CN117236201A - A downscaling method based on Diffusion and ViT - Google Patents

Abstract

The invention discloses a downscaling method based on Diffusion and ViT, which includes the following steps: S1 establishes low-resolution numerical model precipitation forecast and high-resolution precipitation observation samples, and performs preprocessing; S2 constructs Diffusion-Vision-Transformer precipitation forecast model; S3 trains the model until the error of the Diffusion‑Vision‑Transformer converges, saves the model and makes predictions; the present invention greatly improves the training efficiency of the model and reduces the usage of the model by using the Vision Transformer model to replace the U‑Net structure in the original Diffusion model. at the predicted time.

Description

A downscaling method based on Diffusion and ViT

Technical field

The invention relates to the technical field of weather forecasting, and specifically relates to a downscaling method based on Diffusion and ViT.

Background technique

Most of the traditional statistical downscaling methods are based on linear framework models, which are difficult to handle complex and high-dimensional meteorological field data and characterize the nonlinear dynamic processes of the atmosphere. The rise of deep learning provides a new direction for representation learning of high-dimensional and highly nonlinear complex data such as meteorological element fields. By using an efficient spatial feature extraction module to extract key information of high-dimensional spatial data and establish a statistical model from low-resolution input to high-resolution output, the deep learning model can be effectively used in scenarios such as image denoising and image resolution improvement. , and such methods are generally called "super-resolution" models. However, how to efficiently transfer it to the downscaling problem of meteorology while further improving the computational efficiency and forecast accuracy of the model still requires further research and exploration.

Contents of the invention

Purpose of the invention: The purpose of the invention is to provide a downscaling method based on Diffusion and ViT to solve the problems of insufficient spatial resolution and large forecast errors in numerical model precipitation forecasts.

Technical solution: a downscaling method based on Diffusion and ViT according to the present invention, including the following steps:

S1: Establish low-resolution numerical model precipitation forecast and high-resolution precipitation observation samples, and perform preprocessing;

S2: Construct the Diffusion-Vision-Transformer precipitation forecast model; including the following steps:

S21: Perform forward noise on high-resolution precipitation observation samples in the Diffusion model;

S22: Use the Vision-Transformer model to extract high-order spatial features of low-resolution numerical model precipitation forecast;

S23: Denoise the results obtained in step S21 in the Diffusion model, and introduce the high-order spatial features obtained in step S22 as condition information to obtain a downscaled high-resolution precipitation forecast;

S3: Train the model until the error of Diffusion-Vision-Transformer converges, save the model and make predictions.

Further, in the step S1, the preprocessing includes: performing logarithmic and normalization operations on the data set.

Further, the specific process of step S21 is as follows:

Assume that the preprocessed high-resolution precipitation observation sample at a certain time , gradually add Gaussian noise to the original observations in T times/> ,get/> ;Then the data distribution at time t/> previous moment The formula is as follows:

;

in, is a preset constant hyperparameter, ranging between 0 and 1;

The data distribution at the last t-th time It can be obtained from the data at time 0/> The distribution is obtained, and the formula is as follows:

;

in, , and for/> , then/> .

Further, the details of S22 are as follows: input paired high-resolution precipitation observation samples and low-resolution numerical model precipitation forecast/> , and determine the number of steps T for forward noise addition, and the variance hyperparameter of the added random Gaussian noise. .

Further, the step S23 includes the following steps:

S231: Divide the low-resolution numerical model precipitation forecast into several blocks, and then linearly map the divided blocks;

S232: Use position coding to represent the position information of different blocks, and use the processed coding information as the input of N groups of self-attention modules;

S233: Use spatial self-attention module to replace convolution operation.

Further, the formula of step S231 is as follows:

;

in, is a group of divided tiles,/> is the weight coefficient to be trained,/> is the truncation coefficient to be trained,/> is a set of vectors after linear mapping.

Further, the position encoding in step S232 is a two-dimensional position embedding method.

Further, the details of step S233 are as follows:

Let a group of divided tiles be , using three sets of weights, namely query weights/> , key value weight/> , and numerical weight/> , divide the original data into three features: query matrix/> , key value matrix/> , numerical matrix/> ;then/> Corresponding self-attention/> The formula is as follows:

;

in, for/> The square root of the dimension.

