Climate data downscaling method and device, equipment and storage medium

Abstract

The invention belongs to the field of electric power, and discloses a downscaling method, device and equipment for climate data and a storage medium, and the method comprises the steps: obtaining a low-resolution climate data sequence containing multiple time scales, and generating model input data; model input data is input to an encoder module, and the encoder module adopts a 3D causal convolutional network to perform feature extraction and downsampling on the input data so as to capture local structural characteristics of the low-resolution climate data sequence in time and space dimensions and map the local structural characteristics to a potential space; features of the potential space are input to a decoder module, which up-samples over a 3D causal convolutional network to generate high resolution climate data. According to the invention, by introducing the 3D causal convolutional network, the features of climate data in space and time dimensions can be captured, the time causal law is strictly followed, the smoothness and continuity of the generated high-resolution data in a space-time sequence are ensured, and a time mutation phenomenon is avoided.

Description

Technical Field

[0001] This invention belongs to the field of power, and particularly relates to a method, apparatus, equipment and storage medium for downscaling climate data. Background Technology

[0002] High-resolution climate data is crucial for accurate weather forecasting, climate research, and environmental management. However, due to limitations in observation equipment and computational costs, directly acquiring large-scale, long-term high-resolution climate data is extremely difficult. Climate downscaling techniques aim to address this problem by converting widely available low-resolution data into high-resolution data.

[0003] Traditional downscaling methods are mainly divided into two categories: statistical downscaling and dynamic (physical) downscaling. Statistical downscaling methods establish statistical relationships between low-resolution and high-resolution data based on historical data, but they struggle to capture complex nonlinear relationships and spatiotemporal causal relationships in the climate system, especially exhibiting poor generalization ability in data-scarce regions. Dynamic downscaling methods perform regional climate simulations based on physical equations. While possessing clear physical meaning, they are computationally extremely expensive and highly sensitive to initial conditions and parameterization schemes, making it difficult to meet the demands for high efficiency and real-time performance in practical applications.

[0004] In related technologies, deep learning-based methods, such as convolutional neural networks (CNNs), have been applied to climate downscaling tasks and have shown potential to outperform traditional methods. However, existing deep learning methods mostly focus on spatial super-resolution at a single time step, lacking effective modeling of continuity and causal relationships in the temporal dimension. This leads to potential discontinuities or abrupt changes in the generated data sequences over time, resulting in insufficient spatiotemporal consistency and affecting their reliability in time series prediction and analysis. Summary of the Invention

[0005] In view of this, the present invention discloses a method, apparatus, device and storage medium for downscaling climate data, which can solve the shortcomings of related technologies.

[0006] To achieve the above objectives, the present invention discloses the following technical solution:

[0007] According to a first aspect of the present invention, a method for downscaling climate data is proposed, the method comprising: Acquire low-resolution climate data sequences containing multiple time scales and generate model input data; The model input data is input to the encoder module, which uses a 3D causal convolutional network to extract features and downsample the input data in order to capture the local structural characteristics of the low-resolution climate data sequence in the time and space dimensions and map them to the latent space. The features of the latent space are input into the decoder module, which upsamples the data through a 3D causal convolutional network to generate high-resolution climate data.

[0008] According to a second aspect of the present invention, a climate data downscaling device is provided, the device comprising: Acquisition Unit: Acquires low-resolution climate data sequences containing multiple time scales and generates model input data; Downsampling unit: Inputs the model input data to the encoder module. The encoder module uses a 3D causal convolutional network to extract features and downsample the input data in order to capture the local structural characteristics of the low-resolution climate data sequence in the time and space dimensions and map them to the latent space. Upsampling unit: Inputs the features of the latent space into the decoder module, which performs upsampling through a 3D causal convolutional network to generate high-resolution climate data.

[0009] According to a third aspect of the present invention, an electronic device is provided, comprising: processor; Memory used to store processor-executable instructions; The processor implements the steps of the method as described in the first aspect by running the executable instructions.

[0010] According to a fourth aspect of the invention, a computer-readable storage medium is provided having computer instructions stored thereon that, when executed by a processor, implement the steps of the method as described in the first aspect.

