High-resolution rainfall ensemble forecasting method, system, equipment and medium

Abstract

The invention discloses a high-resolution rainfall ensemble forecasting method, system and equipment and a medium. The method comprises the following steps: acquiring a historical atmosphere reanalysis data set and a multi-source fusion rainfall data set; the constructed deterministic neural network learning model is trained to output target duration accumulated average rainfall; training a probability precipitation model constructed by a variational auto-encoder model and a denoising diffusion probability model; acquiring meteorological data at a to-be-predicted moment, and inputting the meteorological data into the deterministic rainfall model to obtain an accumulated average rainfall prediction result within a target duration; sampling from Gaussian distribution to obtain noise, and performing denoising processing in a hidden space by using the trained denoising diffusion model; the denoising result is restored by the trained variational auto-encoder model and added with the output of the deterministic model to obtain the accumulated rainfall of the target time length; and constructing ensemble forecast based on the multiple groups of target duration accumulated rainfall forecast. According to the method, high-resolution rainfall distribution can be generated, the extreme rainfall prediction accuracy is improved, and the uncertainty of rainfall is evaluated.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of high-resolution precipitation prediction, and in particular to a high-resolution precipitation ensemble prediction method, system, device and medium. BACKGROUND

[0002] Currently, the horizontal resolution of the traditional numerical weather prediction (NWP) model is 10-25km. Due to the high complexity and randomness of small-scale atmospheric processes, the traditional prediction method is difficult to explicitly analyze small-scale convection and microphysical processes, and needs to use a simplified mathematical formula to represent the physical processes of the sub-grid scale through a physical parameterization scheme. However, such a parameterization method is difficult to accurately describe the actual atmospheric behavior, and is prone to introduce systematic errors, which limits the prediction effect of the NWP model in complex weather processes such as severe convective precipitation. Although increasing the resolution of the NWP can help to enhance the resolution of the precipitation process, the "gray scale" problem and the sharp increase in computing cost make high-resolution numerical simulation face a bottleneck. Therefore, achieving high-resolution precipitation prediction is still a major challenge in the field of weather prediction.

[0003] With the development of deep learning, data-driven prediction methods have shown great potential in weather prediction. Deep neural network models such as Huawei Pangu-Weather, Google GraphCast and FuXi can perform high-resolution weather prediction at a relatively low computing cost. These models use 39 years of global atmospheric reanalysis data (ERA5) reanalysis dataset, and combine neural network architectures such as self-attention mechanism (Transformer) and graph neural networks to automatically learn atmospheric patterns from data. Although these models perform well in predicting atmospheric variables such as temperature and wind speed, they still face challenges in precipitation prediction. For example, the above models produce overly smooth precipitation prediction results in long-term prediction, making it difficult to capture small-scale severe convective precipitation processes and resulting in the loss of spatial details of precipitation. In addition, the above single deterministic deep learning model is usually difficult to handle the uncertainty in predicting precipitation processes, and the prediction of extreme precipitation events is not accurate enough. SUMMARY

[0004] The present application provides a high-resolution precipitation ensemble prediction method, system, device and medium to solve the problem that current deep learning models are not accurate in predicting high-resolution precipitation and are difficult to quantify the uncertainty of the rainfall prediction process.

[0005] The purpose of the present application and the solution to its technical problems are achieved by using the following technical solutions.

[0006] The first aspect of this invention provides a high-resolution precipitation ensemble forecasting method, comprising: acquiring historical atmospheric reanalysis datasets and multi-source fused precipitation datasets for learning, training, and predicting cumulative precipitation over a target duration; constructing a deterministic precipitation model with a three-dimensional spatial attention mechanism neural network architecture, training the model using the atmospheric reanalysis dataset to output the cumulative average precipitation over the target duration; subtracting the precipitation data in the atmospheric reanalysis dataset from the normalized multi-source fused precipitation data to obtain residual precipitation, and using the residual precipitation as the target to train a probabilistic precipitation model constructed from a variational autoencoder model and a denoised diffusion probability model; acquiring meteorological data for the time to be predicted, and inputting the meteorological data... The cumulative average precipitation prediction for the target duration is obtained by inputting the data into a deterministic precipitation model. Noise is then sampled from a Gaussian distribution, and this noise, combined with the average precipitation prediction and relevant meteorological input variables, is used as conditional information. This conditional information is then input into a trained denoising diffusion probability model for multi-step denoising. The final denoised result is then input into the decoder of a trained variational autoencoder for reconstruction, yielding a set of residual precipitation prediction values. By sampling the noise multiple times and repeating the above denoising and reconstruction process, multiple sets of different residual precipitation prediction values ​​are obtained. Each set of residual precipitation prediction results is superimposed with the average precipitation output from the deterministic precipitation model to obtain multiple sets of cumulative precipitation prediction results for the target duration. Based on these prediction results, a high-resolution precipitation ensemble forecast is constructed.

[0007] According to one embodiment of the present invention, acquiring historical atmospheric reanalysis datasets and multi-source fused precipitation datasets for learning, training and predicting cumulative precipitation within a target duration includes: acquiring surface variable data, upper atmospheric variable data and static variable data of the target area in the historical atmospheric reanalysis dataset within a preset target duration.

[0008] According to one embodiment of the present invention, acquiring historical atmospheric reanalysis datasets and multi-source fused precipitation datasets for learning, training and predicting cumulative precipitation within a target time period further includes: creating atmospheric reanalysis datasets for each target time, and performing standardization processing to generate training datasets, validation datasets and test datasets; wherein, the atmospheric reanalysis dataset for each target time includes the atmospheric reanalysis dataset for the previous target time and the atmospheric reanalysis dataset for the current target time.

[0009] According to one embodiment of the present invention, the structural similarity index and mean square error are used as weighting indices of the deterministic precipitation model, and the weighting indices are used as the target loss function for training; training is completed after validation based on the validation dataset, and the weight parameters of the deterministic precipitation model are saved after training.

