DEVELOPMENT LSTM-BASED DEEP LEARNING MODEL FOR REAL TIME MULTIPOINT RIVER STAGE-DISCHARGE PREDICTION

Abstract

 Recent rash of severe flooding events have increased the number of victims due to delayed escapes in Japan. To reduce the victims, it is necessary to develop a system to predict the water levels and discharge efficiently at multiple points in a river basin. In this study, for the Saba River in the Yamaguchi Prefecture, we built a deep-learning model with LSTM to accurately predict water level and discharge 3 hours ahead. From the analyzed results, it was shown that the water level 3 hours ahead can be predicted with high accuracy at multiple points in the basin. On the other hand, the logarithmic conversion was applied to the discharge values to improve the learning efficiency against data of high dynamic range. With this modification, the accuracy of the model prediction was greatly improved on the peak discharge rate and delay time of occurrence compared to the case without the logarithmic conversion.