[0011] As can be seen from the above technical solutions, the climate data downscaling method disclosed in this invention is as follows: On the one hand, by introducing a 3D causal convolutional network, this invention can simultaneously capture the features of climate data in both spatial and temporal dimensions, and strictly follow the temporal causality law, ensuring the smoothness and continuity of the generated high-resolution data in the spatiotemporal sequence and avoiding abrupt changes in time. On the other hand, by combining an autoregressive module and multi-task loss functions (such as Charbonnier loss and LPIPS perceptual loss), the model not only approximates real data at the pixel level, but also retains complex climate structure features (such as cloud systems and precipitation belts) at the perceptual level, improving the accuracy and visual quality of the downscaling results. Attached Figure Description

[0012] Figure 1 This is a flowchart illustrating a method for downscaling climate data, provided in an exemplary embodiment. Figure 2 This is a schematic diagram of a downscaling model provided in an exemplary embodiment; Figure 3 This is a schematic structural diagram of a device provided in an exemplary embodiment; Figure 4 This is a block diagram of a climate data downscaling device provided in an exemplary embodiment. Detailed Implementation

[0013] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with one or more embodiments of the present invention. Rather, they are merely examples of apparatuses and methods consistent with some aspects of one or more embodiments of the present invention.

[0014] It should be noted that in other embodiments, the corresponding methods are not necessarily performed in the order shown and described in this invention. The method comprises steps. In some other embodiments, the method may include more or fewer steps than those described in this invention. Furthermore, a single step described in this invention may be broken down into multiple steps in other embodiments; and multiple steps described in this invention may be combined into a single step in other embodiments.

[0015] High-resolution climate data is crucial for accurate weather forecasting, climate research, and environmental management. However, due to limitations in observation equipment and computational costs, directly acquiring large-scale, long-term high-resolution climate data is extremely difficult. Climate downscaling techniques aim to address this problem by converting widely available low-resolution data into high-resolution data.

[0016] Traditional downscaling methods are mainly divided into two categories: statistical downscaling and dynamic (physical) downscaling. Statistical downscaling methods establish statistical relationships between low-resolution and high-resolution data based on historical data, but they struggle to capture complex nonlinear relationships and spatiotemporal causal relationships in the climate system, especially exhibiting poor generalization ability in data-scarce regions. Dynamic downscaling methods perform regional climate simulations based on physical equations. While possessing clear physical meaning, they are computationally extremely expensive and highly sensitive to initial conditions and parameterization schemes, making it difficult to meet the demands for high efficiency and real-time performance in practical applications.

[0017] In related technologies, deep learning-based methods, such as convolutional neural networks (CNNs), have been applied to climate downscaling tasks and have shown potential to outperform traditional methods. However, existing deep learning methods mostly focus on spatial super-resolution at a single time step, lacking effective modeling of continuity and causal relationships in the temporal dimension. This leads to potential discontinuities or abrupt changes in the generated data sequences over time, resulting in insufficient spatiotemporal consistency and affecting their reliability in time series prediction and analysis.

[0018] To address the shortcomings of related technologies, this invention proposes a method, apparatus, device, and storage medium for downscaling climate data.

[0019] Figure 1 This is a flowchart illustrating a method for downscaling climate data, as provided in an exemplary embodiment. Figure 1 As shown, the method may include the following steps: Step 101: Obtain low-resolution climate data sequences containing multiple time scales and generate model input data.

[0020] The input data is a low-resolution climate data series containing multiple time scales. The data is preprocessed and transformed into tensors suitable for model input. The encoder module uses a 3D causal convolutional network to extract features and downsample the input data, capturing the local structural characteristics of the climate data in the temporal and spatial dimensions and mapping them into the latent space. The decoder module gradually recovers the features in the latent space into high-resolution climate data through sampling on the 3D causal convolutional network. To enhance the model's consistency across time scales, during the training phase, the autoregressive module performs autoregressive predictions on the generated high-resolution data, resulting in a high-resolution climate data sequence containing multiple time scales. By using supervision through multiple loss functions, the consistency and accuracy of the model across time steps and spatial scales can be further improved.