[0010] According to one embodiment of the present invention, the residual precipitation is obtained by subtracting the normalized multi-source fused precipitation data from the precipitation data in the atmospheric reanalysis dataset. The training of a variational autoencoder model and a denoised diffusion probability model using the residual precipitation as the target includes: interpolating the cumulative precipitation within the target time period of the atmospheric reanalysis dataset and the multi-source fused precipitation dataset to obtain the cumulative precipitation within the target time period of the same grid; and subtracting the cumulative precipitation within the target time period of the multi-source fused precipitation dataset and the cumulative precipitation within the target time period of the atmospheric reanalysis dataset after normalization to obtain the residual precipitation.

[0011] According to one embodiment of the present invention, the residual precipitation is encoded and trained based on the constructed variational autoencoder model in the encoding stage, the residual precipitation is compressed into the latent space, and the output is reconstructed in the decoding stage. The parameters of the variational autoencoder model are optimized through joint training. After the variational autoencoder model is trained, the parameters of the encoder and decoder of the trained variational autoencoder model are fixed for processing subsequent residual precipitation.

[0012] According to one embodiment of the present invention, residual precipitation is compressed into the latent space based on the trained variational autoencoder model, and the constructed diffusion denoising probability model is trained to generate residual precipitation; wherein, the diffusion denoising probability model is trained with atmospheric conditions and cumulative average precipitation within the target duration in the historical atmospheric reanalysis dataset as training constraints, and the model parameters of the diffusion denoising probability model are saved.

[0013] According to one embodiment of the present invention, the diffusion denoising probability model generates the denoising result in the latent space, which is then restored to the pixel space by the trained variational autoencoder model to obtain residual precipitation.

[0014] According to one embodiment of the present invention, a predetermined time for prediction is set, and a standardized meteorological dataset is input into a trained deterministic precipitation model to obtain the cumulative average precipitation at a predetermined resolution within the predicted target duration. Sampled noise, the cumulative average precipitation, and meteorological data are input into a diffusion denoising probability model for multi-step denoising processing to generate the predicted residual precipitation in the latent space. The predicted residual precipitation is restored to the pixel space based on a variational autoencoder model to obtain a set of residual precipitation prediction values. By sampling the noise multiple times and repeating the above denoising and restoration, multiple sets of different residual precipitation prediction values ​​are obtained. Each set of residual precipitation prediction results is superimposed with the predicted cumulative average precipitation at a predetermined resolution within the target duration to obtain multiple sets of cumulative precipitation predictions at a predetermined resolution within the target duration. An ensemble forecast of rainfall within the target duration is constructed based on these multiple sets of cumulative precipitation predictions within the target duration. The diffusion denoising probability model is a fast denoising diffusion probability model.

[0015] A second aspect of this invention provides a high-resolution precipitation ensemble forecasting system, comprising: an acquisition unit for acquiring historical atmospheric reanalysis datasets and multi-source fused precipitation datasets, used for training and predicting cumulative precipitation over a target duration; a training unit for constructing a deterministic precipitation model with a neural network architecture based on a three-dimensional spatial attention mechanism, training the model using the atmospheric reanalysis dataset to output the cumulative average precipitation over the target duration; obtaining residual precipitation by subtracting the precipitation data in the atmospheric reanalysis dataset from the normalized multi-source fused precipitation data, and using the residual precipitation as the target to train a probabilistic precipitation model constructed from a variational autoencoder model and a denoised diffusion probability model; and a forecasting unit for acquiring meteorological data for the time to be predicted. Meteorological data is input into a deterministic precipitation model to obtain the cumulative average precipitation prediction result within the target duration. Noise is then sampled from a Gaussian distribution. This noise, combined with the average precipitation prediction result and relevant meteorological input variables, is used as conditional information. This conditional information is then input into a trained denoising diffusion probability model for multi-step denoising. The final denoised result is input into the decoder of a trained variational autoencoder for reconstruction, resulting in a set of residual precipitation prediction values. By sampling the noise multiple times and repeating the above denoising and reconstruction, multiple sets of different residual precipitation prediction values ​​are obtained. Each set of residual precipitation prediction results is superimposed with the average precipitation output by the deterministic precipitation model to obtain multiple sets of cumulative precipitation prediction results within the target duration, thus constructing a high-resolution precipitation ensemble forecast.

[0016] A third aspect of the present invention provides an electronic device comprising: a memory for storing non-transitory computer-readable instructions; and a processor for executing the computer-readable instructions such that when the computer-readable instructions are executed by the processor, the above-described high-resolution precipitation ensemble forecasting method is implemented.

[0017] A fourth aspect of the present invention provides a computer-readable storage medium comprising: computer instructions that, when executed on a device, cause the device to perform the aforementioned high-resolution precipitation ensemble forecasting method.

[0018] This invention has significant advantages and beneficial effects compared with the prior art. Through the above technical solution, this invention has at least one of the following advantages and beneficial effects:

[0019] I. This invention provides a high-resolution precipitation ensemble forecasting method. This method obtains high-resolution residual precipitation through a probabilistic precipitation model composed of a variational autoencoder model (VAE) and a denoised diffusion probability model (DDPM). By superimposing the high-resolution residual precipitation with the predicted cumulative average precipitation, a high-resolution precipitation forecast is obtained. On the one hand, this improves the ability to express high-resolution atmospheric spatial details and enhances the ability to capture extreme precipitation conditions. On the other hand, by obtaining high-resolution residual precipitation through a probabilistic precipitation model, the uncertainty of the precipitation process is quantified. This allows the generation of ensemble forecasts to be guided by large-scale atmospheric conditions and deterministic precipitation as constraints.

[0020] II. This invention provides a high-resolution precipitation ensemble forecasting system. The system acquires a training dataset through its acquisition unit, and trains a deterministic precipitation model with a three-dimensional spatial attention mechanism neural network architecture to obtain cumulative average precipitation. It also trains a probabilistic precipitation model composed of a variational autoencoder model and a denoised diffusion probability model to obtain high-resolution residual precipitation. The forecasting unit then superimposes the cumulative average precipitation predicted by the deterministic precipitation model and the residual precipitation predicted by the probabilistic precipitation model to obtain a high-resolution precipitation forecast. In use, this system enhances the ability to represent atmospheric spatial details at high resolution and improves the ability to capture extreme precipitation conditions. Furthermore, by obtaining multiple sets of high-resolution residual precipitation data through the probabilistic precipitation model, it quantifies the uncertainty of the precipitation process. This allows the system to use large-scale atmospheric conditions and deterministic precipitation as constraints to guide the probabilistic precipitation model in generating residual precipitation, ultimately constructing a precipitation ensemble forecast.

