Adaptive precipitation nowcasting using deep learning and ensemble modeling

Abstract

Accurate rainfall nowcasting is necessary for real-time flood management in urban areas. In this paper, some deep neural networks (DNNs) are developed for rainfall nowcasting with lead-time of 5 min. To increase the accuracy of the DNNs, their outputs are fused with the predictions of some numerical weather prediction (NWP) models using three ensemble models (i.e., bagging, random forest, and adaboost). The bias of NWPs is corrected using the quantile mapping technique. The ensemble models are compared with the DNNs and NWP models in terms of the accuracy of predictions. To reduce the run time of the proposed rainfall nowcasting models to be applicable for real-time urban flood management, they are executed using a parallel computing algorithm on the central processing unit (CPU) and graphics processing unit (GPU). To evaluate the efficiency of the proposed models, they were applied to the eastern drainage catchment (EDC) of Tehran city in Iran. More than 29,000 observed rainfall data were used for training and cross validating the proposed models. The results illustrated that the accuracy, coefficient of determination, of the ensemble models is at least 10% better than the DNNs in most rainfall events.