Multistep-ahead flood forecasts by neuro-fuzzy networks with effective rainfall-run-off patterns

The purpose of this study is to construct a multistep-ahead flood forecasting model based on the precise information of the rainfall-runoff process in a watershed during typhoon events through neuro-fuzzy networks. To achieve this goal, the nonparametric Kendall trend test was implemented for identifying appropriate rainfall lag times, and then the multistep-ahead flood forecasting was carried out by the adaptive neuro-fuzzy inference system (ANFIS)-based hydrological models with different input combinations. Hydrological data collected during 13 typhoon events in the Shihmen Reservoir watershed of Taiwan were used to train and validate the forecasting models. Results reveal that rainfall and inflow had similar patterns with a time shift of 5 up to 7 h, and the ANFIS-based model with inputs that involved effective time-delayed rainfall identified by the Kendall trend test performed better than the other comparative models. Results demonstrate that accurate inflow forecasts can be achieved up to a lead time of 5 h, which is very valuable information on real-time reservoir operation for flood control.