A downscaling correction method for climate model projections based on generative deep learning

Abstract

This invention discloses a generative deep learning-based downscaling correction method for climate model predictions. This method uses an unsupervised generative adversarial network to identify biases in global climate model hydrological and meteorological variable predictions. Using adversarial training of a generator and a discriminator, the adversarial loss is continuously reduced, enabling the trained generative deep learning network to handle complex nonlinear relationships, achieving a comprehensive optimization of downscaling and bias correction. This method can be used to correct the average bias of ensemble predictions of multi-model variables (precipitation, temperature, etc.) and improve their resolution. Inventive examples demonstrate that this method can effectively downscale multi-model ensemble predictions and achieves superior downscaling performance compared to commonly used quantile mapping and convolutional neural networks.

Description

Technical Field

[0001] The present invention relates to a climate model prediction method, and specifically to a climate model prediction downscaling correction method based on generative deep learning. Background Art

[0002] As global climate change becomes increasingly serious, climate models have become a key tool for studying climate change trends and predicting extreme weather events. However, due to their low spatial resolution, GCMs (Global Climate Models) (GCMs) struggle to accurately represent subtle changes in hydrological and meteorological variables at the regional scale. This is particularly true in regions with complex terrain and drastic climate change, where their simulations often fall short of the requirements for sophisticated applications. Therefore, climate model downscaling has become an important means of improving the applicability of climate data. Currently, downscaling methods are primarily divided into dynamical downscaling and statistical downscaling. Dynamical downscaling involves nesting GCMs within regional climate models (RCMs), which can capture local climate characteristics but is computationally expensive. Statistical downscaling, on the other hand, leverages the statistical relationship between historical observations and GCMs to convert low-resolution climate simulations into high-resolution data. Common methods include linear regression, quantile mapping (QM), support vector machines (SVM), and random forests (RF). However, traditional statistical downscaling methods often struggle to fully capture the nonlinear characteristics of the climate system and can produce significant bias when processing high-dimensional, complex climate data. Summary of the Invention

[0003] Purpose of the invention: In order to overcome the above-mentioned shortcomings, the present invention discloses a climate model prediction downscaling correction method based on generative deep learning, which can generate sufficiently accurate and reliable hydrological and meteorological variable downscaling data, thereby improving the adaptability of climate model predictions in the region.

[0004] Technical solution: The climate model prediction downscaling correction method based on generative deep learning described in the present invention includes: preprocessing climate model prediction data output by a global climate model, and correcting downscaling bias of the preprocessed climate model prediction data through a generative adversarial network downscaling correction model;

[0005] The generative adversarial network downscaling correction model includes a generator and a discriminator. The input of the generator is preprocessed climate model prediction data. The goal of the generator is to minimize the distribution difference between the generator's output data and the reference data; the output data of the upsampling layer in the generator is used as the input of the discriminator. The discriminator is used to distinguish the distribution difference between the output data of the upsampling layer in the generator and the reference data, and calculate the downscaling loss function according to the judgment result, and adjust the parameters of the generator and the discriminator according to the result of the downscaling loss function; the output data of the generator is downsampled, and the difference between the downsampled generator output data and the reference data is quantified by the deviation correction loss function, and the parameters of the generator are adjusted according to the result of the deviation correction loss function.

[0006] The preprocessing includes: using bilinear interpolation to adjust the temporal resolution of the climate model prediction data to be consistent with the temporal resolution of the reference data, and checking the missing values and outliers in the data; for missing values, using interpolation to fill them, and for outliers, using statistical methods to eliminate or correct them.

[0007] The climate model prediction data are monthly grid data of hydrological and meteorological variables, and the reference data are corresponding ERA5-Land reanalysis data.

[0008] The calculation formula of the downscaling loss function is:

[0009] ,

[0010] in, represents the downscaling loss, G Represents a generator, D represents the discriminator, x Represents the monthly data of the Sixth Coupled Model Intercomparison Project (CMIP6). z Represents reference data, Indicates data x From the reference data distribution P data expectations, Indicates data z Output data distribution P from the upsampling layer in the generator z expectations.

[0011] The calculation formula of the deviation correction loss function is:

[0012] ,

[0013] Where, represents the bias correction loss, x Represents the monthly data of the Sixth Coupled Model Intercomparison Project (CMIP6). represents the generator output data after downsampling, Represents the generator output data after downsampling Compared with the monthly data of CMIP6 The L1 norm of .

[0014] Training the generative adversarial network downscaling correction model includes: inputting preprocessed multiple global climate model monthly-scale hydrological and meteorological variable data and corresponding ERA5-Land reanalysis data into the generative adversarial network downscaling correction model, calculating the downscaling loss and the gradient of the downscaling loss for each parameter in the generator and the discriminator, adjusting the parameters of the generator and the discriminator, and then calculating the bias correction loss. When the downscaling loss and the bias correction loss no longer decrease, the model training is completed.