[0021] III. The present invention provides an electronic device that, by inputting a dataset and executing non-transitory computer-readable instructions stored in the memory of a computer electronic device through a processor, enables the implementation of the aforementioned high-resolution precipitation ensemble forecasting method when the computer-readable instructions are executed by the processor, thereby obtaining high-resolution precipitation data. In use, this electronic device of the present invention, on the one hand, enhances the ability to express high-resolution atmospheric spatial details and improves the ability to capture extreme precipitation conditions; on the other hand, it quantifies the uncertainty of the precipitation process by obtaining high-resolution residual precipitation through a probabilistic precipitation model, realizing the use of large-scale atmospheric conditions and deterministic precipitation as constraints to guide the probabilistic precipitation model to generate residual precipitation, ultimately constructing a precipitation ensemble forecast.

[0022] IV. The present invention provides a computer-readable storage medium comprising computer instructions. When the computer instructions are executed on a device, the device performs the aforementioned high-resolution precipitation ensemble forecasting method and obtains high-resolution precipitation data. When this computer-readable storage medium is in operation, it enhances the ability to express high-resolution atmospheric spatial details and capture extreme precipitation conditions. Furthermore, by obtaining high-resolution residual precipitation through a probabilistic precipitation model, it quantifies the uncertainty of the precipitation process, achieving the use of large-scale atmospheric conditions and deterministic precipitation as constraints to guide the probabilistic precipitation model in generating residual precipitation, ultimately constructing a precipitation ensemble forecast.

[0023] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described in detail below with reference to the accompanying drawings. Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating a high-resolution precipitation ensemble forecasting method based on deep learning according to an embodiment of the present invention.

[0025] Figure 2 This is a schematic diagram of the framework structure of a high-resolution precipitation ensemble forecasting model according to an embodiment of the present invention.

[0026] Figure 3 This is a schematic diagram of the deep learning training process for a high-resolution precipitation ensemble forecasting model according to an embodiment of the present invention.

[0027] Figure 4 This is a schematic diagram of the precipitation inference forecast process of a high-resolution precipitation ensemble forecast model according to an embodiment of the present invention.

[0028] Figure 5 This is a schematic diagram of a high-resolution precipitation ensemble forecasting system according to an embodiment of the present invention.

[0029] Figure 6 This is a schematic diagram of the structure of an electronic device for high-resolution precipitation ensemble forecasting according to an embodiment of the present invention.

[0030] Figure 7 This is a schematic diagram of the cumulative precipitation distribution of a multi-source fused precipitation dataset within a target duration, according to an embodiment of the present invention.

[0031] Figure 8 This is a schematic diagram of the cumulative forecast precipitation distribution over a target duration according to an embodiment of the present invention.

[0032] Figure 9 This is another set of schematic diagrams showing the cumulative forecast precipitation distribution over a target duration according to an embodiment of the present invention.

[0033] Figure 10 This is another set of schematic diagrams showing the cumulative forecast precipitation distribution over a target duration according to an embodiment of the present invention.

[0034] Figure 11 This is another set of schematic diagrams showing the cumulative forecast precipitation distribution over a target duration according to an embodiment of the present invention.

[0035] Figure 12 This is another set of schematic diagrams showing the cumulative forecast precipitation distribution over a target duration according to an embodiment of the present invention. Detailed Implementation

[0036] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0037] This invention provides a high-resolution precipitation ensemble forecasting method, system, equipment, and medium, such as... Figure 1 As shown, this high-resolution precipitation ensemble forecasting method includes the following process:

[0038] S1 acquires historical atmospheric reanalysis datasets and multi-source fused precipitation datasets for learning, training, and predicting cumulative precipitation over the target time period.

[0039] It should be noted that atmospheric reanalysis datasets can be selected from sources such as the Global Atmospheric Reanalysis Data (ECMWFReanalysis v5, ERA5), and multi-source fused precipitation datasets can be selected from sources such as the China Merged Precipitation Analysis (CMPA) dataset. In addition to the above datasets, datasets composed of meteorological data from different sources can also be used to obtain the cumulative precipitation dataset for the training and prediction target time period. These can include data from at least one meteorological analysis source among different surface regions, different upper-air atmospheric environments, and different surface static variables, or a combination of multiple meteorological analysis data sources. Data standardization can be performed by standardizing data of each category with the same name within the same region, one category at a time. Alternatively, the selected input data can be standardized in a custom manner. The target time period can be one hour.

[0040] S2. Construct a deterministic precipitation model with a neural network architecture based on a three-dimensional spatial attention mechanism. Train the model using an atmospheric reanalysis dataset to output the cumulative average precipitation over the target duration after training. Subtract the precipitation data in the atmospheric reanalysis dataset from the normalized multi-source fused precipitation data to obtain residual precipitation. Use this residual precipitation as the target to train a probabilistic precipitation model constructed from a variational autoencoder model and a denoising diffusion probability model.

[0041] It should be noted that the deterministic precipitation model using a 3D spatial attention mechanism neural network architecture can be a deterministic precipitation model with an architecture such as the 3D Swin Transformer. The Variational Auto-Encoder (VAE) model is not limited to a specific network structure; its encoder and decoder parts can be constructed using various units such as Convolutional Neural Networks (CNN), Attention mechanisms, and ResNet. In the embodiments of this invention, the U-Net network architecture, which integrates the above structural features, is used. The Variational Auto-Encoder performs learning and training for encoding and decoding residual precipitation, as well as post-training encoding and decoding processing. The Denoising Diffusion Probability Model (DDPM) is used to learn and train to generate high-resolution residual precipitation data and predict high-resolution residual precipitation data within the target duration in subsequent ensemble forecasts. In addition to the aforementioned Denoising Diffusion Probabilistic Model (DDPM), a fast denoising model (Denoising Diffusion Implicit M Models, DDIM) can also be used. DDIM can achieve more efficient training and prediction of precipitation sample data through a fast sampling process and by introducing a non-Markov chain diffusion process. It can also perform deep learning training tasks on precipitation images with a neural network architecture that combines a three-dimensional spatial attention mechanism. After training, it can predict residual precipitation within the target duration in a high-efficiency, stable and high-resolution form, and generate high-resolution cumulative precipitation within the target duration by combining the cumulative average precipitation within the target duration.

