Two novel error-updating model frameworks for short-to-medium range streamflow forecasting using bias-corrected rainfall inputs: Development and comparative assessment

Abstract

Accurate inflow forecasts with sufficient lead-time are highly crucial for efficient reservoir operation, for which, this study advocates the popular MIKE11-NAM-HD (MIKE) standalone conceptual hydrological model and its two error-updating frameworks for 1–5 days ahead inflow forecasting. These two modelling frameworks integrate the recently developed Wavelet-based Nonlinear AutoRegressive neural network with eXogenous inputs (WNARX) and a novel nested smoothing-based Long Short-Term Memory (LSTM) network as the error-updating models, with the standalone MIKE11-NAM-HD model, respectively. The available rainfall forecasts are bias-corrected using the Gaussian and Archimedean copulas, enhanced Kohonen Self-Organizing Maps (eKSOM) and hybrid Copula–eKSOM (Cop-SOM) approaches. The MIKE11-NAM-HD forced with the raw India Meteorological Department–Multi-model Ensemble (IMD-MME) and the selected bias-corrected rainfall forecasts are, hereafter named as rMIKE and bcMIKE, respectively. These proposed frameworks are tested for daily streamflow forecasting into the Hirakud multi-purpose reservoir in eastern India. The results reveal that the hybrid Cop-SOM could provide improved rainfall forecast skills with up to 5 days lead-time as compared to the standalone Copula and eKSOM. The standalone rMIKE and bcMIKE forecasting frameworks are useful for up to 2 days lead-time; whereas the bcMIKE-WNARX and bcMIKE-LSTM error-updating frameworks are useful up to 5 days lead-time with the Nash-Sutcliffe Efficiency (NSE) of 0.67–0.88 and 0.81–0.92, respectively. Overall, the LSTM proves to be a robust error-forecasting model at 1–5 days lead-time with reliable reproduction of peak flows; and the MIKE-LSTM framework forced with the Cop-SOM based bias-corrected rainfall forecasts has the lowest model prediction uncertainty with the narrowest confidence bands.