Comparative Assessment of Artificial Neural Networks (ANNs), Long Short Term Memory Network (LSTM) and Hydrologic Engineering Centre-Hydrologic Modelling System (HEC-HMS) for Runoff Modelling

Rainfall runoff modelling is very crucial and efficient for basin management. In the current exercise, ANNs as shallow network, LSTM as Deep Learning Network and HEC-HMS as conceptual/numerical technique were used for forecasting one day in advance runoff at Shivade station located in upper Krishna basin, Maharashtra, India. Main focus of the present study is to emphasize comparative nature of the research and its goal of forecasting one-day ahead runoff using rainfall runoff process. The outcomes of two data-driven approaches (ANN and LSTM) are compared to a hard computing methodology (HEC-HMS) to determine best way for forecasting one-day runoff. Accuracy of the all models was judged by different error measures like correlation coefficient (r), Root Mean square error (RMSE), Mean Absolute error (MAE) and Coefficient of efficiency (CE). Present work is novel because model development uses limited basin data. Current work suggests that LSTM (r-0.75, RMSE-145.24 m3/s and CE-0.56) and ANN ((r-0.87, RMSE-129.78 m3/s and CE-0.63)) can be used at an operational level as a supportive (if not alternative) tool with a conceptual model like HEC-HMS which is basically based on physics of the process and it require exogeneous data. Another conclusion came from noticing similar results produced from both ANN and LSTM models is that ANN is sufficient to address the nonlinearity associated with the natural problem like rainfall runoff. Therefore it can said that additional hidden layers with larger memory structures, such as in LSTM may be required for modelling the rainfall runoff process at least in present work. However, LSTM requires additional investigations in different basins with more basin data to generalize this observation.