[0042] It should be further explained that the above-mentioned probabilistic precipitation model can be constructed by combining a variational autoencoder (VAE) and a denoised diffusion probability model (DDPM). In the embodiments of the present invention, the VAE is first trained, the parameters of the VAE are fixed, and then the latent vector of residual precipitation is obtained using the VAE model. The DDPM model is then trained in the latent space.

[0043] S3. Acquire meteorological data for the time to be predicted and input it into a deterministic precipitation model to obtain the cumulative average precipitation prediction result for the target duration. Then, sample noise from a Gaussian distribution and combine the noise with the average precipitation prediction result and relevant meteorological input variables as conditional information. Input the conditional information into a trained denoising diffusion probability model for multi-step denoising. Input the final denoising result into the decoder of a trained variational autoencoder for restoration to obtain a set of residual precipitation prediction values. By sampling the noise multiple times and repeating the above denoising and restoration process, multiple sets of different residual precipitation prediction values ​​are obtained. The residual precipitation prediction results of each set are superimposed with the average precipitation output of the deterministic precipitation model to obtain multiple sets of cumulative precipitation prediction results for the target duration. Based on the prediction results, a high-resolution precipitation ensemble forecast is constructed.

[0044] It should be noted that after the VAE and DDPM models are trained, the residual precipitation within the target time period is predicted using the trained VAE and DDPM models.

[0045] This invention decomposes the cumulative precipitation over a high-resolution predicted target time period into cumulative average precipitation over a mesoscale grid. And uncertain precipitation TP′ at the convective scale. Both types of precipitation are used as deterministic precipitation models by constructing a three-dimensional spatial attention mechanism neural network architecture, and a high-resolution probabilistic precipitation model is constructed at the same time. The two are combined to simulate and predict the cumulative precipitation TP within the target duration, as shown in Equation (1):

[0046]

[0047] In this invention, such as Figure 2 As shown, solid lines with arrows represent the training process, and dashed lines with arrows represent the prediction inference process. In the prediction model framework for cumulative precipitation (TP) within the target duration, the deterministic precipitation model uses surface variables (t2m, mslp, u10, v10) and upper atmospheric variables (t, z, u, v, q, 13 pressure layers) at 0.25-degree resolution from the atmospheric reanalysis dataset (ECMWF Reanalysis v5, ERA5), as well as static variables (lsm, soil type, z0, lat, lon) and time variables within the preset target duration, such as the number of hours in a day and the number of days in a year, as inputs to the deterministic precipitation model. The cumulative precipitation within the preset target duration of ERA5 serves as the training objective, allowing the construction of a deterministic precipitation model (3D Swin Transformer) with a neural network architecture based on a three-dimensional spatial attention mechanism. The specific meteorological variables required as inputs to the deterministic precipitation model are shown in Table 1.

[0048] Table 1 shows the training dataset used.

[0049]

[0050] Note: The variable names in Table 1 are specifically represented as follows: t2m represents the air temperature at 2m above the ground, mslp represents the mean sea level pressure, u10 and v10 represent the zonal and meridional wind speeds at 10m above the ground, respectively, tp represents the total surface precipitation; t represents the air temperature of each pressure layer, z represents the geopotential of each pressure layer, q represents the specific humidity of each pressure layer, u and v represent the zonal and meridional wind speeds of each pressure layer, respectively; lsm represents the land-sea distribution, soil type represents the soil type, z0 represents the surface geopotential, lat represents the latitude, and lon represents the longitude; hour of day represents the number of hours in a day, and day of year represents the number of days in a year; among them, the upper atmospheric variable data are pressure layer variable data, and the selectable pressure layers are 50, 100, 150, 200, 250, 300, 400, 500, 600, 700, 850, 925 and 1000 hPa.

[0051] It should be further noted that the input variable data for the aforementioned deterministic precipitation model can include surface variable data, upper-air atmospheric variable data, and a subset of static variable data. For example, selecting fewer types of variable data can improve the efficiency of deterministic precipitation model training. For instance, using only upper-air variables and precipitation at 200hPa, 500hPa, 700hPa, 850hPa, and 925hPa for training can quickly establish a general gradient framework for deterministic precipitation model training.

[0052] The probabilistic precipitation model uses residual precipitation—namely, high-resolution precipitation from the Multi-Source Fusion Precipitation Dataset (CMPA) and ERA5 precipitation (interpolated to the same grid using averaging and nearest neighbor interpolation methods, respectively)—as the training target after normalization. First, a VAE is trained to compress the residual precipitation, reducing the computational burden on training the Denoising Diffusion Probabilistic Model (DDPM). Then, atmospheric field and precipitation at a 0.25-degree resolution are used as constraints for denoising training of the DDPM model within the latent space. This invention introduces the Denoising Diffusion Probabilistic Model (DDPM) applied to image generation as a probabilistic model for predicting precipitation, generating high-quality, clear samples of uncertain precipitation TP′ at the convective scale through an inverse denoising process. This denoising diffusion probabilistic model framework has unique advantages in capturing the probability distribution of complex uncertain precipitation TP′, and is particularly suitable for handling small-scale phenomena with high uncertainty. In the field of weather forecasting, the Denoising Diffusion Probability Model (DDPM) is applied to precipitation forecasting. This type of model can capture the development and changes of uncertain precipitation TP′ in convective systems within a preset target duration (such as a target duration with different hourly cycles throughout the day) and provide ensemble forecasts to quantify the uncertainty. Compared with traditional deep learning precipitation models, this method can better capture the randomness in the precipitation process and provide more accurate forecasts of extreme precipitation events. The high-resolution precipitation ensemble forecast model proposed in this invention combines the advantages of deterministic cumulative average precipitation models and conditional probabilistic convective precipitation models for high-resolution precipitation prediction. The deterministic precipitation model is responsible for capturing the evolution of mesoscale atmospheric precipitation, while the Denoising Diffusion Probability Model (DDPM) handles the uncertain precipitation forecasts related to small-scale convective precipitation. This invention uses basic meteorological variables such as temperature, pressure, humidity, and wind as input variables to generate high-resolution precipitation ensemble forecasts, better addressing the uncertainty problem of extreme precipitation events.

