Enhancing Dam Inflow Prediction Performance of LSTM using Improved Inflow Fluctuation of the Simple Water Balance Method by Applying Event Identification and Exponential Smoothing

The performance of LSTMs in analyzing rainfall-runoff dynamics has demonstrated better performance over the physically-based hydrological models. However, data quality remains crucial to further improve the performance of LSTM. In this study, we applied the Integration of Event Identification & Exponential Smoothing (EI & ES) methods to reduce the fluctuations in inflow estimation, which are typically calculated by the Simple Water Balance method. The preprocessed inflow data was then used to train a LSTM, demonstrating a noticeable improvement with 12 different inflow preprocessing cases in performance metrics such as NSE, RMSE, and MSE for hourly dam inflow prediction. This research presents not only new deep learning methods but also data-centric research that is critical to improving the performance of deep learning as an example of rainfall-runoff analysis.