Rainfall downscaling method and system based on deep learning network model fusing rainfall priori knowledge

By integrating a deep learning network model that incorporates prior knowledge of precipitation, and combining convolutional neural networks and long short-term memory networks, the problems of uncertainty quantification and physical inconsistency in the precipitation downscaling process in existing technologies have been solved, achieving high-precision precipitation downscaling and uncertainty quantification.

CN121637394APending Publication Date: 2026-03-10YANCHENG INST OF TECH
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202511739756.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-11-25
Publication Date
2026-03-10

AI Technical Summary

Technical Problem

Existing technologies struggle to accurately capture localized heavy precipitation driven by complex terrain during precipitation downscaling and cannot effectively quantify the uncertainties in the precipitation generation process. This results in inconsistencies between downscaling results and atmospheric dynamic processes, making them unsuitable for disaster risk assessment.

Method used

A deep learning network model that integrates prior knowledge of precipitation is adopted, combining convolutional neural networks and long short-term memory networks to construct a precipitation spatiotemporal correlation model. A precipitation loss function with weighted negative log-likelihood is established using the likelihood function of coupled censored data, Box-Cox transformation and Gaussian distribution of time-varying variance. High-resolution precipitation distribution is obtained through Monte Carlo sampling.

Benefits of technology

It achieves precise downscaling of precipitation data, improves downscaling accuracy, and quantifies the uncertainty of precipitation downscaling, thus better reflecting the impact of topographic and surface features on precipitation.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121637394A_ABST
    Figure CN121637394A_ABST
Patent Text Reader

Abstract

The invention discloses a rainfall downscaling method and system based on a deep learning network model fusing rainfall priori knowledge, and the method comprises the steps: firstly collecting the topographic data and low-resolution day-by-day rainfall data of a target region, and taking the data as input data; a short-term high-resolution precipitation field generated in a mesoscale weather forecast WRF mode is used as training truth value data; according to the method, the function of accurately downscaling the rainfall data in combination with the convolutional neural network and the long and short term memory network is realized, the spatial-temporal correlation of rainfall is fully considered in the downscaling process, and meanwhile, a likelihood function combined with coupled censored data, Box-Cox conversion and time variation variance Gaussian distribution is adopted as a rainfall loss function; the method not only can represent zero expansibility, skewness and heterovariance characteristics of rainfall, but also can improve the rainfall downscaling precision and quantify the uncertainty of rainfall downscaling, and is suitable for wide popularization and use.
Need to check novelty before this filing date? Find Prior Art

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of precipitation downscaling, in particular to a precipitation downscaling method and system based on a deep learning network model fusing precipitation prior knowledge. BACKGROUND

[0002] Precipitation downscaling refers to converting large-scale precipitation data with relatively low spatial resolution into high spatial resolution precipitation data using various techniques and methods, so as to better reflect the influence of local factors such as topographic features and surface features on precipitation. High-resolution precipitation data is the core basis for understanding water cycle, managing water resources and warning flood and drought disasters.

[0003] At present, the spatial resolution of the existing mainstream precipitation data is usually greater than 10 km, which cannot capture complex topography-driven local heavy precipitation. Although the kilometer resolution level precipitation forecast can be realized by the mesoscale weather forecast WRF model dynamic downscaling model, it requires a large amount of computing resources and is affected by the uncertainty of the parameterization scheme, making it difficult to be applied in business. At the same time, the relationship between the existing statistical downscaling model assumption factor and precipitation is linear, which makes it difficult to depict the high nonlinearity and spatial correlation in the precipitation generation process. Most of the existing deep learning model architectures lack targeted design for the fusion of meteorological physical variable characteristics, resulting in physical inconsistency between the downscaling results and the atmospheric dynamic process. At the same time, the existing deep learning model generally adopts a deterministic output framework, which can only provide a single prediction value and cannot quantify the prediction uncertainty, which makes it difficult to accurately capture the spatial distribution of extreme precipitation events and cannot provide reliable confidence intervals for disaster risk assessment in the application of precipitation downscaling. Therefore, it is necessary to design a precipitation downscaling method and system based on a deep learning network model fusing precipitation prior knowledge. SUMMARY

[0004] The present application aims to overcome the shortcomings of the prior art and provide a precipitation downscaling method and system based on a deep learning network model fusing precipitation prior knowledge, which realizes the function of accurately downscaling precipitation data by combining convolutional neural networks and long short-term memory networks, and fully considers the spatiotemporal correlation of precipitation in the downscaling process. At the same time, the likelihood function combining coupling censored data, Box-Cox transformation and time-varying variance Gaussian distribution is used as the precipitation loss function, which can not only represent the zero inflation, skewness and heteroscedasticity characteristics of precipitation, but also improve the accuracy of precipitation downscaling and quantify the uncertainty of precipitation downscaling.

