Abstract
A novel smoothing-based long short-term memory (Smooth-LSTM) framework for up to 5 days ahead flood forecasting is proposed, and compared with the benchmark LSTM (LSTM) model, an ANN model, and the conceptual MIKE11 NAM-HD (MIKE) hydrological model. This framework was tested in the typical middle Mahanadi River basin (India) having a tropical monsoon-type climatology. Variation of training loss indicated that the LSTM network has a higher learning ability at smaller network and batch sizes. The Smooth-LSTM model could predict the streamflow with higher Nash-Sutcliffe efficiency of 0.82–0.87 up to 5 days lead-time with a better reproduction of the observed crucial high peak floods; whereas the corresponding MIKE, ANN and LSTM model-based forecasts were acceptable only up to 4, 3 and 1-day lead-times, respectively. Overall, the Smooth-LSTM model is found to be robust in operational flood forecasting characterized with lower uncertainty and the least sensitivity to any redundant input information.

This publication has 108 references indexed in Scilit: