Applicability of Long Short Term Memory Network to Flow Discharge Modeling at Watershed in Kyushu Region, Japan

It is important for water management to accurately estimate flow discharge at a watershed with precipitation data. Machine learning approaches have been utilized for such rainfall-runoff modeling. Nowadays, new machine learning approaches, which are called deep learning, are getting more popular in many research fields. Among deep learning approaches, a recurrent-type neural network suitable for time-series analysis. Actually, a long short term memory network (LSTM), which is categorized into recurrent-type neural network was successfully applied to rainfall-runoff modeling. However, applications of LSTM for rainfall runoff modeling are still limited. LSTM has many hyper-parameters, but sensitivities of these hyper-parameters for rainfall-runoff modeling are not known well. Therefore, this study investigated the sensitives of hyper-parameters of LSTM. As a study watershed, the Kikuchi River Basin, which is located in the Kyushu region, Japan, was selected. A rainfall-runoff model with LSTM This study employed Keras, a machine learning framework for python. Only daily precipitation data were used as the inputs for rainfall-runoff modeling in this study. Then, sensitivity analysis of hyper-parameters of LSTM was conducted. The results show the great capability of LSTM for rainfall-runoff model. Meanwhile, the results of the sensitivity analysis indicate that hyper-parameters should carefully be selected in order to improve flow discharge simulations.