Further, the details of step S3 are as follows:

Assume that the result obtained through steps S21-S22 is: ,in, is the model obtained in steps S21-S22,/> For low-resolution numerical model precipitation forecast,/> For paired high-resolution precipitation observation samples,/> is the preset hyperparameter in step S21, and T is the number of steps of forward noise addition in step S21; then the prediction error of the Diffusion-Vision-Transforme model in step S3/> The formula is as follows:

;

in, is a random Gaussian distribution, then/> ;

When the forecast error of the Diffusion-Vision-Transforme model When convergence occurs, the T steps are extrapolated backward until the model prediction is obtained/> ; Among them, the previous step's /> From the next step/> Obtained, the formula is as follows:

;

in, is the model obtained in steps S21-S22,/> For low-resolution numerical model precipitation forecast,/> is the hyperparameter preset in step S21,/> is a random Gaussian distribution, then/> .

A device according to the present invention includes a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, it implements any one of the Diffusion and Diffusion-based methods. Steps in ViT’s downscaling method.

Beneficial effects: Compared with the existing technology, the present invention has the following significant advantages: (1) By using the Diffusion model to improve the degree of refinement of downscaling forecasts, it is especially advantageous in tasks with downscaling multiples exceeding 4; (2) ) By using the Vision Transformer model to replace the U-Net structure in the original Diffusion model, the training efficiency of the model is greatly improved and the time for the model to be used for prediction is reduced.

Description of drawings

Figure 1 is a general flow chart of the present invention;

Figure 2 is a schematic diagram of the Diffusion-ViT model training process;

Figure 3 is a schematic diagram of the Diffusion model;

Figure 4 is a schematic diagram of the Vision-Transformer model.

Implementation

The technical solution of the present invention will be further described below with reference to the accompanying drawings.

As shown in Figure 1, the embodiment of the present invention provides a downscaling method based on Diffusion and ViT, which includes the following steps:

S1: Establish low-resolution numerical model precipitation forecast and high-resolution precipitation observation samples, and perform preprocessing; preprocessing includes: logarithmizing and normalizing the data set.

As shown in Figure 2, S2: Construct the Diffusion-Vision-Transformer precipitation forecast model; including the following steps:

S21: Perform forward noise on high-resolution precipitation observation samples in the Diffusion model; the details are as follows: As shown in Figure 3, assume that the high-resolution precipitation observation samples are preprocessed at a certain time , gradually add Gaussian noise to the original observations in T times/> ,get/> ;Then the data distribution at time t/> Previous moment/> The formula is as follows:;Then the data distribution at time t/> Previous moment/> The formula is as follows:

;

in, is a preset constant hyperparameter, ranging between 0 and 1;

The data distribution at the last t-th time It can be obtained from the data at time 0/> The distribution is obtained, and the formula is as follows:

;

in, , and for/> , then/> .

S22: Use the Vision-Transformer model to extract high-order spatial features of low-resolution numerical model precipitation forecasts; details are as follows: input paired high-resolution precipitation observation samples and low-resolution numerical model precipitation forecast/> , and determine the number of steps T for forward noise addition, and the variance hyperparameter of the added random Gaussian noise/> .

Denoise the results obtained in step S21 in the Diffusion model, and introduce the high-order spatial features obtained in step S22 as condition information to obtain a downscaled high-resolution precipitation forecast;

Includes the following steps:

S231: As shown in Figure 4, the low-resolution numerical model precipitation forecast is divided into several blocks, and then the divided blocks are linearly mapped; the formula is as follows:

;

in, is a group of divided tiles,/> is the weight coefficient to be trained,/> is the truncation coefficient to be trained,/> is a set of vectors after linear mapping.

S232: Use position coding to represent the position information of different blocks, and use the processed encoding information as the input of N groups of self-attention modules; where the position coding is a two-dimensional position embedding method, specifically: by comparing each block to Coding is performed on the X-axis and Y-axis positions, and different position codes are used to represent different blocks.

S233: Use spatial self-attention module to replace convolution operation. details as follows:

Let a group of divided tiles be , using three sets of weights, namely query weights/> , key value weight/> , and numerical weight/> , divide the original data into three features: query matrix/> , key value matrix/> , numerical matrix/> ;then/> Corresponding self-attention/> The formula is as follows:

;

in, for/> The square root of the dimension. The spatial self-attention module consists of a regular layer, multi-head self-attention, residual structure, and feed-forward neural network.

S3: Train the model until the error of Diffusion-Vision-Transformer converges, save the model and make predictions. details as follows:

Assume that the result obtained through steps S21-S22 is: ,in, is the model obtained in steps S21-S22,/> For low-resolution numerical model precipitation forecast,/> For paired high-resolution precipitation observation samples,/> is the preset hyperparameter in step S21, and T is the number of steps of forward noise addition in step S21; then the prediction error of the Diffusion-Vision-Transforme model in step S3/> The formula is as follows:

;

in, is a random Gaussian distribution, then/> ;

When the forecast error of the Diffusion-Vision-Transforme model When convergence occurs, the T steps are extrapolated backward until the model prediction is obtained/> ; Among them, the previous step's /> From the next step/> Obtained, the formula is as follows:

;

in, is the model obtained in steps S21-S22,/> For low-resolution numerical model precipitation forecast,/> is the hyperparameter preset in step S21,/> is a random Gaussian distribution, then/> .

