Short-term rainfall forecasting using multi-task learning and Weibull based postprocessing technique

Reliable short-term rainfall forecast plays a key role for flood forecasting which can prevent or mitigate life and financial loses. In this study, a new framework integrating maximum overlap discrete wavelet transforms as a data preprocessing technique, long short-term memory as deep learning algorithm with a multitask learning approach, and a postprocessing technique gaining Weibull distribution are employed to achieve reliable rainfall forecasts from 1-h up to 12-h ahead. The model inputs include real time observations from a synoptic station (Aliabad) in Golestan Province. The multitask learning approach combine continuous rainfall forecasts from regression and rainfall detection from classification. Overall, the proposed framework showed efficiency to improve probability of detection and false alarm ratio as well as correlation between forecasted and real values. The postprocessing technique was applied to improve the model forecasts for extreme events since they were generally underestimated. The results demonstrate that the proposed methodology can be successfully for rainfall forecasts within various time windows accordingly. Furthermore, it only considers real time rainfall observations as the model inputs which is promising for regions with data shortage of other parameters such as temperature and humidity.