[0015] The trained generative adversarial network downscaling correction model is evaluated, including calculating the mean absolute error, correlation coefficient, and Kling-Gupta efficiency coefficient. The calculation formulas are:

[0016] ,

[0017] ,

[0018] ,

[0019] Among them, MAE represents the mean absolute error, CC represents the correlation coefficient, and KGE represents the Kling-Gupta efficiency coefficient. It represents the meteorological element correction value output after the meteorological element is downscaled and corrected by the generative adversarial network downscaling correction model. The meteorological element observation values represented by n Describe the number of samples, represents the mean value of the meteorological elements output after downscaling correction, represents the average value of meteorological element observations, represents the variance of the meteorological element correction value, Represents the variance of meteorological element observations.

[0020] The generator includes a sequentially connected input layer, a residual layer, and an upsampling layer; the residual layer includes multiple convolutional layers for feature extraction; the upsampling layer is used to adjust the resolution of the data; the discriminator includes a sequentially connected input layer, a residual layer, a downsampling layer, and a nonlinear layer, wherein the residual layer is used for feature extraction, and the nonlinear layer is used to enhance the nonlinear capability of the model; the generator and the discriminator are optimized through an adversarial training mechanism to ultimately achieve a dynamic balance.

[0021] The present invention provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the above-mentioned generative deep learning-based climate model prediction downscaling correction method is implemented.

[0022] The present invention provides a computer program product, comprising a computer program / instruction, which, when executed by a processor, implements the climate model prediction downscaling correction method based on generative deep learning.

[0023] Beneficial Effects: Compared to existing technologies, this invention offers the following advantages: Based on a generative adversarial network (GAN), it proposes a novel downscaling and bias correction scheme. The proposed GAN architecture combines downscaling and bias correction, effectively capturing high-order features and complex nonlinear relationships between climate models and reference data while simultaneously performing bias correction to provide accurate downscaling results. Compared to commonly used quantile mapping and convolutional neural network methods, GAN downscaling performs superiorly. This method has strong generalization capabilities and can adapt to climate model downscaling requirements in different regions and with varying climatic factors, thus possessing broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 This is a training flowchart of the generative adversarial network downscaling correction model described in the present invention;

[0025] Figure 2 This is a structural diagram of the generative adversarial network downscaling correction model described in the present invention;

[0026] Figure 3 A comparison of the spatial distribution of the downscaling results of the multi-mode average global land temperature using generative adversarial networks, convolutional neural networks, and quantile mapping in the embodiment. DETAILED DESCRIPTION

[0027] The present invention will be further explained below with reference to the accompanying drawings and embodiments.

[0028] The climate model prediction downscaling correction method based on generative deep learning of the present invention comprises: preprocessing climate model prediction data output by a global climate model, and performing downscaling bias correction on the preprocessed climate model prediction data through a generative adversarial network downscaling correction model;

[0029] The generative adversarial network downscaling correction model includes a generator and a discriminator. The input of the generator is preprocessed climate model prediction data. The goal of the generator is to minimize the distribution difference between the generator's output data and the reference data; the output data of the upsampling layer in the generator is used as the input of the discriminator. The discriminator is used to distinguish the distribution difference between the output data of the upsampling layer in the generator and the reference data, and calculate the downscaling loss function according to the judgment result, and adjust the parameters of the generator and the discriminator according to the result of the downscaling loss function; the output data of the generator is downsampled, and the difference between the downsampled generator output data and the reference data is quantified by the deviation correction loss function, and the parameters of the generator are adjusted according to the result of the deviation correction loss function.

[0030] In this example, the generator consists of one convolutional layer (input), six residual layers, and two upsampling layers. The discriminator includes two convolutional layers, three residual layers, three downsampling layers, and a nonlinear layer. These layers are managed by PyTorch's built-in model building library, Module. The generator is followed by two downsampling layers.

[0031] The calculation formula of the downscaling loss function used is:

[0032] ,

[0033] in, represents the downscaling loss, G Represents a generator, D represents the discriminator, x Represents the monthly data of the Sixth Coupled Model Intercomparison Project (CMIP6). z Represents reference data, Indicates data x From the reference data distribution P data expectations, Indicates data z Output data distribution P from the upsampling layer in the generator z expectations.

[0034] The calculation formula of the bias correction loss function used is:

[0035] ,

[0036] Where, represents the bias correction loss, x Represents the monthly data of the Sixth Coupled Model Intercomparison Project (CMIP6). represents the generator output data after downsampling, Represents the generator output data after downsampling Compared with the monthly data of CMIP6 The L1 norm of .