[0005] To achieve the above purpose, the technical solution adopted by the present application is: A precipitation downscaling method based on a deep learning network model fusing precipitation prior knowledge, comprising the following steps, Step A, collecting terrain data and low-resolution daily precipitation data of a target area as input data; Step B, using short-term high-resolution precipitation fields generated by a mesoscale weather forecast WRF model as training true value data; Step C, constructing a deep learning network model architecture fusing precipitation prior knowledge by using the spatiotemporal correlation of precipitation; Step D, establishing a precipitation loss function based on a weighted negative log-likelihood constructed based on a conditional censored Gaussian distribution according to the zero inflation, heteroscedasticity and skewness characteristics of precipitation, and then applying the precipitation loss function to the deep learning network model architecture to obtain a deep learning downscaling model; Step E, inputting the input data and the training true value data into the deep learning downscaling model to obtain high-resolution precipitation distribution data, and then using a Monte Carlo sampling method to extract samples from the high-resolution precipitation distribution data as the downscaling results.

[0006] The aforementioned precipitation downscaling method based on a deep learning network model fusing precipitation prior knowledge, in step A, the terrain data is resampled to the same resolution as the low-resolution daily precipitation data and then normalized.

[0007] The aforementioned precipitation downscaling method based on a deep learning network model fusing precipitation prior knowledge, in step C, the deep learning network model architecture fusing precipitation prior knowledge is constructed by using the spatiotemporal correlation of precipitation, and the deep learning network model architecture comprises a shallow feature extraction module, a deep feature extraction module and a high-resolution reconstruction module, the shallow feature extraction module is composed of a convolution block, the convolution block comprises a first convolutional neural network Conv2D, a BatchNorm layer and a ReLU layer, the deep feature extraction module is composed of a residual attention module and a long short-term memory network LSTM, the residual attention module comprises a second convolutional neural network Conv2D and a spatial attention mechanism, and the high-resolution reconstruction module is composed of an upsampling module and a third convolutional neural network Conv2D, the upsampling module comprises a fourth convolutional neural network Conv2D, a pixel transformation layer and a ReLU layer.

[0008] The aforementioned precipitation downscaling method based on a deep learning network model fusing precipitation prior knowledge, in step D, a precipitation loss function based on a weighted negative log-likelihood constructed based on a conditional censored Gaussian distribution is established according to the zero inflation, heteroscedasticity and skewness characteristics of precipitation, and then the precipitation loss function is applied to the deep learning network model architecture to obtain a deep learning downscaling model, wherein the precipitation loss function is shown in formula (1), (1) wherein, is the precipitation loss, , and are the length of time series, the length of image and the width of image respectively, is the time at which, and are the horizontal and vertical coordinates of the precipitation grid respectively, is the empirical weight set according to the size of the precipitation intensity, is the training true value data, is the training true value data after Box-Cox transformation, is the Jacobian transformation, is the sign of the Gaussian cumulative distribution function, is the sign of the Gaussian probability function is the display representation, is the deletion threshold.

[0009] The aforementioned precipitation downscaling method based on the fusion of precipitation prior knowledge deep learning network model, step E, inputs the input data and the training true value data into the deep learning downscaling model and obtains the high-resolution precipitation distribution data, and then uses the Monte Carlo sampling method to extract samples from the high-resolution precipitation distribution data as the downscaling result, and the specific steps are as follows, Step E1, the low-resolution daily precipitation data in the input data and the training true value data are processed, if the value of the low-resolution daily precipitation data and the training true value data is ≤ 0.1 mm / d, then a uniform distribution U(0, 0.1) is used for random filling; Step E2, the precipitation data is processed using the Box-Cox formula, as shown in formula (2), (2) wherein, is the precipitation variable after Box-Cox transformation, is the precipitation variable, and are both constants; Step E3, the Box-Cox transformed precipitation data is modeled using left-censored Gaussian distribution, and the deletion threshold is set to the corresponding value of 0.1 mm / d after transformation, when the predicted value , the predicted value belongs to the censored data, at this time the predicted value Gaussian cumulative distribution is as shown in formula (3), (3) where, is the Box-Cox transformation of the low-resolution daily precipitation data, is the Box-Cox transformation of the low-resolution daily precipitation data, is the Box-Cox transformation of the low-resolution daily precipitation data, and are the Gaussian distribution mean and variance of the neural network output when the input is ; and Step E4, when the predicted value , then the predicted value probability density function is shown in equation (4), (4).