[0053] In an optional embodiment of the present invention, the specific implementation process for training the deterministic precipitation model is as follows: Figures 2 to 4 The training steps shown are as follows:

[0054] In this invention, acquiring historical atmospheric reanalysis datasets and multi-source fusion precipitation datasets for learning and training cumulative precipitation within the target duration includes: acquiring surface variable data, upper atmospheric variable data, and static variable data of the target area in the atmospheric reanalysis dataset within the preset target duration; wherein, the surface variable data includes at least one or more types of surface variable data such as near-surface temperature, mean sea level pressure, near-surface zonal wind speed, and near-surface meridional wind speed; the upper atmospheric variable data includes at least one or more types of upper atmospheric variable data such as temperature of each pressure layer in the preset upper atmospheric pressure layer, geopotential of each pressure layer, zonal wind speed of each pressure layer, meridional wind speed of each pressure layer, and specific humidity of each pressure layer; the static variable data includes at least one or more types of static variable data such as land-sea distribution data, soil type data, surface geopotential data, latitude data, and longitude data. Specifically, (1) acquiring the atmospheric reanalysis dataset (ERA5). For example, the atmospheric reanalysis dataset (ERA5) for the period 2010 to 2021 can be obtained from the European Centre for Medium-Range Weather Forecasts (ECMWF) website. To ensure that the obtained atmospheric reanalysis dataset covers the target area (such as China), the precipitation resolution for the target duration can be selected as 1 hour. The downloaded input variable data includes surface variables, upper-air variables, and static variables as listed in Table 1. In this invention, obtaining the standardized atmospheric reanalysis dataset and the multi-source fused precipitation dataset for learning and training the cumulative precipitation within the target duration also includes: creating an atmospheric reanalysis dataset for each target time and performing standardization processing to generate a training dataset, a validation dataset, and a test dataset; wherein, the atmospheric reanalysis dataset for each target time includes the atmospheric reanalysis dataset for the previous target time and the atmospheric reanalysis dataset for the current target time. Specifically, as in step (2), before training the deterministic precipitation model, it is necessary to preprocess the above-downloaded input variable data and create the input training dataset. For example, create an input dataset for training ERA5 containing the previous time and the current time for each target time. These input data include surface variable data, upper atmospheric variable data, and static variable data at the current moment. The training objective is to predict the cumulative precipitation within the corresponding 1 hour at the target moment (i.e., the total cumulative precipitation within 1 hour from the target moment). Thus, the ERA5 dataset can be divided into training dataset (2010 to 2019), validation dataset (2020), and test dataset (2021) in chronological order. (3) Construct variable data with ERA5 variable data (such as temperature, humidity, wind speed, etc.) as input for training, set the target duration of training output (i.e., output the cumulative precipitation within 1 hour after the current moment), and train the constructed deterministic precipitation model.The deterministic precipitation model can be a deterministic precipitation model with a three-dimensional spatial attention mechanism neural network architecture, which is a three-dimensional spatial attention mechanism neural network architecture (3D Swin Transformer). Before training, the input variable data and the target variable data of the training are standardized to ensure the reliability and stability of the training data and training structure. At the same time, mean squared error (MSE) and structural similarity index (SSIM) are used as weighting indices of the deterministic precipitation model, and their weighting indices are used as the target loss function to train the deterministic precipitation model with a three-dimensional spatial attention mechanism neural network architecture for deep learning. Among them, MSE is used to measure pixel-level error, and SSIM is more sensitive to changes in brightness, contrast and texture, so it is more effective in preserving the spatial distribution and intensity of features such as heavy precipitation. The specific formulas of mean squared error (MSE), structural similarity index (SSIM) and target loss function are shown in equations (2) to (4) below:

[0055]

[0056]

[0057] Loss = w1MSE + w2(1-SSIM) (4)

[0058] Where n is the number of sample points, x i It predicts precipitation, y i It is the target precipitation, where i is a positive integer ranging from 1 to n, and μ x and μ y It is the average of the predicted precipitation and the target precipitation. and σ is the variance between predicted precipitation and target precipitation. xy The covariances between the predicted and target precipitation are C1 = 0.01, C2 = 0.04, C3 = 0.02, and w1 and w2 = 0.5 and 1.5, respectively.

[0059] (4) Saving the deterministic precipitation model of the neural network architecture with three-dimensional spatial attention mechanism after training. Specifically, in this embodiment of the invention, an early stopping strategy is adopted, that is, when the target loss function on the validation set does not show a decreasing trend in 8 consecutive evaluations, the model is determined to have converged and training is terminated. After the deterministic precipitation model is validated on the validation set, the weight parameters of the deterministic precipitation model are saved; otherwise, the training of the deterministic precipitation model continues until the validation set validation meets the validation requirements, and then the weight parameters of the deterministic precipitation model are saved. This deterministic precipitation model can directly predict the cumulative precipitation (resolution of 0.25 degrees) of the current target time within the next target duration (1 hour) based on the ERA5 input data of the previous and current time of each target time.

