Dam inflow prediction using large-scale climate variability and deep learning approach: a case study in South Korea

Accurate prediction of dam inflows is essential for effective water resources management in terms of both water quantity and quality. This study aims to develop a Long Short-Term Memory (LSTM) deep learning-based monthly dam inflow prediction model using large-scale climate indices. Six climate indices, Atlantic multidecadal oscillation (AMO), El Nino-southern oscillations (ENSO), North Atlantic oscillation (NAO), Pacific decadal oscillation (PDO), Nino 3.4, and Southern Oscillation Index (SOI) for the period of 1981-2020, were used as input variables of the model. The proposed model was trained with 29 years of data (1981-2009) and tested with 12 years of data (2009-2020). We investigated 29 input data combinations to evaluate the predictive performance according to different input datasets. The model showed the average values of metrics ranged from 0.5 to 0.6 for CC and from 40 to 80 cm for root mean square error (RMSE) at three dams. The prediction results from the model showed lower performance as the lead time increased. Also, each dam showed different prediction results for different seasons. For example, Soyangriver/Daecheong dams have better accuracy in prediction for the wet season than the dry season, whereas the Andong dam has a high prediction ability during the dry season. These investigations can be used for better efficient dam management using a data-driven approach.