A robust recurrent ANFIS for modeling multi-step-ahead flood forecast of Three Gorges Reservoir in the Yangtze River

Abstract. Accurate and robust multi-step-ahead flood forecast during flood season is extremely crucial to reservoir flood control. A modified hybrid learning algorithm, which fuses the Least Square Estimator (LSE) with Genetic Algorithm (GA), is proposed for optimizing the parameters of recurrent ANFIS (R-ANFIS) model to overcome the instability and local minima problems as well as improve model’s generalization and robustness. A coherent set of evaluation criteria is used to fully explore the model's accuracy (MAE, RMSE, CC & CE) and robustness (reliability, vulnerability & resilience). Three types of ANFIS (i.e. Classic, Recurrent, and Modified Recurrent) models with their optimal input variables identified by the Gamma Test are utilized for modeling multi-step-ahead flood forecast of Three Gorges Reservoir in the Yangtze River, respectively. Taking the horizon t + 12 (three days ahead), for example, the comparison analysis between C-ANFIS and R-ANFIS indicates that the R-ANFIS model can largely improve the CE, CC, reliability and resilience by 38.09 %, 17.36 %, 28.30 % & 140.26 % as well as significantly reduce the MAE, RMSE, vulnerability by 68.03 %, 47.98 % & 13.32 %. The comparison analysis between R-ANFIS and MR-ANFIS shows that the MR-ANFIS model can further enhance the CE, CC, reliability and resilience by 2.04 %, 2.04 %, 5.05 %, and 3.61 %, respectively, as well as decrease the MAE, RMSE, vulnerability by 9.91 %, 13.79 %, and 9.92 %, respectively. Such results evidently promote data-driven model's generalization (accuracy & robustness) and leads to better decisions on real-time reservoir operation during flood season.