A novel two-stage multi-step dynamic error correction model for improving streamflow forecast accuracy

The occurrences of floods in the recent past have significantly increased due to climate change and anthropogenic activities. Hence, reliable streamflow forecasts are crucial for minimizing the detrimental effects of flooding. However, forecast accuracy deteriorates besides elevated uncertainty when the lead time increases. Therefore, streamflow forecast should have improved accuracy with simultaneous uncertainty quantification to increase the model confidence for effective decision-making. The study proposes a novel two-stage multi-step dynamic error correction model to forecast up to 7 days ahead of streamflow, with the objective of no significant deterioration in accuracy. The framework is developed by integrating the process-based hydrological HBV model with the Bayesian-based Particle filter (PF) and machine learning-based Random Forest algorithm (RF). This facilitates combining the advantages of each model, i.e., process understanding ability of the HBV model, robust uncertainty quantifying ability of the PF technique, and relatively superior predictive ability of the RF algorithm. The model performance is quantified through several statistical performance error measures and uncertainty indices, with graphical performance indicators. The framework tested on the Beas and Sunkoshi river basins of India and Nepal exemplified the NSE of 0.94 and 0.98 in calibration and 0.95 and 0.99 in validation respectively for the 7-day ahead streamflow forecast. Hence, the proposed dynamic modeling framework can be considered as a potential tool to forecast streamflow without significant deterioration in the model accuracy even at increased lead times.