An embodiment of the present invention also provides a device, including a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, it implements any one of the Diffusion-based and steps in ViT’s downscaling method.

Claims (10)

1. A downscaling method based on Diffusion and ViT, which is characterized by including the following steps: S1: Establish low-resolution numerical model precipitation forecast and high-resolution precipitation observation samples, and perform preprocessing; S2: Construct the Diffusion-Vision-Transformer precipitation forecast model; including the following steps: S21: Perform forward noise on high-resolution precipitation observation samples in the Diffusion model; S22: Use the Vision-Transformer model to extract high-order spatial features of low-resolution numerical model precipitation forecast; S23: Denoise the results obtained in step S21 in the Diffusion model, and introduce the high-order spatial features obtained in step S22 as condition information to obtain a downscaled high-resolution precipitation forecast; S3: Train the model until the error of Diffusion-Vision-Transformer converges, save the model and make predictions. 2. A downscaling method based on Diffusion and ViT according to claim 1, characterized in that, in step S1, preprocessing includes: performing logarithmic and normalization operations on the data set. 3. A downscaling method based on Diffusion and ViT according to claim 1, characterized in that the specific process of step S21 is as follows: Assume that the preprocessed high-resolution precipitation observation sample at a certain time , gradually add Gaussian noise to the original observations in T times/> ,get/> ;Then the data distribution at time t/> Previous moment/> The formula is as follows: ; in, is a preset constant hyperparameter, ranging between 0 and 1; The data distribution at the last t-th time It can be obtained from the data at time 0/> The distribution is obtained, and the formula is as follows: ; in, , and for/> , then/> . 4. A downscaling method based on Diffusion and ViT according to claim 1, characterized in that the S22 is as follows: input paired high-resolution precipitation observation samples and low-resolution numerical model precipitation forecast/> , and determine the number of steps T for forward noise addition, and the variance hyperparameter of the added random Gaussian noise/> . 5. A downscaling method based on Diffusion and ViT according to claim 1, characterized in that step S23 includes the following steps: S231: Divide the low-resolution numerical model precipitation forecast into several blocks, and then linearly map the divided blocks; S232: Use position coding to represent the position information of different blocks, and use the processed coding information as the input of N groups of self-attention modules; S233: Use spatial self-attention module to replace convolution operation. 6. A downscaling method based on Diffusion and ViT according to claim 4, characterized in that the formula of step S231 is as follows: ; in, is a group of divided tiles,/> is the weight coefficient to be trained,/> is the truncation coefficient to be trained,/> is a set of vectors after linear mapping. 7. A downscaling method based on Diffusion and ViT according to claim 4, characterized in that the position encoding in step S232 is a two-dimensional position embedding method. 8. A downscaling method based on Diffusion and ViT according to claim 4, characterized in that the step S233 is as follows: Let a group of divided tiles be , using three sets of weights, namely query weights/> , key value weight/> , and numerical weight/> , divide the original data into three features: query matrix/> , key value matrix/> , numerical matrix/> ;then/> Corresponding self-attention/> The formula is as follows: ; in, for/> The square root of the dimension. 9. A downscaling method based on Diffusion and ViT according to claim 1, characterized in that the step S3 is as follows: Assume that the result obtained through steps S21-S22 is: , where,/> is the model obtained in steps S21-S22,/> For low-resolution numerical model precipitation forecast,/> For paired high-resolution precipitation observation samples,/> is the preset hyperparameter in step S21, and T is the number of steps of forward noise addition in step S21; then the prediction error of the Diffusion-Vision-Transforme model in step S3/> The formula is as follows: ; in, is a random Gaussian distribution, then/> ; When the forecast error of the Diffusion-Vision-Transforme model When convergence occurs, the T steps are extrapolated backward until the model prediction is obtained/> ; Among them, the previous step's /> From the next step/> Obtained, the formula is as follows: ; in, is the model obtained in steps S21-S22,/> For low-resolution numerical model precipitation forecast,/> is the hyperparameter preset in step S21,/> is a random Gaussian distribution, then/> . 10. A device, including a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that when the processor executes the program, the program as claimed in any one of claims 1 to 9 is implemented. The steps in the described downscaling method based on Diffusion and ViT.