meteorological variables and geographical variables; The meteorological variables include at least one of temperature, humidity, and wind field components; The geographic variables include at least one of elevation and land-sea boundary data.

4. The unified precipitation downscaling method based on a multi-flow analytical diffusion model according to claim 3, characterized in that, The analytical diffusion model described in step S3 includes a forward diffusion process and a reverse diffusion process, as detailed below: Forward diffusion process: based on high-resolution precipitation baseline values Starting with the noise customization scheme preset in the analytical diffusion model, the following steps are taken: Gaussian noise is added progressively at each time step. The precipitation deviation sample is generated as follows: ； In the formula, For time steps Precipitation deviation sample, For time steps Precipitation deviation sample, For noise scheduling parameters, Indicates time step noise, express Follows a standard normal distribution. Indicates random noise; Back diffusion process: at each time step It follows the Gaussian transition mechanism as follows, specifically as shown in the following formula: ； In the formula, ) is the noise estimate of the output of the hybrid attention U-shaped network. It is the time-varying variance obtained from the analytical diffusion model. For multimodal conditions; Represents the transfer distribution of back diffusion. Indicates a normal distribution. For time steps Precipitation deviation sample, For time steps Precipitation deviation samples; The analytical diffusion model predicts the noise introduced at each time step. The prediction loss function of the analytical diffusion model is constructed as follows: ； In the formula, To analyze the prediction loss function of the diffusion model, ) is the noise estimate of the output of the hybrid attention U-shaped network. Indicates noise. It expresses expectation.

5. A unified precipitation downscaling method based on a multi-flow analytical diffusion model according to claim 4, characterized in that, The hybrid attention U-Net is a U-Net network based on an encoder-decoder structure, which incorporates channel attention modules and local importance attention modules in the skip connection paths between the encoder and decoder; The channel attention module has a two-branch structure after the input layer. One branch is connected in sequence with a max pooling layer, a linear layer, a ReLU activation function layer, and another linear layer. The other branch is connected in sequence with an average pooling layer, a linear layer, a ReLU activation function layer, and another linear layer. The outputs of the linear layers of the two branches are concatenated, passed through a sigmoid activation function layer, and the input is weighted by calculating the tensor product between the output of the sigmoid activation function layer and the input of the channel attention module. The output is then passed through the output layer. The local importance attention module, after its input layer, sequentially passes through a two-dimensional soft pooling layer and a convolutional kernel. The two-dimensional convolutional layer and convolutional kernel are The two-dimensional convolutional layer and convolutional kernel are The system consists of a 2D convolutional layer, a sigmoid activation function layer, and a bilinear interpolation function layer. The tensor product of the output of the bilinear interpolation function layer and the output of the local importance attention module is calculated and then output through the output layer.

6. The unified precipitation downscaling method based on a multi-flow analytical diffusion model according to claim 5, characterized in that, In step S4, during the training of the precipitation downscaling model, the precipitation bias prediction loss function is constructed as follows: ； in, The precipitation bias prediction loss function is used in the precipitation downscaling model. Let be the predicted precipitation bias value for the i-th sample. Let be the actual precipitation deviation value for the i-th sample; As weight, The total number of samples; The training loss function for constructing the precipitation downscaling model is as follows: ； in, Let be the training loss function for the precipitation downscaling model. To analyze the prediction loss function of the diffusion model, The parameters are used to balance the diffusion mission and the precipitation deviation prediction mission.

7. A unified precipitation downscaling method based on a multi-flow analytical diffusion model according to claim 6, characterized in that, The specific steps of step S5 are as follows: Step S5.1: Collect meteorological data of the target area and normalize it using the Z-score method; Step S5.2: Using the back diffusion process of the analytical diffusion model, from randomly generated noise samples that follow a standard normal distribution... Initially, the analytical diffusion model is used at each time step. Perform reverse diffusion transformation: ； In the formula, It is the noise estimate of the output of the hybrid attention U-shaped network. It is a multimodal condition. To analyze the time-varying variance obtained from the diffusion model, after a preset number of iterations, the analytical diffusion model outputs the predicted normalized precipitation bias value. ; For random noise, Represents random noise term Follows a standard normal distribution. For time step Precipitation deviation sample, For time step Precipitation deviation samples; Step S5.3: By performing inverse Z-fractional normalization, the predicted normalized precipitation bias value is... To convert back to the original scale, the formula is: ； In the formula, This represents the precipitation deviation value at the original scale. and These are the mean and standard deviation of precipitation deviations; Super-resolution precipitation is obtained by superimposing the original-scale precipitation bias values ​​onto the bicubic interpolated low-resolution precipitation field, as shown in the formula: ； in, This is a low-resolution precipitation field after bicubic interpolation. This represents the precipitation deviation value at the original scale. Super-resolution precipitation; Step S5.4: Perform threshold correction on the super-resolution precipitation, as follows: ； in, Threshold-corrected super-resolution precipitation It is a function for maximizing the value.

Images