[0021] Specifically, the data series contains at least two climate variables at different time scales, such as daily average precipitation and hourly temperature fields. It is multi-level, specifically including but not limited to the precipitation field, temperature field and wind speed field at the ground level, as well as various meteorological elements at the pressure layer. There are a total of 13 pressure layers (50, 150, 200, 250, 300, 400, 500, 600, 700, 850, 925, 1000, unit: hPa) containing a total of 70 meteorological element channels. The low-resolution climate data sequence is preprocessed, including normalizing climate variables at different scales and reshaping the sequence into tensors suitable for input to a 3D Causal CNN. Its dimensions are ,in For time step, and This refers to the number of climate variable channels (i.e., 70 meteorological elements). and These are low-resolution spatial dimensions to ensure the spatiotemporal consistency of the input data; Input data into the model The input data is fed into an encoder module, which uses a stacked 3D causal convolutional network to extract features and downsample the input data. The 3D causal convolutional kernels are spatially... It possesses the feature capture capability of standard convolution, while in the time dimension ( Look only forward to ensure the current moment. The output features depend only on the input sequence The encoder maps the spatiotemporal features to a latent space based on data from the previous time point and earlier. The feature extraction and downsampling processes follow causal constraints, and their convolution operations adhere to the following principles: the temporal depth of the convolution kernel is designed to be asymmetric, using only... The time step kernels are weighted and summed to ensure strict causal consistency in the time dimension; Features of the latent space The data is input to a decoder module, which upsamples the data using a stacked 3D causal deconvolutional network (or transposed 3D convolutional network) to progressively recover high-resolution climate data from the features of the latent space. Its dimensions are The upsampling process incorporates skip connections to fuse multi-scale features from the encoder module, thereby effectively recovering and generating refined spatial details.

[0022] Step 102: Input the model input data into the encoder module. The encoder module uses a 3D causal convolutional network to extract features and downsample the input data in order to capture the local structural characteristics of the low-resolution climate data sequence in the time and space dimensions and map them to the latent space. Step 103: Input the features of the latent space into the decoder module, which upsamples the data through a 3D causal convolutional network to generate high-resolution climate data.

[0023] In this embodiment, by introducing a 3D causal convolutional network, the present invention can simultaneously capture the features of climate data in both spatial and temporal dimensions, and strictly follow the temporal causality law, ensuring the smoothness and continuity of the generated high-resolution data in the spatiotemporal sequence, and effectively avoiding the phenomenon of temporal abrupt changes.

[0024] In one embodiment, generating model input data includes: preprocessing the low-resolution climate data sequence to transform the low-resolution climate data sequence into model input data suitable for model input.

[0025] The low-resolution climate data series is preprocessed, including at least the following: data cleaning to remove outliers and missing values ​​from the observations; unified spatiotemporal registration to align climate variables of different time scales and spatial resolutions (e.g., daily precipitation and hourly temperature) to a unified low-resolution grid and time step using linear or bilinear interpolation; and normalization to map the range of all climate variables to a uniform low-resolution grid. or Within the interval, its normalization operation It can be represented as: ; in, For raw climate data, The mean of the training set, The standard deviation of the training set. To prevent extremely small constants from being divided by zero, this operation aims to eliminate the influence of different dimensions between variables, thereby enhancing the convergence speed and stability of the model; finally, the normalized data sequence is transformed into a four-dimensional tensor. Its dimensions are .

[0026] In one embodiment, the method further includes: training a pre-trained model to construct a downscaling model; wherein the downscaling model includes an input module, an encoder module, and a decoder module, and the input module is used to receive and preprocess low-resolution climate data sequences.

[0027] Furthermore, the downscaling model also includes an autoregressive module; training the pre-training module includes: inputting the high-resolution climate data sequence generated by the decoder module into the autoregressive module, which predicts the climate data for the next time step based on the Transformer architecture to enhance the consistency of the generated data in the time dimension; and jointly optimizing the model using a multi-task loss function, which includes at least: a reconstruction loss for measuring pixel-level differences and a perceptual loss for measuring perceptual similarity.