[0010] A precipitation downscaling system based on a deep learning network model fusing precipitation prior knowledge, comprising a data acquisition unit, a data generation unit, a model construction unit, a model obtaining unit and a downscaling unit, the data acquisition unit is used for acquiring terrain data and low-resolution daily precipitation data of a target area and taking them as input data; the data generation unit is used for taking short-term high-resolution precipitation fields generated by a mesoscale weather forecast WRF model as training true value data; the model construction unit is used for constructing a deep learning network model architecture fusing precipitation prior knowledge by using precipitation spatiotemporal correlation; the model obtaining unit is used for establishing a precipitation loss function based on a weighted negative log-likelihood of a conditional censored Gaussian distribution according to zero inflation, heteroscedasticity and skewness characteristics of precipitation, then applying the precipitation loss function to the deep learning network model architecture and obtaining a deep learning downscaling model; the downscaling unit is used for inputting the input data and the training true value data into the deep learning downscaling model and obtaining high-resolution precipitation distribution data, then using a Monte Carlo sampling method to extract samples from the high-resolution precipitation distribution data as downscaling results.

[0011] The beneficial effects of this invention are as follows: This invention provides a precipitation downscaling method and system based on a deep learning network model that integrates prior precipitation knowledge. First, topographic data and low-resolution daily precipitation data of the target area are collected as input data. Then, a short-term high-resolution precipitation field generated by a mesoscale weather forecasting (WRF) model is used as the training ground data. Next, a deep learning network model architecture integrating prior precipitation knowledge is constructed using the spatiotemporal correlation of precipitation. Subsequently, based on the zero-expansion, heteroscedasticity, and skewness characteristics of precipitation, a precipitation loss function based on a conditionally censored Gaussian distribution and weighted negative log-likelihood is established. This precipitation loss function is then applied to the deep learning network model architecture to obtain a deep learning downscaling model. Finally, the input data is processed... Input data and training ground data are fed into a deep learning downscaling model to obtain high-resolution precipitation distribution data. Then, Monte Carlo sampling is used to extract samples from the high-resolution precipitation distribution data as the downscaling result. This effectively realizes that the precipitation downscaling method and system can accurately downscale precipitation data by combining convolutional neural networks and long short-term memory networks. Moreover, the spatiotemporal correlation of precipitation is fully considered in the downscaling process. At the same time, the likelihood function combining coupled censored data, Box-Cox transformation and Gaussian distribution of time-varying variance is used as the precipitation loss function, which can not only characterize the zero expansion, skewness and heteroscedasticity of precipitation, but also improve the accuracy of precipitation downscaling and quantify the uncertainty of precipitation downscaling. Attached Figure Description

[0012] Figure 1 This is an overall flowchart of a precipitation downscaling method based on a deep learning network model that integrates prior precipitation knowledge, according to the present invention. Figure 2 This is a schematic diagram illustrating the downscaling principle of a precipitation downscaling system based on a deep learning network model that integrates prior precipitation knowledge, according to the present invention. Figure 3 This is a spatial distribution map of precipitation before and after downscaling in an embodiment of the present invention. Detailed Implementation

[0013] The present invention will now be further described with reference to the accompanying drawings.

[0014] like Figure 1 As shown, the present invention provides a precipitation downscaling method based on a deep learning network model that integrates prior precipitation knowledge, comprising the following steps: Step A: Collect topographic data and low-resolution daily precipitation data of the target area as input data. The topographic data is resampled to the same resolution as the low-resolution daily precipitation data and then normalized.

