Improving ECMWF short-term intensive rainfall forecasts using generative adversarial nets and deep belief networks

Short-term intensive rainfall (3-h rainfall amount > 30 mm) is a destructive weather phenomenon that is poorly predicted using traditional forecasting methods. In this study, we propose a model using European Center for Medium-Range Weather Forecasts (ECMWF) data and a machine learning framework to improve the ability of short-term intensive rainfall forecasting in Fujian Province, China. ECMWF forecast data and ground observation station data (2015-2018) were interpolated using a radial basis function, outliers were processed, and the data were blocked according to the monthly cumulative rainfall and forecast window. Subsequently, the box difference index was used to select features for each data block. As short-term intensive rainfall events are rare, a data processing method based on the K-means and generative adversarial nets was used to address data imbalances in the rainfall distribution. Finally, focal loss object detection was combined with a deep belief network to construct the short-term intensive rainfall classification model. The results show that the data preprocessing method and resampling method used in this study were effective. Furthermore, the classification model was superior to other machine learning methods for predicting short-term intensive rainfall.

Automated summary

This study proposed a model using ECMWF data and a machine learning framework to improve the forecasting of short-term intensive rainfall in Fujian Province, China. The data preprocessing and resampling methods used in this study were effective. The classification model combining focal loss object detection and deep belief network outperformed other machine learning methods in predicting short-term intensive rainfall.