[0028] like Figure 2As shown, the model includes an input module, an encoder module, a decoder module, and an autoregressive module.

[0029] The autoregressive module employs a Transformers-based architecture to perform time-series predictions on generated high-resolution climate data, thereby enhancing the model's consistency and accuracy across time scales. This module captures the temporal dependencies of climate data through a multi-head self-attention mechanism, while incorporating location encoding information to ensure the sequentiality and correlation between time steps in the time series. During the inference phase, the autoregressive module does not participate in inference. During the training phase, it uses a pre-trained model to predict the next time step of the high-resolution climate data sequence generated by the decoder, generating high-resolution climate data containing future time steps. Time-series data is used for supervision, thus improving the temporal consistency of the downscaling model.

[0030] The reconstruction loss can be Charbonnier loss. Charbonnier loss is a smooth L1 norm approximation loss function, commonly used in deep learning tasks such as image restoration, denoising, and super-resolution. Compared to directly using L1 loss, Charbonnier loss is smooth at the origin, facilitating gradient calculation and avoiding the instability of gradients when they approach zero.

[0031] The Charbonnier loss function is used to calculate the pixel-level difference between generated high-resolution climate data and real high-resolution climate data, and is defined as follows: ; in, For the high-resolution data at the predicted time t+1, This is the actual high-resolution data at time t+1. This is a preset minimum constant.

[0032] This perceptual loss can be LPIPS loss. The application of LPIPS loss in meteorological super-resolution tasks aims to assess the difference between model-generated meteorological data and high-resolution real data through perceptual similarity. Compared to traditional pixel-level losses (such as L1 or L2), LPIPS loss can better capture complex structural and detailed features in meteorological data, such as cloud texture and precipitation distribution, thereby improving the perceptual quality of the model's generated results.

[0033] The LPIPS loss function is used to calculate the perceptual similarity between generated high-resolution climate data and real high-resolution climate data in the feature space, and is defined as: ; Where I1 and I2 are the generated data and the real data, respectively. This represents the feature map of the l-th layer in the pre-trained neural network. represents the adaptive weights of the l-th layer.

[0034] This invention can employ a multi-task loss function to jointly optimize the model, wherein the multi-task loss function includes at least: a reconstruction loss for measuring pixel-level differences. and perceptual loss used to measure perceptual similarity The loss function is calculated by jointly utilizing the difference between the high-resolution data predicted by the autoregressive module for the next time step and the actual high-resolution data, as well as the difference between the generated data and the actual data at the current time step, thereby achieving joint supervision of the downscaling and time prediction tasks. The multi-task loss function aims to optimize the accuracy of spatial details and the smoothness of the time series in the generated data; the joint optimization form is as follows: ; in, and Reconstruction loss and perceived loss The weighting coefficients.

[0035] In this embodiment, by combining an autoregressive module and multi-task loss functions (such as Charbonnier loss and LPIPS perceptual loss), the model not only approximates real data at the pixel level, but also preserves complex climate structure features (such as cloud systems and precipitation belts) at the perceptual level, thereby improving the accuracy and visual quality of the downscaling results.

[0036] Furthermore, the method also includes: evaluating the pre-trained model based on multiple metrics, and determining the pre-trained model as the downscaling model after the pre-trained model reaches the multiple metrics; wherein the multiple metrics include: root square error and critical success index.

[0037] The model is trained end-to-end, employing joint optimization of multi-task loss functions to improve model performance and generation quality. The training process uses low-resolution climate data as input to learn and generate high-resolution, multi-timescale climate data. In the loss function design, the model combines Charbonnier loss and perceptual loss (LPIPS). Charbonnier loss, as a reconstruction loss, measures the pixel-level difference between generated data and real high-resolution data, exhibiting smoothness and stability. Perceptual loss ensures improved perceptual quality of the generated climate data by comparing the distance between generated and real data in the high-level feature space. Furthermore, to guarantee causal consistency across time and space, the model incorporates an autoregressive strategy for sequence prediction during training to capture the dynamic changes of time series, while multi-scale feature modeling enhances the generation effect in terms of spatial resolution. During the evaluation phase, model performance is measured using various metrics, including Root Mean Square Error (RMSE) and the Critical Success Index (CSI, or TS score). Experimental results show that the model can generate high-quality, high-resolution data in complex climate time-space scale tasks, significantly outperforming traditional methods. It can efficiently generate accurate and consistent high-resolution climate data, providing reliable technical support for practical applications.

