Multi-Step-Ahead Rainfall-Runoff Modeling: Decision Tree-Based Clustering for Hybrid Wavelet Neural- Networks Modeling

Abstract

This paper introduces a novel hybrid approach for predicting the rainfall-runoff (r-r) phenomenon across different data division scenarios (50%-50%, 60%-40%, and 75%-25%) within two distinct watersheds, encompassing both monthly and daily scales. Additionally, the effectiveness of this newly proposed hybrid method is evaluated in multi-step ahead prediction (MSAP) scenarios. The proposed method comprises three primary steps. Initially, to address the non-stationarity of the runoff and rainfall time series, these series are decomposed into multiple sub-time series using the wavelet (WT) decomposition method. Subsequently, in the second step, the decomposed sub-series are utilized as input data for the M5 model tree, a decision tree-based model. The M5 model tree classifies the samples of decomposed runoff and rainfall time series into distinct classes. Finally, each class is modeled using an artificial neural network (ANN). The results demonstrate the superior efficiency of the proposed WT-M5-ANN method compared to other available hybrid methods. Specifically, the calculated R 2 was 0.93 for the proposed WT-M5-ANN method, whereas it was 0.89 and 0.81 for the WT-ANN and WT-M5 methods, respectively, for the Lobbs Hole Creek watershed at the daily scale.