[0060] In an optional embodiment of the present invention, the specific implementation process for training the probabilistic precipitation model is as follows: Figures 2 to 4 The training steps shown are as follows:

[0061] (1) Obtain the Multi-Source Fusion Precipitation Dataset (CMPA). The original resolution of this training dataset is 0.01 degrees, which can be interpolated to 0.05 degrees for training and as a reference for real precipitation (e.g., verifying the cumulative precipitation at the target time within the target duration in the dataset). (2) Create a probabilistic precipitation model training dataset. Interpolate the 0.25 resolution precipitation data from the Atmospheric Reanalysis dataset to the same 0.05 degree resolution data grid as the Multi-Source Fusion Precipitation Dataset (CMPA) using the nearest neighbor interpolation method. After grid alignment, normalize the interpolated Atmospheric Reanalysis dataset and the Multi-Source Fusion Precipitation Dataset respectively, and subtract the normalized CMPA precipitation from the normalized ERA5 precipitation to obtain the residual precipitation for training. (3) Train a variational autoencoder model such as a variational autoencoder (VAE). Based on the variational autoencoder model, the encoder maps the residual precipitation between the high-resolution CMPA precipitation and ERA5 precipitation into the latent space. Specifically, the residual precipitation is compressed from the pixel space to a low-dimensional latent space during the encoding stage of the encoder using VAE for training. After training, the parameters of the trained VAE encoder and decoder are fixed for use in the subsequent training, verification, and testing stages of the denoising diffusion probability model (DDPM) to process the residual precipitation. (4) Training the denoising diffusion probability model (DDPM). Based on the VAE encoder trained in the previous step, the precipitation residuals obtained from CMPA precipitation and ERA5 precipitation are compressed into the latent space, and the diffusion denoising probability model (DDPM) is trained to generate high-resolution forecasts of precipitation residuals. The diffusion denoising probability model is trained with atmospheric conditions and cumulative average precipitation within the target duration in the historical atmospheric reanalysis dataset as training constraints, and the model parameters of the diffusion denoising probability model are saved. The probabilistic precipitation model takes ERA5 surface variable data (t2m, mslp, u10, v10) and upper atmospheric variables (t, z, u, v, q, 13 pressure layers) as input, and ERA5 0.25 degree resolution precipitation as constraints. It generates high-resolution residual precipitation through stepwise denoising (the high-resolution residual precipitation generated after denoising is also called residual precipitation). (5) Reconstruction of residual precipitation. The denoised results in the latent space are generated according to the DDPM model. (6) Implement an early stopping strategy, that is, when the target loss function on the validation set does not show a decreasing trend in 8 consecutive evaluations, the model is judged to have converged and training is terminated. The model parameters of the DDPM model are saved for subsequent probabilistic precipitation models to predict uncertain convective precipitation.

[0062] In an optional embodiment of the invention, high-resolution precipitation ensemble forecasting includes using a deterministic precipitation model to forecast the cumulative average precipitation over a target duration at a target time (e.g., average precipitation over a mesoscale grid). The document describes the prediction of cumulative convective precipitation (such as uncertain precipitation TP′ at the convective scale) within a target duration at a target time using a probabilistic precipitation model, and the specific implementation process of ensemble forecasting of high-resolution precipitation generated by superimposing the above two types of precipitation over multiple target durations. Figures 2 to 4 The prediction steps shown are as follows:

[0063] (1) Preprocessing of the input data for prediction. The input variable data required for precipitation prediction are standardized respectively. The input variable data required for precipitation prediction can be the data in the test dataset mentioned above. The test dataset includes the atmospheric reanalysis dataset before prediction (such as data from the atmospheric reanalysis dataset) or the predicted atmospheric dataset (such as atmospheric data predicted by the forecast model).

[0064] (2) Predicting deterministic precipitation. Input the input variable data required for precipitation prediction into a trained deterministic model (such as a deterministic precipitation model with a neural network architecture based on three-dimensional spatial attention mechanism) to obtain a deterministic precipitation forecast with a resolution of 0.25 degrees.

[0065] (3) Predicting uncertain precipitation. Based on the probabilistic precipitation model, noise is generated by sampling. The generated noise, temperature, pressure, humidity and wind variables, and the cumulative precipitation within the predicted target time of the deterministic precipitation model (a neural network architecture with a three-dimensional spatial attention mechanism) are used as constraints for the DDPM model. The sampling schedule of the fast denoising diffusion probabilistic model (DDIM) is configured as 300 steps to generate residual precipitation in the latent space.

[0066] (4) Predicting uncertain precipitation. The residual precipitation in the latent space is input into the decoder of the VAE model, and the decoder based on the VAE model recovers the predicted residual precipitation into the pixel space.

[0067] (5) Overlaying the average precipitation and the predicted residual precipitation. The predicted residual precipitation within the target duration at the predicted target time is overlaid with the cumulative average precipitation output from the deterministic precipitation forecast to obtain the final high-resolution cumulative precipitation forecast within the target duration at the predicted target time. It should be noted that the target duration at the above target time corresponds to the hourly duration of the corresponding hour within a day.

[0068] (6) Ensemble forecast. An ensemble forecast of precipitation within the target duration is constructed based on multiple sets of cumulative precipitation forecasts within the target duration. Specifically, by repeating steps 3 to 5 above, multiple sets of cumulative precipitation forecasts within the target duration based on the precipitation forecast model can be obtained, and a high-resolution precipitation ensemble forecast can be constructed through multiple sets of cumulative precipitation forecasts within the target duration.

[0069] This invention also provides a high-resolution precipitation ensemble forecasting system, such as...Figure 5 As shown, the high-resolution precipitation ensemble forecasting system 200 includes: an acquisition unit 210, used to acquire standardized atmospheric reanalysis datasets and multi-source fused precipitation datasets for training and predicting cumulative precipitation over a target duration; a training unit 220, used to construct a deterministic precipitation model with a neural network architecture based on a three-dimensional spatial attention mechanism, training the deterministic precipitation model using the atmospheric reanalysis dataset and outputting the trained cumulative average precipitation over the target duration; residual precipitation is obtained by subtracting the atmospheric reanalysis precipitation data from the normalized multi-source fused precipitation datasets, and the residual precipitation is used as the target to train a probabilistic precipitation model constructed by a variational autoencoder model (VAE) and a denoised diffusion probability model (DDPM); and a forecasting unit 230, used to acquire the time to be predicted. Meteorological data is input into a deterministic precipitation model to obtain the cumulative average precipitation prediction result within the target duration. Noise is then sampled from a Gaussian distribution, and this noise, combined with the average precipitation prediction result and relevant meteorological input variables, is used as conditional information. This conditional information is then input into a trained denoising diffusion probability model for multi-step denoising. The final denoised result is input into the decoder of a trained variational autoencoder for restoration, yielding a set of residual precipitation prediction values. By sampling the noise multiple times and repeating the above denoising and restoration process, multiple sets of different residual precipitation prediction values ​​are obtained. Each set of residual precipitation prediction results is superimposed with the average precipitation output of the deterministic precipitation model to obtain multiple sets of cumulative precipitation prediction results within the target duration. Based on these prediction results, a high-resolution precipitation ensemble forecast is constructed.

