Author(s): Masayuki Hitokoto; Takeru Araki

Linked Author(s):

Keywords: Deep learning; Dam inflow prediction; Data augmentation; Extreme floods

Abstract: There are many studies on dam inflow prediction and river water level prediction using machine learning. However, these methods have a major problem that they are less applicable to inexperienced large-scale floods. The authors have been proposed to improve the prediction accuracy of dam inflow prediction using machine learning by data augmentation based on the theory of rainfall-runoff. The proposed method improves the accuracy of the inflow prediction model by adding virtual large-scale floods asaugmented data to the learning data. Specifically, we assume a steady-state condition of constant rainfall, and use atheoretical dataset of virtual floods as the augmented data, such that the total amount of rainfall in the basin is equalto the dam inflow. In this study, we studied the applicability of the proposed method in seven dam basins in Japan. Case studies were conducted on several flood events, including these large-scale floods. We used a feed forward deep neural networkas the machine learning model. For all dams, it was confirmed that the prediction accuracy improved by applying the proposed method. We analyzed the hydrological characteristics of past flood events used for learning, and analyzed under what conditions the proposed method would be most effective.

DOI: https://doi.org/10.3850/iahr-hic2483430201-276

Year: 2024