Potential of hybrid wavelet-coupled data-driven-based algorithms for daily runoff prediction in complex river basins

Accurate prediction of daily runoff's dynamic nature is necessary for better watershed planning and management. This study analyzes the applicability of artificial neural network (ANN), wavelet-coupled artificial neural network (WANN), adaptive neuro-fuzzy inference system (ANFIS), and wavelet-coupled adaptive neuro-fuzzy inference system (WANFIS) models for daily runoff prediction of Koyna River basin, India. Gamma test (GT) was used to select the best input vector to avoid the time-consuming and tedious trial and error input selection methods. Original daily rainfall and runoff time series data were decomposed into different multifrequency sub-signals using three types (Haar, Daubechies, and Coiflet) of mother wavelets. The decomposed sub-signals were fed to ANN and ANFIS as inputs for developing hybrid WANN and WANFIS models, respectively. The quantitative and qualitative performance evaluation criteria were used for assessing the prediction accuracy of developed models. An uncertainty analysis was employed to study the reliability of the developed models. It was observed that hybrid data-driven models (WANN/WANFIS) outperformed simple data-driven models (ANN/ANFIS). Finally, it was found that the Coiflet wavelet-coupled ANFIS model can be successfully applied for daily runoff prediction of the highly dynamic and complex Koyna River basin. The sensitivity analysis was also carried out to detect the most crucial variable for daily runoff prediction. The sensitivity analysis indicated that the previous 1-day runoff (Q(t-1)) is the most crucial variable for daily runoff prediction.