[0037] The training process of the generative adversarial network downscaling correction model is as follows: Figure 1 As shown, the following steps are included:

[0038] S1: Ten global climate model models were used as downscaled models for the global land area (excluding Antarctica), with a resolution of 1° × 1°. ERA5-Land data were used as reference data, with a parameter resolution of 0.1° × 0.1°. This reference data was used for downscaling and bilinearly interpolated to a resolution of 1° × 1°, which served as reference data for bias correction. Using the Python pytorch library, the data was first converted from array format to torch format. Table 1 shows detailed information for the ten global climate model models used in this example.

[0039] Table 1 Detailed information of the 10 global climate models

[0040]

[0041] The data is then segmented and dimensionally transformed to convert the data dimension into: ,

[0042] Wherein: t represents time (in this embodiment, the number of months from 1950 to 2014), m represents model (in this embodiment, 10, representing 10 models), x represents latitude (in this embodiment, 12), and y represents longitude (in this embodiment, 33).

[0043] At the same time, we also segmented and transformed the dimensions of the ERA5-Land reference data. The data dimensions were converted to: ,

[0044] Wherein, t represents time (in this embodiment, the number of months from 1950 to 2014), x represents latitude (in this embodiment, 120), and y represents longitude (in this embodiment, 330).

[0045] Use torch.utils.data.DataLoader to make model data and observation data into iterators.

[0046] S2: Use the torch.nn library and torch.nn.Module to build a generative adversarial network downscaling correction model (including torch.nn.Conv2d, torch.nn.Maxpool2D, torch.nn.ReLu, torch.nn.Linear, torch.nn.ConvTranspose2d). The generative adversarial network downscaling correction model is as follows Figure 2As shown in the figure, for the generator, the initial layer is implemented using torch.nn.Conv2d (with a kernel size of 3, stride of 3, zero padding, and 64 channels). The residual layer consists of two convolutional layers and activation functions, namely torch.nn.Conv2d and torch.nn.ReLu. The input is directly added to the output of the two convolutional and activation layers (with a kernel size of 3, stride of 3, zero padding, and channels set to 128 and 256, respectively). Furthermore, after every three residual layers, a downsampling layer (torch.nn.Maxpool2D with a kernel size of 2 and stride of 2) is added, followed by two upsampling layers (torch.nn.ConvTranspose2d with a kernel size of 2). For the discriminator, the initial convolutional layer is also constructed using torch.nn.Conv2d, followed by three residual layers and three downsampling layers, with the same parameters as the generator. Finally, a convolutional layer (with channels set to 1) is added.

[0047] S3: Randomly extract sample batch0 from the data iterator constructed by S1. In this example, the shape and size of the climate model sample are [16, 1, 12, 33], and the shape and size of the ERA5-Land reference data sample are [16, 1, 120, 330]. After upsampling the climate model sample generator, the result is a shape of [16, 1, 120, 330]. This result and the corresponding reference data are input into the discriminator to obtain two discriminant results, both of which are [16, 1, 30, 55]. Both are input into the formula , thereby calculating the downscaling loss. After that, the upsampled result is downsampled to obtain a result similar to the original shape ([16, 1, 12, 33]), which is compared with the reference data with a resolution of [16, 1, 12, 33] to calculate the bias correction loss, using the formula , calculating the bias correction loss. We add the two losses to obtain the total loss and calculate the gradient of the total loss with respect to the two generator parameters. We use only the downscaling loss to calculate the gradient of the two discriminator parameters. We then use the Adam optimizer to perform gradient descent on both the generator and discriminator simultaneously, thereby updating the gradients of the generator and discriminator. After updating the gradients, we randomly sample batch1 from the data iterator constructed by S1 and repeat the above steps until the total loss stabilizes. At this point, we can consider the generator and discriminator optimization complete.

[0048] S4: The difference between the downscaled multi-model average temperature forecast and the temperature reference data is evaluated using indicators such as mean absolute error (MSE), correlation coefficient (CC), and Kling-Gupta efficiency coefficient (KGE). The results are shown in Table 2. The results of the multi-model ensemble average downscaling method of the present invention are superior to the traditional convolutional neural network and quantile mapping method in all indicators.

[0049] Table 2 Multi-index results of temperature downscaling correction using three downscaling methods

[0050]

[0051] like Figure 3 Shown is a comparison of the spatial distribution of the downscaling results of the multi-model average global land temperature using generative adversarial networks (generative adversarial network downscaling correction model), convolutional neural networks, and quantile mapping.

[0052] The above analysis demonstrates that the generative deep learning approach used in this paper to perform bias correction for downscaling and fusion of climate model predictions can effectively improve the accuracy and reliability of multi-model forecasts. Compared to other techniques, this approach effectively captures the complex nonlinear relationships between climate models and reference data, exhibits strong generalization capabilities, and offers higher computational accuracy and reliability, promising broad application prospects.