[0070] The present invention also provides an electronic device, such as Figure 6 As shown, the electronic device 300 includes: a memory 310 for storing non-transitory computer-readable instructions 330; and a processor 320 for executing the computer-readable instructions 330, such that when the computer-readable instructions 330 are executed by the processor 320, the above-described high-resolution precipitation ensemble forecasting method is implemented.

[0071] The present invention also provides a computer-readable storage medium comprising computer instructions that, when executed on a device, cause the device to perform the above-described high-resolution precipitation ensemble forecasting method.

[0072] like Figures 7 to 12 As shown, this is a schematic diagram of the cumulative precipitation in the multi-source fused precipitation dataset within the target time and target duration (e.g., Figure 7 ),in, Figure 7 This is the hourly cumulative precipitation dataset for precipitation fusion as of 01:00 UTC on August 1, 2021. Figures 8 to 12 The diagram shown illustrates the cumulative precipitation over the target time period predicted by multiple forecasts (such as forecast members 1 to 5). Figure 8It can provide a high-resolution precipitation forecast for member 1 as of 01:00 UTC on August 1, 2021. Figure 9 This can provide a high-resolution precipitation forecast for Member 2 as of 01:00 UTC on August 1, 2021. Figure 10 It can provide high-resolution precipitation forecasts for member 3 as of 01:00 UTC on August 1, 2021. Figure 11 This can provide a high-resolution precipitation forecast for member 4 as of 01:00 UTC on August 1, 2021. Figure 12 This can provide a high-resolution precipitation forecast for member 5 as of 01:00 UTC on August 1, 2021. From the above... Figure 7 Comparing the cumulative precipitation within the target duration shown by forecast members 1 to 5, it can be seen that... Figures 8 to 12 The forecast members 1 to 5 are shown to generate detailed, high-resolution precipitation forecasts and form an ensemble forecast.

[0073] In the above-described implementation process of the present invention, Figures 8 to 12 The multiple forecasts shown can be constructed into a high-resolution ensemble forecast. By statistically analyzing the results of multiple ensemble forecasts, the probability, confidence interval, and distribution range of precipitation exceeding the threshold in a specific area can be assessed, thereby improving the ability to predict extreme precipitation events and achieving effective assessment of the uncertainty of precipitation processes.

[0074] The high-resolution precipitation ensemble forecasting method of this invention differs from traditional numerical precipitation models. Unlike traditional precipitation models, this method does not use complex modules of precipitation physics parameterization schemes to describe precipitation processes. Instead, it directly employs a deterministic precipitation model. Based on a neural network architecture with a three-dimensional spatial attention mechanism, it learns, predicts, and diagnoses average precipitation under current atmospheric conditions. By applying SSIM and MSE weighted loss functions, it improves the deterministic model's ability to forecast heavy precipitation, significantly enhancing its ability to capture spatial details and localized heavy precipitation events, thus improving forecast accuracy. Furthermore, by introducing a denoising diffusion model as a probabilistic precipitation model, this invention not only generates precipitation with clear textures but also, as an uncertain module, learns and trains on random precipitation at the convective scale and generates high-resolution precipitation ensemble forecasts through multiple sampling predictions. This invention utilizes a combined architecture of deterministic and probabilistic precipitation models, employing a denoising and diffusion model guided by large-scale atmospheric field variable data and precipitation data as target conditions. This enables the generation of multiple precipitation forecasts and quantifies the uncertainties in the precipitation process.

[0075] It should be noted that any process or method description in the flowchart or otherwise described herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of the preferred embodiments of the invention includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order according to the functions involved, as should be understood by those skilled in the art to which the embodiments of the invention pertain.

[0076] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which programs can be printed, because programs can be obtained electronically, for example, by optically scanning the paper or other media, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0077] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0078] Those skilled in the art will understand that all or part of the steps included in the methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0079] Furthermore, the functional units in the various embodiments of the present invention can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be partially implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0080] The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. Although embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present invention.

[0081] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A high-resolution precipitation ensemble forecasting method, characterized in that, include: Historical atmospheric reanalysis datasets and multi-source fused precipitation datasets are acquired for training and prediction of cumulative precipitation over a target time period. A deterministic precipitation model with a three-dimensional spatial attention mechanism neural network architecture is constructed. The model is trained using an atmospheric reanalysis dataset to output the cumulative average precipitation over a target duration. The residual precipitation is obtained by subtracting the precipitation data in the atmospheric reanalysis dataset from the normalized multi-source fused precipitation data. The residual precipitation is then used as the target to train a probabilistic precipitation model constructed from a variational autoencoder model and a denoised diffusion probability model. Meteorological data for the time to be predicted is obtained and input into a deterministic precipitation model to obtain the cumulative average precipitation prediction result within the target duration. Then, noise is sampled from the Gaussian distribution, and the noise is combined with the average precipitation prediction result and relevant meteorological input variables as conditional information. The conditional information is then input into the trained denoising diffusion probability model for multi-step denoising. The final denoising result is input into the decoder of the trained variational autoencoder for restoration, resulting in a set of residual precipitation prediction values. By sampling the noise multiple times and repeating the above denoising and restoration, multiple sets of different residual precipitation prediction values ​​are obtained. The residual precipitation prediction results of each group are superimposed with the average precipitation output of the deterministic precipitation model to obtain multiple cumulative precipitation prediction results within the target time period. Based on the prediction results, a high-resolution precipitation ensemble forecast is constructed.

