Application of RandomForests for Real-time River Stage Forecasting

Chia-Wen Kuo

The purpose of this study is to use the RandomForests (RFs) method for constructing a real-time flood stage forecasting model at Yilan River basin in Taiwan. Before using RF model, the cross correlation analysis is introduced for understanding the hydrological concept of the time of response between the upstream raingagues and the river stage station for all calibrated events. Then this study did the input variables analysis and parameters analysis. Based on the aforeward analysis, the RF model is constructed and the uncertainty of RF model is analyzed. Final, the validation results revealed that (1) the RF model can reasonable forecast the flood river stage forecasts one-three-ahead, (2) the parameters analysis results is matched with original work, (3) the uncertainty analysis results is revealed that the variance of decision tree can represent the uncertainty of the RF model.

Key words: RandomForests, river stage forecasting, flood warning.