Daily Streamflow Forecasting in the Mahanadi River Basin using a Novel Deep Learning-based Model

Flood is one of the most devastating natural disasters accounting for the loss of life and property of millions of people every year. Since 2000s, floods have become more frequent in some parts of the world, especially in the tropical region. In India, many frequent extreme floods are found to occur recently. While the structural measures of flood management are not always feasible, the non-structural measures, such as flood forecasting plays a vital role in developing early flood warning systems. In the present study, a novel deep learning model, namely Smoothing-based Long Short-Term Memory (Smooth-LSTM) model is developed for daily streamflow forecasting at the head of the delta region in the Mahanadi River basin, eastern India. This modelling framework integrates smoothing filters and the traditional LSTM networks to predict the daily streamflow foreacasts up to 5-days lead-time. This model follows a sequence-to-single output approach, with the time-lagged streamflows as the only input variable. The Smooth-LSTM model is able to predict the streamflows reasonably well with a Nash-Sutcliffe Efficiency of 0.87–0.82 up to a lead-time of 5-days. The overall model performance is found to be satisfactory with the ability to capture the observed streamflows within the 90% uncertainty bands.