Comparison of a fuzzy control and the data-driven model for flood forecasting

Abstract

A novel hybrid of adaptive neuro-fuzzy inference system (ANFIS) and two-dimensional Mamdani fuzzy controller was developed for accurately forecasting the water level at the Three Georges Reservoir during the flood season in China. Using statistical approaches, nine input variables were selected based on the upper water levels in the reservoir and the quantity of interval rainfall. Since rainfall is an important input variable in flood forecasting during the flood season, ANFIS was modified to account for the influence of rainfall. Two sub-models were written, ANFIS 1 with rainfall and ANFIS 2 without rainfall, due to the weak cross-correlation function between the interval rainfall and the forecasted water levels. These two sub-models were trained by adjusting the number of the membership functions and the fuzzy rules. The number of membership functions and fuzzy rules was as selected 5 for ANFIS1, because of the relatively better results obtained based on the evaluation criterion in comparison with the other groups. The two-dimensional Mamdani fuzzy controller was regarded as an updating process for ANFIS forecasting, which controlled the error rate between the observed and forecasted amounts to within 0.05 %. The final forecasted results were acquired through error feedback and proved to be very close to the observed results. These results verified that this novel model has accurate predictive capabilities.