QPESUMS Real-time Rainfall Forecasting Using Machine Learning Techniques

Szu-Yin Chen

The purpose of this study is to develop a real-time rainfall forecasting model to predict lead-time 1~3 hours rainfall by using two machine learning techniques, Random Forests (RF) and Support Vector Machine (SVM), with QPESUMS data as input. In this study, the data were collected from six typhoon events for three various reservoir catchments (i.e., Feitsui, Deji, and Zengwen reservoir catchment). A number of variables including QPESUMS data, gird XY position, grid elevation and typhoon information were examined for finding suitable input variables in rainfall forecasting. Besides, two model structures were also tested: (1) single-mode and (2) multiple-mode for deciding a better model structure. Based on two performance indexes (i.e., correlation coefficient and root mean squared error), the results suggest that the single-mode model structure gives a better performance and both machine learning techniques show reasonable performance. However, for lead-time 2~3 hours, RF will underestimate in low rainfall amount and SVM will overestimate or underestimate in high rainfall amount. Moreover, because of the extended of prediction time, time lag and the decrease of accuracy occur. Overall, SVM-based rainfall forecasting model can give better results than RF-based model. In order to verify the ability and reliability of the proposed real-time rainfall forecasting model, the study compared the predictions and rain gauge data. The results indicate that SVM performs better than RF.

Key words: QPESUMS, Machine Learning Techniques, Real-time Rainfall Forecasting