[0053] In one embodiment, a computer device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the above-mentioned generative deep learning-based climate model prediction downscaling correction method is implemented.

[0054] In one embodiment, a computer program product is provided, comprising a computer program / instruction, which, when executed by a processor, implements the generative deep learning-based climate model prediction downscaling correction method.

[0055] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0056] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0057] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0058] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

Claims

1. A downscaling correction method for climate model prediction based on generative deep learning, characterized by: include: The multi-model climate model prediction data output by the global climate model are preprocessed, and the preprocessed multi-model climate model prediction data are downscaled and bias corrected through the generative adversarial network downscaling correction model; The generative adversarial network downscaling correction model includes a generator and a discriminator. The input of the generator is the preprocessed multi-model climate model prediction data. The goal of the generator is to minimize the distribution difference between the generator's output data and the reference data; the output data of the upsampling layer in the generator is used as the input of the discriminator. The discriminator is used to distinguish the distribution difference between the output data of the upsampling layer in the generator and the reference data, and calculate the downscaling loss function according to the judgment result, and adjust the parameters of the generator and the discriminator according to the result of the downscaling loss function; the output data of the generator is downsampled, and the difference between the downsampled generator output data and the reference data is quantified by the deviation correction loss function, and the parameters of the generator are adjusted according to the result of the deviation correction loss function.

2. The method for downscaling climate model prediction based on generative deep learning according to claim 1, characterized in that: The preprocessing includes: using bilinear interpolation to adjust the temporal resolution of the climate model prediction data to be consistent with the temporal resolution of the reference data, and checking the missing values and outliers in the data; for missing values, using interpolation to fill them, and for outliers, using statistical methods to eliminate or correct them.

3. The method for downscaling climate model prediction based on generative deep learning according to claim 1, characterized in that: The climate model prediction data are monthly grid data of hydrological and meteorological variables, and the reference data are corresponding ERA5-Land reanalysis data.

4. The method for downscaling climate model prediction based on generative deep learning according to claim 1, wherein: The calculation formula of the downscaling loss function is: ， in, represents the downscaling loss, G Represents a generator, D represents the discriminator, x Represents the monthly data of the Sixth Coupled Model Intercomparison Project (CMIP6). z Represents reference data, Indicates data x From the reference data distribution P data expectations, Indicates data z Output data distribution P from the upsampling layer in the generator z expectations.

5. The method for downscaling climate model prediction based on generative deep learning according to claim 1, wherein: The calculation formula of the deviation correction loss function is: ， Where, represents the bias correction loss, x Represents the monthly data of the Sixth Coupled Model Intercomparison Project (CMIP6). represents the generator output data after downsampling, Represents the generator output data after downsampling Compared with the monthly data of CMIP6 The L1 norm of .

6. The method for downscaling climate model prediction based on generative deep learning according to claim 1, characterized in that: Training the generative adversarial network downscaling correction model includes: inputting preprocessed multiple global climate model monthly-scale hydrological and meteorological variable data and corresponding ERA5-Land reanalysis data into the generative adversarial network downscaling correction model, calculating the downscaling loss and the gradient of the downscaling loss for each parameter in the generator and the discriminator, adjusting the parameters of the generator and the discriminator, and then calculating the bias correction loss. When the downscaling loss and the bias correction loss no longer decrease, the generative adversarial network downscaling correction model is trained.

7. The method for downscaling climate model prediction based on generative deep learning according to claim 6, characterized in that: The trained generative adversarial network downscaling correction model is evaluated, including calculating the mean absolute error, correlation coefficient, and Kling-Gupta efficiency coefficient. The calculation formulas are: ， ， ， Among them, MAE represents the mean absolute error, CC represents the correlation coefficient, and KGE represents the Kling-Gupta efficiency coefficient. It represents the meteorological element correction value output after the meteorological element is downscaled and corrected by the generative adversarial network downscaling correction model. The meteorological element observation values represented by n represents the number of samples, represents the mean value of the meteorological elements output after downscaling correction, represents the average value of meteorological element observations, represents the variance of the meteorological element correction value, Represents the variance of meteorological element observations.

8. The method for downscaling climate model prediction based on generative deep learning according to claim 1, characterized in that: The generator includes an input layer, a residual layer, and an upsampling layer connected in sequence; the residual layer includes multiple convolutional layers for feature extraction; the discriminator includes an input layer, a residual layer, a downsampling layer, and a nonlinear layer connected in sequence.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the climate model prediction downscaling correction method based on generative deep learning according to any one of claims 1 to 8 is implemented.

10. A computer program product comprising a computer program and / or instructions, characterized in that When the computer program and / or instructions are executed by a processor, the climate model prediction downscaling correction method based on generative deep learning according to any one of claims 1 to 8 is implemented.

Images