2. The high-resolution precipitation ensemble forecasting method according to claim 1, characterized in that, Historical atmospheric reanalysis datasets and multi-source fused precipitation datasets were acquired for training and prediction of cumulative precipitation over the target time period, including: Acquire surface variable data, upper atmospheric variable data, and static variable data for the target area in the historical atmospheric reanalysis dataset over a preset target time period.

3. The high-resolution precipitation ensemble forecasting method according to claim 2, characterized in that, Acquiring historical atmospheric reanalysis datasets and multi-source fused precipitation datasets for training and predicting cumulative precipitation over the target time period also includes: Create atmospheric reanalysis datasets for each target time point and perform standardization to generate training, validation, and test datasets; The atmospheric reanalysis dataset for each target time includes the atmospheric reanalysis dataset for the previous target time and the atmospheric reanalysis dataset for the current target time.

4. The high-resolution precipitation ensemble forecasting method according to claim 3, characterized in that, The structural similarity index and mean squared error are used as weighting indices for the deterministic precipitation model, and these weighting indices are used as the target loss function for training. After validation on the validation dataset, the training is completed, and the weight parameters of the deterministic precipitation model are saved after training.

5. The high-resolution precipitation ensemble forecasting method according to claim 4, characterized in that, The residual precipitation is obtained by subtracting the precipitation data from the atmospheric reanalysis dataset after normalization of the multi-source fused precipitation data. This residual precipitation is then used as the target to train a variational autoencoder model and a denoised diffusion probability model, including: The cumulative precipitation within the target duration of the atmospheric reanalysis dataset and the multi-source fusion precipitation dataset are interpolated separately to obtain the cumulative precipitation within the target duration of the same grid. The cumulative precipitation within the target duration of the multi-source fusion precipitation dataset and the cumulative precipitation within the target duration of the atmospheric reanalysis dataset are normalized separately and then subtracted to obtain the residual precipitation.

6. The high-resolution precipitation ensemble forecasting method according to claim 5, characterized in that, The constructed variational autoencoder model encodes and trains residual precipitation during the encoding stage, compresses the residual precipitation into the latent space, and reconstructs the output during the decoding stage. The variational autoencoder model parameters are optimized through joint training. After the variational autoencoder model is trained, the parameters of the encoder and decoder of the trained variational autoencoder model are fixed for subsequent processing of residual precipitation.

7. The high-resolution precipitation ensemble forecasting method according to claim 5, characterized in that, The residual precipitation is compressed into the latent space based on the trained variational autoencoder model, and the constructed diffusion denoising probability model is trained to generate residual precipitation. The diffusion denoising probability model is trained using atmospheric conditions and cumulative average precipitation within the target time period from the historical atmospheric reanalysis dataset as training constraints, and the model parameters of the diffusion denoising probability model are saved.

8. The high-resolution precipitation ensemble forecasting method according to claim 7, characterized in that, The diffusion denoising probability model generates denoising results in the latent space, which are then restored to the pixel space by the trained variational autoencoder model to obtain residual precipitation.

9. The high-resolution precipitation ensemble forecasting method according to claim 8, characterized in that, The standardized meteorological dataset is input into the trained deterministic precipitation model at a preset time to be predicted, and the cumulative average precipitation at a preset resolution within the target time period is obtained. The sampled noise, the cumulative average precipitation, and meteorological data are input into the diffusion denoising probability model, and multi-step denoising processing is performed to generate the predicted residual precipitation in the latent space. The predicted residual precipitation is recovered and restored to the pixel space based on the variational autoencoder model to obtain a set of residual precipitation prediction values; By sampling the noise multiple times and repeating the above denoising and restoration, multiple sets of different residual precipitation prediction values ​​are obtained. The residual precipitation prediction results of each set are superimposed with the cumulative average precipitation at a preset resolution within the predicted target time to obtain multiple sets of cumulative precipitation predictions at a preset resolution within the target time. An ensemble forecast of rainfall within a target duration is constructed based on cumulative precipitation predictions over multiple target durations. Among them, the diffusion denoising probability model is the fast denoising diffusion probability model.

10. A high-resolution precipitation ensemble forecasting system, characterized in that, include: The acquisition unit is used to acquire historical atmospheric reanalysis datasets and multi-source fused precipitation datasets for learning, training and predicting cumulative precipitation over the target time period. The training unit is used to construct a deterministic precipitation model with a neural network architecture based on a three-dimensional spatial attention mechanism. The model is trained using an atmospheric reanalysis dataset to output the cumulative average precipitation over a target time period. The residual precipitation is obtained by subtracting the precipitation data in the atmospheric reanalysis dataset from the normalized multi-source fused precipitation data. The residual precipitation is then used as the target to train a probabilistic precipitation model constructed from a variational autoencoder model and a denoised diffusion probability model. The forecasting unit acquires meteorological data for the time to be predicted, inputs the meteorological data into a deterministic precipitation model, and obtains the cumulative average precipitation prediction result within the target duration. Then, it samples noise from the Gaussian distribution, combines the noise with the average precipitation prediction result and relevant meteorological input variables as conditional information, and inputs the conditional information into the trained denoising diffusion probability model for multi-step denoising. The final denoising result is input into the decoder of the trained variational autoencoder for restoration, resulting in a set of residual precipitation prediction values. By sampling the noise multiple times and repeating the above denoising and restoration, multiple sets of different residual precipitation prediction values ​​are obtained. The residual precipitation prediction results of each group are superimposed with the average precipitation output of the deterministic precipitation model to obtain multiple cumulative precipitation prediction results within the target time period. Based on the prediction results, a high-resolution precipitation ensemble forecast is constructed.

11. An electronic device, characterized in that, include: Memory is used to store non-transitory computer-readable instructions; as well as A processor for running computer-readable instructions such that, when executed by the processor, the computer-readable instructions implement the high-resolution precipitation ensemble forecasting method as described in any one of claims 1 to 9.

12. A computer-readable storage medium, characterized in that, Includes computer instructions that, when executed on the device, cause the device to perform the high-resolution precipitation ensemble forecasting method as described in any one of claims 1 to 9.

Images