EVALUATION OF THE APPLICABILITY OF DATA AUGMENTATION FOR DAM INFLOW PREDICTION USING DEEP LEARNING

Abstract

 We attempted to improve the accuracy of dam inflow prediction using deep learning by augmenting the training data and validated the applicability of the proposed method to multiple basins in Japan. The proposed data augmentation method assumes a steady-state condition of constant rainfall and uses a virtual rainfall-runoff dataset that is based on the theory of hydrology as the augmentation data such that the total rainfall and dam inflow into the watershed are equal (runoff coefficient = 1.0). The applicability of the data augmentation method to recent large-scale floods was validated in case studies for four dam basins: the Terauchi, Miyagase, Nomura, and Kanayama dams. The prediction accuracy of the data augmentation method was confirmed for each dam. However, when the test flood was much larger than the training floods, the prediction improvement was limited.