[0015] Step B uses the short-term high-resolution precipitation field generated by the mesoscale weather forecast WRF model as the training ground truth data. Step C involves constructing a deep learning network model architecture that integrates prior knowledge of precipitation, utilizing the spatiotemporal correlation of precipitation. This deep learning network model architecture includes a shallow feature extraction module, a deep feature extraction module, and a high-resolution reconstruction module. The shallow feature extraction module is composed of convolutional blocks, each containing a first convolutional neural network (Conv2D), a BatchNorm layer, and a ReLU layer. The deep feature extraction module is composed of a residual attention module and a Long Short-Term Memory (LSTM) network, with the residual attention module containing a second convolutional neural network (Conv2D) and a spatial attention mechanism. The high-resolution reconstruction module is composed of an upsampling module and a third convolutional neural network (Conv2D), with the upsampling module containing a fourth convolutional neural network (Conv2D), a pixel transformation layer, and a ReLU layer.

[0016] Preferably, the shallow feature extraction module consists of a convolutional block, which includes a first convolutional neural network (Conv2D), a BatchNorm layer, and a ReLU layer; the deep feature extraction module consists of multiple residual attention modules and a two-layer long short-term memory (LSTM) network, which includes two second convolutional neural networks (Conv2D) and a spatial attention mechanism; the high-resolution reconstruction module consists of two upsampling modules and a third convolutional neural network (Conv2D), which includes a fourth convolutional neural network (Conv2D), two pixel transformation layers, and a ReLU layer.

[0017] Step D: Based on the zero expansion, heteroscedasticity, and skewness characteristics of precipitation, a precipitation loss function based on a conditionally censored Gaussian distribution and weighted negative log-likelihood is established. This precipitation loss function is then applied to a deep learning network model architecture to obtain a deep learning downscaling model. The precipitation loss function is shown in Equation (1). (1) in, For rainfall loss, , and Time series length, image length, and width, respectively. For the current time, and These are the x and y coordinates of the precipitation grid points, respectively. The empirical weights are set based on the magnitude of precipitation intensity. To train on ground truth data, The training ground truth data is after Box-Cox transformation. For Jacobi transformation, The symbol for the Gaussian cumulative distribution function is... The symbol for the Gaussian probability function The display indicates that, This is the censoring threshold.

[0018] Step E involves inputting the input data and training ground data into a deep learning downscaling model to obtain high-resolution precipitation distribution data. Then, Monte Carlo sampling is used to extract samples from the high-resolution precipitation distribution data as the downscaling result. The specific steps are as follows: Step E1: Process the low-resolution daily precipitation data in the input data. and training ground data Processing is required for low-resolution daily precipitation data. and training ground data If the value is ≤ 0.1 mm / d, then a uniform distribution U(0, 0.1) is used for random filling; Step E2: The precipitation data is processed using the Box-Cox formula, as shown in formula (2). (2) in, The precipitation variable is after Box-Cox transformation. For precipitation variables, and All are constants; Step E3: Model the precipitation data after Box-Cox transformation using a left-censored Gaussian distribution and set a censoring threshold. The transformed value corresponds to 0.1 mm / d, when the predicted value... If the predicted value belongs to the censored data, then the Gaussian cumulative distribution of the predicted value is as shown in formula (3). (3) in, The symbol for Gaussian cumulative distribution function The display indicates that, This is low-resolution daily precipitation data after Box-Cox transformation. and The inputs are respectively The mean and variance of the Gaussian distribution output of the time-space neural network; Step E4, when the predicted value When the predicted value probability density function is as shown in formula (4), (4).

[0019] like Figure 2As shown, a precipitation downscaling system based on a deep learning network model that integrates prior precipitation knowledge includes a data acquisition unit, a data generation unit, a model building unit, a model acquisition unit, and a downscaling unit. The data acquisition unit collects topographic data and low-resolution daily precipitation data of the target area as input data. The data generation unit uses a short-term high-resolution precipitation field generated by a mesoscale weather forecasting (WRF) model as training ground data. The model building unit constructs a deep learning network model architecture that integrates prior precipitation knowledge using the spatiotemporal correlation of precipitation. The model acquisition unit establishes a precipitation loss function based on a conditionally censored Gaussian distribution and weighted negative log-likelihood, based on the zero expansion, heteroscedasticity, and skewness characteristics of precipitation. This precipitation loss function is then applied to the deep learning network model architecture to obtain a deep learning downscaling model. The downscaling unit inputs the input data and training ground data into the deep learning downscaling model to obtain high-resolution precipitation distribution data. Monte Carlo sampling is then used to extract samples from the high-resolution precipitation distribution data as the downscaling result.

