A fundamental challenge in numerical weather prediction is to efficiently produce high-resolution forecasts. A common solution is applying downscaling methods, which include dynamical downscaling and statistical downscaling, to the outputs of global models. This work focuses on statistical downscaling, which establishes statistical relationships between low-resolution and high-resolution historical data using statistical models. Deep learning has emerged as a powerful tool for this task, giving rise to various high-performance super-resolution models, which can be directly applied for downscaling, such as diffusion models and Generative Adversarial Networks. This work relies on a diffusion-based downscaling framework named CorrDiff. In contrast to the original work of CorrDiff, the region considered in this work is nearly 40 times larger, and we not only consider surface variables as in the original work, but also encounter high-level variables (six pressure levels) as target downscaling variables. In addition, a global residual connection is added to improve accuracy. In order to generate the 3km forecasts for the China region, we apply our trained models to the 25km global grid forecasts of CMA-GFS, an operational global model of the China Meteorological Administration (CMA), and SFF, a data-driven deep learning-based weather model developed from Spherical Fourier Neural Operators (SFNO). CMA-MESO, a high-resolution regional model, is chosen as the baseline model. The experimental results demonstrate that the forecasts downscaled by our method generally outperform the direct forecasts of CMA-MESO in terms of MAE for the target variables. Our forecasts of radar composite reflectivity show that CorrDiff, as a generative model, can generate fine-scale details that lead to more realistic predictions compared to the corresponding deterministic regression models.
Gridded meteorological forecasts are important in various fields such as transportation, energy sector, agriculture, and scientific research. In particular, high spatial resolution forecasts are crucial for local studies and risk assessment. Traditionally, global gridded forecasts are obtained from numerical weather prediction models. Generating km-resolution forecasts is challenging for numerical weather prediction models due to the limitation of computation time. Currently, there are also data-driven deep learning models that generate global forecasts [2,13,16,6]. However, up to now, there are few global high-resolution gridded reanalysis data that can be used to train such data-driven models. Instead of getting global high-resolution forecasts, a practical alternative is to get regional high-resolution forecasts from the low-resolution output of a global model by downscaling.
Downscaling methods can be categorized into three types: dynamical downscaling, statistical downscaling, and combined methods. Dynamical downscaling, similar to numerical weather prediction models, is based on a set of atmospheric dynamical equations. It derives high-resolution forecasts by solving these equations using initial and lateral boundary conditions provided by global models. Statistical downscaling establishes statistical relationships between lowresolution variables of global models and high-resolution variables using historical data. These relationships are then applied for future forecasting. Compared to dynamical downscaling, statistical downscaling offers advantages, including simpler implementation, lower computational cost, and potentially higher accuracy.
Many machine learning models have been applied for statistical downscaling, such as multiple linear regression [17], support vector machine [5], random forest [7], and artificial neural networks [12]. Among these models, artificial neural networks, which have evolved into deep learning [19], are likely the most promising approach.
The fast development in deep learning over the past decade has led to numerous active research areas, including super-resolution in computer vision. Various high-performance super-resolution models have been proposed, such as Generative Adversarial Networks (GANs) and diffusion models. Super-resolution is similarity to downscaling: the input for both is a low-resolution grid, and the output is a high-resolution grid. However, there is also a difference between super-resolution and downscaling: the goal of super-resolution is to generate visually realistic images, while meteorological downscaling must ensure accuracy and physical consistency. Given the similarity between super-resolution and downscaling, many researchers have investigated the application of such superresolution models to downscaling [21,1]. In parallel, there are also works that developed specific neural network structures for downscaling [22].
This work investigates a diffusion-based downscaling model named Corrective Diffusion (CorrDiff) [15]. CorrDiff is a two-step approach that includes the training of a regression model and the training of a diffusion model to improve the predictions of the regression model. In [15], CorrDiff is applied to the Taiwan region, the resolution of the inputs is 25km and the resolution of the outputs is 2km. The size of the 2km high-resolution grid is 448 × 448. In this study, we apply CorrDiff on the China region, the resolutions of the inputs and the outputs are 25km and 3km respectively. The size of our 3km high-resolution grid is 1600 × 2400, which is nearly 20 times the size in [15]. Our models are trained on reanalysis data, including 25km ECMWF Reanalysis v5 (ERA5) [9] (as low-resolution inputs of the downscaling models) and 3km reanalysis data (as high-resolution labels) that are produced by China Meteorological Administration Regional Reanalysis Atmospheric System (CMA-RRA). In contrast to [15], which focuses mainly on surface variables, in this work multiple variables are considered for downscaling, including surface variables and variables at six pressure levels. The prediction of radar composite reflectivity is also investigated. Different combinations of input and output variables are examined in order to understand the intervariable dependencies in the downscaling task. By connecting our downscaling models to global forecasts, 3km regional forecasts for the China region are obtained, which are evaluated through comparison with CMA-MESO. CMA-MESO is a high-resolution regional numerical weather prediction model of the China Meteorological Administration (CMA) that generates 3km and 1km resolution forecasts of the China region. Two global forecasts are considered: CMA-GFS, which is an operational global model of CMA, and SFF, which is a deep learning-based weather model. The experimental results demonstrate that, for the mean absolute error (MAE), our forecasts generally outperform those of CMA-MESO for the target variables
This content is AI-processed based on open access ArXiv data.
