Enhancing real-time streamflow forecasts with wavelet-neural network based error-updating schemes and ECMWF meteorological predictions in Variable Infiltration Capacity model

Abstract

As a non-structural river basin management option, development of a reliable inflow forecasting system using the hydrological models is essential to aid in real-time reservoir operation. For real-time streamflow forecasting, this study evaluates the capability of the Variable Infiltration Capacity (VIC) in integration with the linear AutoRegressive (ARu), AutoRegressive Moving Average with eXogenous inputs (ARMAXu), nonlinear static wavelet-based neural network (WNNu), and dynamic wavelet-based Non-linear AutoregRessive with eXogenous inputs (WNARXu) variants of error-correction sub-models in comparison with the predictions by the standalone VIC model forced with European Centre for Medium Range Weather Forecast (ECMWF) and ERA-Interim reanalysis products. These VIC-standalone, VIC-ARu, VIC-ARMAXu, VIC-WNNu, and VIC-WNARXu frameworks are field-tested for 1- to 5-days lead-time streamflow forecasting in the Hirakud reservoir catchment of the Mahanadi River basin in eastern India. The performance evaluation measures and quantile regression (QR)-based predictive uncertainty (PU) analysis reveal that the VIC-WNARXu model is the best approach for short to medium range flood forecasting with up to 7 days lead-time having the potential to be used in other world-river basins as a component of the early flood warning system.