[0020] To better illustrate the effectiveness of the present invention, a specific embodiment of using the method of the present invention for precipitation downscaling is described below.

[0021] This study selected the Tibetan Plateau as the research area and used the WRF model to simulate high-resolution precipitation data from 2012 to 2021. The WRF model employed a four-layer nested grid configuration with horizontal resolutions of 27 km, 9 km, 3 km, and 1 km, respectively, and the simulation results of the innermost 1 km resolution were taken as the true precipitation values. The coarse-resolution precipitation data were obtained from ERA5-Land reanalysis data, with a spatial resolution of approximately 9 km.

[0022] In this embodiment, the data from 2012 to 2021 are divided into training, validation, and testing periods based on 70%, 20%, and 10% of the data length, respectively, for deep learning probabilistic precipitation downscaling model calculation.

[0023] like Figure 3 As shown in the figure, this embodiment provides a spatial distribution map of precipitation before and after downscaling, which shows that the details of the spatial distribution of precipitation have been greatly improved.

[0024] In summary, the precipitation downscaling method and system of the present invention, based on a deep learning network model that integrates prior precipitation knowledge, first collects topographic data and low-resolution daily precipitation data of the target area as input data. Then, it uses a short-term high-resolution precipitation field generated by a mesoscale weather forecast WRF model as ground-value training data. Next, it constructs a deep learning network model architecture that integrates prior precipitation knowledge using the spatiotemporal correlation of precipitation. Subsequently, based on the zero-expansion, heteroscedasticity, and skewness characteristics of precipitation, it establishes a precipitation loss function based on a conditionally censored Gaussian distribution and a weighted negative log-likelihood. This precipitation loss function is then applied to the deep learning network model architecture to obtain a deep learning downscaling model. Finally, the input data is... The deep learning downscaling model is input with training ground data to obtain high-resolution precipitation distribution data. Then, Monte Carlo sampling is used to extract samples from the high-resolution precipitation distribution data as the downscaling result. This effectively realizes that the precipitation downscaling method and system can accurately downscale precipitation data by combining convolutional neural networks and long short-term memory networks. The downscaling process fully considers the spatiotemporal correlation of precipitation. At the same time, the likelihood function combining coupled censored data, Box-Cox transformation and Gaussian distribution of time-varying variance is used as the precipitation loss function, which can not only characterize the zero expansion, skewness and heteroscedasticity of precipitation, but also improve the accuracy of precipitation downscaling and quantify the uncertainty of precipitation downscaling.

[0025] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A precipitation downscaling method based on a fusion precipitation prior knowledge deep learning network model, characterized in that: The method comprises the following steps: Step A, collecting terrain data and low-resolution daily precipitation data of a target area as input data; Step B, using short-term high-resolution precipitation fields generated by a mesoscale weather forecast WRF model as training true value data; Step C, constructing a deep learning network model architecture that integrates precipitation prior knowledge by using the spatiotemporal correlation of precipitation; Step D, establishing a precipitation loss function based on a conditional censored Gaussian distribution by constructing a weighted negative log-likelihood according to the zero inflation, heteroscedasticity and skewness characteristics of precipitation, and then applying the precipitation loss function to the deep learning network model architecture to obtain a deep learning downscaling model; Step E, inputting the input data and the training true value data into the deep learning downscaling model to obtain high-resolution precipitation distribution data, and then using a Monte Carlo sampling method to extract samples from the high-resolution precipitation distribution data as the downscaling result.

2. The precipitation downscaling method based on the fusion precipitation prior knowledge deep learning network model according to claim 1, characterized in that: Step A, collecting terrain data and low-resolution daily precipitation data of a target area as input data, wherein the terrain data is resampled to the same resolution as the low-resolution daily precipitation data and then normalized.

