Author(s): Yusuke Nakatani; Motoki Okumura; Shuzo Nishida
Linked Author(s):
Keywords: River water level; XRAIN; CNN; RNN; Deep learning
Abstract: Physical models available for deductively predicting river water levels require a large amount of input data in order to set their boundary conditions and parameters. In this study, we developed a deep learning model comprising a convolutional neural network (CNN) and a recurrent neural network (RNN) with the objective of predicting river water levels based on rainfall distribution data (XRAIN). It was expected that the water level could be predicted using only rainfall data as input data because CNN and RNN can extract spatial and temporal features inherent in the input data, respectively. The developed model was applied to a case study of the Katsura River basin, Kyoto, and we found that the response of the water level to rainfall could be accurately reproduced. The RMSE was 0.23 m, and the coefficient of determination was 0.91. Nonetheless, with respect to the accuracy of reproduction, a small number of problems, such as unstable outputs in the case of ordinary water levels and an underestimation of peak water levels, could be observed. The accuracy of prediction decreased when the spatial information was not considered as a part of the input data. Furthermore, the accuracy decreased when RNN was not used because, in this case, rainfall history was not considered. The study results demonstrate the necessity of using both CNN and RNN for predicting river water level based on rainfall data.
Year: 2020