Author(s): Masayuki Hitokoto; Takeru Araki; Kenta Hakoishi; Yuto Endo

Linked Author(s):

Keywords: Dam inflow prediction; Flood prediction; Deep learning; Data augmentation; Extreme events

Abstract: Machine learning on dam inflow prediction and river water level prediction have a major problem that they are less applicable to inexperienced large-scale floods. This study aims to improve the prediction accuracy of dam inflow using machine learning by adding virtual large-scale floods as augmented data to the training data. The proposed method assumes a steady-state condition of constant rainfall and uses a theoretical dataset of virtual floods as the augmented data. The study applied deep neural networks, ridge regression, lasso regression and support vector regression (SVR) as machine learning models for four dam basins in Japan's first-class river system. For deep neural network, application of the data augmentation improved the prediction accuracy for floods about twice as large as the training floods. However, accuracy was not sufficient for floods three to four times larger than the training floods. For regression models and SVR, prediction accuracy was not as good as deep neural network model. The difference in accuracy between models is an issue to be investigated in the future. The proposed data augmentation method is easy to implement. Further research is necessary to determine the optimal flood scale and data volume for augmented data and examine the differences in applicability based on basin characteristics and historical observation data.

DOI: https://doi.org/10.3850/978-90-833476-1-5_iahr40wc-p1430-cd

Year: 2023