3. The precipitation downscaling method based on the fusion precipitation prior knowledge deep learning network model according to claim 2, characterized in that: Step C, constructing a deep learning network model architecture that integrates precipitation prior knowledge by using the spatiotemporal correlation of precipitation, wherein the deep learning network model architecture comprises a shallow feature extraction module, a deep feature extraction module and a high-resolution reconstruction module, the shallow feature extraction module is composed of a convolution block, the convolution block comprises a first convolutional neural network Conv2D, a BatchNorm layer and a ReLU layer; the deep feature extraction module is composed of a residual attention module and a long short-term memory network LSTM, the residual attention module comprises a second convolutional neural network Conv2D and a spatial attention mechanism; the high-resolution reconstruction module is composed of an upsampling module and a third convolutional neural network Conv2D, the upsampling module comprises a fourth convolutional neural network Conv2D, a pixel transformation layer and a ReLU layer.

4. The precipitation downscaling method based on the fusion precipitation prior knowledge deep learning network model according to claim 3, characterized in that: Step D, establishing a precipitation loss function based on a conditional censored Gaussian distribution by constructing a weighted negative log-likelihood according to the zero inflation, heteroscedasticity and skewness characteristics of precipitation, and then applying the precipitation loss function to the deep learning network model architecture to obtain a deep learning downscaling model, wherein the precipitation loss function is shown in formula (1), (1) wherein, is the precipitation loss, , and are the length of time series, image length and width respectively, is the time of the moment, and are the horizontal and vertical coordinates of the precipitation grid respectively, is the empirical weight set according to the size of the precipitation intensity, is the training true value data, is the training true value data after Box-Cox transformation, is the Jacobian transformation, is the sign of the Gaussian cumulative distribution function, is the sign of the Gaussian probability function is the display representation of, is the deletion threshold.

5. The precipitation downscaling method based on the fusion precipitation prior knowledge deep learning network model according to claim 4, characterized in that: Step E, inputting the input data and the training true value data into the deep learning downscaling model to obtain high-resolution precipitation distribution data, and then using a Monte Carlo sampling method to extract samples from the high-resolution precipitation distribution data as the downscaling result, the specific steps are as follows, Step E1. Process low-resolution daily precipitation data in the input data and the training ground truth data if the value of the low-resolution daily precipitation data and the training ground truth data is ≤ 0.1 mm / d, fill in randomly with a uniform distribution U(0, 0.1). Step E2, processing the precipitation data using the Box-Cox formula, as shown in formula (2), (2) wherein, is the Box-Cox transformed precipitation variable, is the precipitation variable, and are constants; Step E3, modeling the Box-Cox transformed precipitation data using left-truncated Gaussian distribution, setting the truncation threshold For the transformed 0.1 mm / d corresponding value, when the predicted value is greater than the truncation threshold, then the predicted value belongs to the truncated data, and the predicted value Gaussian cumulative distribution is shown in formula (3), (3) wherein, is the display representation of the Gaussian cumulative distribution function symbol , is the Box-Cox transformed low resolution daily precipitation data, and are the Gaussian distribution mean and variance output by the neural network when the input is , respectively. Step E4, when the prediction value is then given by the probability density function of the prediction value as shown in equation (4), (4)。 6. A precipitation downscaling system based on a deep learning network model with fused precipitation prior knowledge, wherein the specific downscaling process of the precipitation downscaling system is based on the precipitation downscaling method of any one of claims 1-5. The system comprises a data acquisition unit, a data generation unit, a model construction unit, a model obtaining unit and a downscaling unit, the data acquisition unit is used to collect terrain data and low-resolution daily precipitation data of a target area as input data; the data generation unit is used to use short-term high-resolution precipitation fields generated by a mesoscale weather forecast WRF model as training true value data; The model construction unit is configured to construct a deep learning network model architecture fusing precipitation prior knowledge by using spatiotemporal correlation of precipitation; The model obtaining unit is configured to establish a precipitation loss function based on a weighted negative log-likelihood constructed by a conditional censored Gaussian distribution according to zero inflation, heteroscedasticity and skewness characteristics of the precipitation, and apply the precipitation loss function to the deep learning network model architecture to obtain a deep learning downscaling model; The downscaling unit is configured to input input data and training true value data into the deep learning downscaling model to obtain high-resolution precipitation distribution data, and extract samples from the high-resolution precipitation distribution data as a downscaling result by using a Monte Carlo sampling method.

Citation Information

Cited By

  • Deep learning and dynamic downscaling fused small watershed future scene climate rapid prediction method

    CN122087740A