[0038] Figure 3 This is a schematic structural diagram of a device provided in an exemplary embodiment. Please refer to... Figure 3 At the hardware level, the device includes a processor 302, an internal bus 304, a network interface 306, memory 308, and non-volatile memory 310, and may also include other hardware required for its functions. One or more embodiments of the present invention can be implemented in software, for example, the processor 302 reads the corresponding computer program from the non-volatile memory 310 into memory 308 and then runs it. Of course, in addition to software implementation, one or more embodiments of the present invention do not exclude other implementation methods, such as logic devices or a combination of hardware and software, etc. That is to say, the execution subject of the following processing flow is not limited to each logic unit, but can also be hardware or logic devices.

[0039] Please refer to Figure 4 A climate data downscaling device can be applied to, for example... Figure 4 The device shown, in order to implement the technical solution of the present invention, includes: The acquisition unit 401 is used to acquire low-resolution climate data sequences containing multiple time scales and generate model input data; The downsampling unit 402 is used to input the model input data to the encoder module. The encoder module uses a 3D causal convolutional network to extract features and downsample the input data in order to capture the local structural characteristics of the low-resolution climate data sequence in the time and space dimensions and map them to the latent space. The upsampling unit 403 is used to input the features of the latent space into the decoder module, which performs upsampling through a 3D causal convolutional network to generate high-resolution climate data.

[0040] Optionally, the acquisition unit 401 is specifically used for: The low-resolution climate data sequence is preprocessed to transform it into model input data suitable for model input.

[0041] Optionally, the device further includes: Training unit 404 is used to train a pre-trained model to construct a downscaling model; wherein the downscaling model includes an input module, an encoder module, and a decoder module, and the input module is used to receive and preprocess low-resolution climate data sequences.

[0042] Furthermore, the downscaling model also includes an autoregressive module; the training unit 404 is specifically used for: The high-resolution climate data sequence generated by the decoder module is input into the autoregressive module, which predicts the climate data for the next time step based on the Transformer architecture to enhance the consistency of the generated data in the time dimension. The model is jointly optimized using a multi-task loss function, which includes at least: a reconstruction loss for measuring pixel-level differences and a perceptual loss for measuring perceptual similarity.

[0043] Optionally, the reconstruction loss is the Charbonnier loss, used to calculate the pixel-level difference between the generated high-resolution climate data and the real high-resolution climate data. The Charbonnier loss function is defined as follows: ; in, For the high-resolution data at the predicted time t+1, This is the actual high-resolution data at time t+1. This is a preset minimum constant.

[0044] Optionally, the perceptual loss is the LPIPS loss, used to calculate the perceptual similarity between the generated high-resolution climate data and the real high-resolution climate data in the feature space. The LPIPS loss function is defined as follows: ; Where I1 and I2 are the generated data and the real data, respectively. This represents the feature map of the l-th layer in the pre-trained neural network. represents the adaptive weights of the l-th layer.

[0045] Optionally, the device further includes: Evaluation unit 405 is used to evaluate the pre-trained model based on multiple indicators, and to determine the pre-trained model as the downscaling model after the pre-trained model reaches the multiple indicators; wherein the multiple indicators include: root square error and critical success index.

[0046] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer, which can take the form of a personal computer, laptop computer, cellular phone, camera phone, smartphone, personal digital assistant, media player, navigation device, email sending and receiving device, game console, tablet computer, wearable device, or any combination of these devices.

[0047] In a typical configuration, a computer includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0048] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0049] Computer-readable media, including both permanent and non-permanent, removable and non-removable media, can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, disk storage, quantum memory, graphene-based storage media or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0050] For any other form of computer-readable medium (or computer-readable storage medium) as described above, computer instructions may be stored thereon, which, when executed by a processor, implement one or more of the above embodiments, thereby realizing the technical solution of the present invention.

