[A-AS04] Machine Learning Techniques in Weather, Climate, Hydrology and Disease Predictions

[AAS04-P02] A reservoir inflow forecasting model using a Wavelet transformed - deep learning algorithm

Trung Duc Tran¹, *Jongho Kim¹ (1.University of Ulsan)

Keywords:Dam inflow prediction, Long short-term memory, Wavelet transform, Input predictor selection, Hyper-parameter optimization

Accurate predictions of inflow would support policy-makers and operators in better performing reservoir operation and management tasks. This study proposed a data-driven model based on deep learning algorithms Long Short-term memory, called SWLSTM, producing accurate daily dam inflow forecasting. SWLSM adopts three main ideas to improve the model's accuracy: (i) it selects an appropriate input variable and a sequence length based on the statistical properties (using partial autocorrelation functions (PACF) and cross-correlation functions (CCF)); (ii) it employs the Wavelet transform (WT) to the selected input predictors to decompose them into sub-series; (iii) it optimizes the hyper-parameters of LSTM using K-fold cross-validation and random search method. The effectiveness of SWLSTM was proved by forecasting the five dams' daily inflow in the Han River watershed (South Korea) with historical data. Different evaluation metrics (i.e., R2, NSE, MAE, PE) are used to generator evaluate the model's accuracy. Overall, SWLSTM outperformed the regular LSTM model in all cases (i.e., evaluation metrics show about 30 to 80% better performance). The results indicate that the selection of the right input variable and the sequence length is effective in reducing noise, increasing efficiency during training a model; the WT enhances the results of forecasting extreme values such as flood peaks; K –fold cross-validation, and random search support setting up model's hyper-parameters more efficient and simple. The results reinforce the potential of a data-driven model for efficient and skillful reservoir inflow forecasting in addressing water-disaster-energy security challenges.
Acknowledgment: This work was supported by KOREA HYDRO & NUCLEAR POWER CO., LTD (No.2019-Tech-11) and the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT)(NRF-2019R1C1C1004833).