Development of a Deep Learning-based Error-Updating Model for Improved Streamflow Forecasting Accuracy of a Hydrological Model

Recent advancements in the deep learning models, such as the Long Short-Term Memory (LSTM) networks are gaining popularity in hydrological applications.In this study, an LSTM-based error-updating model is developed to forecast the streamflow prediction errors of a hydrological model, namely, the MIKE11-NAM-HD (MIKE).The daily raw rainfall forecasts from the ensemble rainfall forecast products (IMD-MME) of the India Meteorological Department (IMD) up to a 3-days lead-time are bias-corrected using a hybrid copula-enhanced Kohonen Self-Organizing Mapbased bias-correction technique.Both the raw and bias-corrected rainfall forecasts, along with the evapotranspiration forecasts are forced as the meteorological inputs to the MIKE model for the upper reaches of the Mahanadi River basin in eastern India.A smoothing-based LSTM (sLSTM) error-updating model is trained and tested using the errors in the MIKE-simulated streamflows.A nested approach is followed in developing the sLSTM model to obtain improved accuracy in the daily streamflow forecasts.The results indicate inter-comparable performance of the MIKE-LSTM models with the raw (NSE=0.92)and bias-corrected (NSE=0.92)rainfall forecasts at 1-day lead time.However, as the lead-time increases from 2-3 days, the performance of the MIKE-LSTM model with the raw rainfall forecasts deteriorates to produce a Nash-Sutcliffe Efficiency (NSE) of 0.87 (2-day) and 0.67 (3-day).The MIKE-LSTM model with the bias-corrected rainfall forecasts outperforms with an NSE of 0.90 and 0.87 at 2-day and 3-day lead-times, respectively.Along with the overall time-series, the MIKE-LSTM model forced with the bias-corrected rainfall forecasts is also able to capture the annual maximum peaks in the testing period with reasonable accuracy.