[0051] The present invention also proposes a computer program that, when executed by a processor, implements one or more of the embodiments described above, thereby realizing the technical solution of the present invention. This computer program may be specifically recorded on the above-described or other computer-readable media, and the present invention does not impose any limitations on this.

[0052] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0053] The foregoing has described specific embodiments of the invention. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps described in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired results. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0054] The terminology used in one or more embodiments of the invention is for the purpose of describing particular embodiments only and is not intended to limit the invention. The singular forms “a,” “the,” and “the” used in one or more embodiments of the invention and in the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more associated listed items.

[0055] It should be understood that although the terms first, second, third, etc., may be used to describe various information in one or more embodiments of the present invention, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first information may also be referred to as second information without departing from the scope of one or more embodiments of the present invention, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."

[0056] The above description is merely a preferred embodiment of one or more embodiments of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of one or more embodiments of the present invention should be included within the protection scope of one or more embodiments of the present invention.

Claims

1. A method for downscaling climate data, characterized in that, The method includes: Acquire low-resolution climate data sequences containing multiple time scales and generate model input data; The model input data is input to the encoder module, which uses a 3D causal convolutional network to extract features and downsample the input data in order to capture the local structural characteristics of the low-resolution climate data sequence in the time and space dimensions and map them to the latent space. The features of the latent space are input into the decoder module, which upsamples the data through a 3D causal convolutional network to generate high-resolution climate data.

2. The method according to claim 1, characterized in that, The input data for the generative model includes: The low-resolution climate data sequence is preprocessed to transform it into model input data suitable for model input.

3. The method according to claim 1, characterized in that, The method further includes: The pre-trained model is trained to construct a downscaling model; wherein the downscaling model includes an input module, an encoder module, and a decoder module, and the input module is used to receive and preprocess low-resolution climate data sequences.

4. The method according to claim 3, characterized in that, The downscaling model further includes an autoregressive module; training the pre-training module includes: The high-resolution climate data sequence generated by the decoder module is input into the autoregressive module, which predicts the climate data for the next time step based on the Transformer architecture to enhance the consistency of the generated data in the time dimension. The model is jointly optimized using a multi-task loss function, which includes at least: a reconstruction loss for measuring pixel-level differences and a perceptual loss for measuring perceptual similarity.

5. The method according to claim 4, characterized in that, The reconstruction loss is the Charbonnier loss, used to calculate the pixel-level difference between the generated high-resolution climate data and the real high-resolution climate data. The Charbonnier loss function is defined as follows: ； in, For the high-resolution data at the predicted time t+1, This is the actual high-resolution data at time t+1. This is a preset minimum constant.

6. The method according to claim 4, characterized in that, The perceptual loss is the LPIPS loss, used to calculate the perceptual similarity between the generated high-resolution climate data and the real high-resolution climate data in the feature space. The LPIPS loss function is defined as follows: ； Where I1 and I2 are the generated data and the real data, respectively. This represents the feature map of the l-th layer in the pre-trained neural network. represents the adaptive weights of the l-th layer.

7. The method according to claim 3, characterized in that, The method further includes: The pre-trained model is evaluated based on multiple metrics, and once the pre-trained model meets all the metrics, it is determined as the downscaling model; wherein, the multiple metrics include: root square error and critical success index.

8. A climate data downscaling device, characterized in that, The device includes: Acquisition Unit: Acquires low-resolution climate data sequences containing multiple time scales and generates model input data; Downsampling unit: Inputs the model input data to the encoder module. The encoder module uses a 3D causal convolutional network to extract features and downsample the input data in order to capture the local structural characteristics of the low-resolution climate data sequence in the time and space dimensions and map them to the latent space. Upsampling unit: Inputs the features of the latent space into the decoder module, which performs upsampling through a 3D causal convolutional network to generate high-resolution climate data.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor implements the steps of the method as described in any one of claims 1-7 by running the executable instructions.

10. A computer-readable storage medium storing computer instructions thereon, characterized in that, When executed by the processor, this instruction implements the steps of the method as described in any one of claims 1-7.

Images