Application of the long short-term memory network for hydrologic time series forecasting

Hydrologic time series forecasting is still a challenging task in water resources management under complex weather condition. Many works including statistical, dynamic, and data-driven approaches associated with dam inflow forecasting have been developed to handle the complexity in the water resources management and planning. Deep learning algorithms are emerging as the most advanced approach in the hydrologic time series forecasting. Therefore, it is necessary to identify the applicability of the deep learning algorithms to the hydrologic time series data. Recently, the applicability of the long short-term memory (LSTM) network has been remarkable in the hydrologic field. The LSTM network is a special kind of recurrent neural network (RNN) which has connections between neurons and form a directed cycle. The LSTM designed to overcome the weakness of the RNN to learn long-term dependencies. In this study, we applied the LSTM network to the monthly dam inflow in South Korea for identifying the applicability in the hydrologic time series forecasting. In addition, the performance of the LSTM network was compared to the conventional hydrologic time series forecasting models. Finally, we drew several conclusions including the optimal conditions for applying the LSTM network in the monthly dam inflow forecasting.