Assessing the Use of Generative Adversarial Networks (GAN) for Augmenting Daecheong Dam Inflow Time-Series Data

This study examined the effectiveness of time-series data augmentation using a Generateve Adversarial Network (GAN) in predicting inflow to the Daecheong Dam. A total of 3,287 observation records from 2013 to 2021, including dam inflow and meteorological, hydrological, and hydraulic variables, were used as the training data. In addition, 1,096 synthetic samples were generated using the GAN, resulting in 4,383 training samples. A multilayer perceptron (MLP) was applied to predict the 2022 inflow, and model performance was evaluated using the mean squared error (MSE). The results showed that the MSE decreased significantly from 6,975.97 (m<sup>3</sup>/s)<sup>2</sup> when usign only observation data to 3,440.20 (m<sup>3</sup>/s)<sup>2</sup> when GAN-augmented data were included. These findings indicate that GAN-based time-series data augmentation can enhance the prediction stability during data-scarce flood periods and suggest its potential applicability in the field of water resource management.