Dynamic neuro-fuzzy systems for rainfall-runoff modeling

Abstract

Urbanization has significant impact on the hydrological processes that have caused an increase in magnitude and frequency of floods; therefore, a reliable rainfall-runoff model will be helpful to estimate discharge for any watershed management plans. Beside physically-based models, the data driven approaches have been also used frequently to model the rainfall-runoff processes. Neuro-fuzzy systems (NFS) as one of the main category of data-driven models are common in hydrological time series modeling. Among the different algorithms, Adaptive network-based fuzzy inference system (ANFIS) is well-practiced in hydrological modeling. ANFIS is an offline model and needs to be retrained periodically to be updated. Therefore, an NFS model that can employ different learning process to overcome such problem is needed. This study developed dynamic evolving neuro fuzzy inference system (DENFIS) model for event based and continuous rainfallrunoff modeling and the results were compared with the existing models to check model capabilities. DENFIS evolves through incremental learning in which the rulebase is evolved after accommodating each individual new input data and benefitted from local learning implemented through the clustering method, Evolving Clustering Method (ECM). In this study, extreme events were extracted from the historical hourly data of selected tropical catchments of Malaysia. The DENFIS model performances were compared with ANFIS, the hydrologic modeling system (HECHMS) and autoregressive model with exogenous inputs (ARX) for event based rainfall-runoff modeling. DENFIS model was also evaluated against ANFIS for continuous rainfall-runoff modeling on a daily and hourly basis, multi-step ahead runoff forecasting and simulation of the river stage. The average coefficients of efficiency (CE) obtained from DENFIS model for the events in testing phase were 0.81, 0.79 and 0.65 for Lui, Semenyih and Klang catchments respectively which were comparable with ANFIS and HEC-HMS and were better than ARX. The CEs obtained from DENFIS model for hourly continuous were 0.93, 0.92 and 0.62 and for daily continuous were 0.73, 0.67 and 0.54 for Lui, Semenyih and Klang catchments respectively which were comparable to the ones obtained from ANFIS. The performances of DENFIS and ANFIS were also comparable for multistep ahead prediction and river stage simulation. This study concluded that less training time and flexibility of the rule-base in DENFIS is an advantage compared to an offline model such as ANFIS despite the fact that the results of the two models are generally comparable. However, the learning algorithm in DENFIS was found to be potentially useful to develop adaptable runoff forecasting tools.