Correcting rainfall forecasts of a numerical weather prediction model using generative adversarial networks

Abstract

In recent years, the use of deep learning techniques to forecast the weather has increased significantly; however, existing machine learning methods based on observed data are only suitable for very short-term forecasting. Numerical models are more stable for short- and medium-term forecasting, but the results may deviate from the observed data. This study proposes a deep learning method to improve the performance of numerical weather prediction models. In this method, the transformation relationship between the output of the numerical model and the observed data is learned by a generative adversarial network, which is then used to correct the forecasts of the numerical model. Experiments on 9 months of paired numerical model data and observed radar data demonstrate that correction of the forecast data using this method improves prediction performance, especially of heavy rainfall events. The proposed method provides a practical approach to combining conventional numerical weather prediction with data-driven deep learning models.