Midterm drought forecasting based on dam storage prediction using deep learning algorithms

Abstract

Effective drought preparedness depends on sufficient lead time to secure water resources and implement adaptive response strategies. In South Korea, three-month drought forecasts are issued at the beginning of each month under the supervision of the Ministry of the Interior and Safety; however, these forecasts remain insufficient to mitigate the risk of persistent drought conditions. Accordingly, this study aims to improve midterm drought forecasting by extending the prediction horizon to six months and modeling dam storage dynamics using deep learning algorithms, thereby enhancing proactive drought preparedness. To achieve this objective, three training datasets were developed through feature engineering applied to observational hydrological data and then combined with four deep learning architectures (Long Short-Term Memory (LSTM), Bidirectional LSTM (Bi-LSTM), Temporal Convolutional Network (TCN), and Convolutional Neural Network–LSTM (CNN-LSTM)), resulting in 12 predictive models. Among the evaluated models, the configuration combining the TCN architecture with training dataset B exhibited the highest performance, and its hyperparameters were subsequently optimized via a stepwise grid search procedure. Evaluation of the final midterm prediction model with optimized hyperparameters yielded an average Root Mean Squared Error (RMSE) of 35.440 and an average error rate of approximately 6.23% relative to total storage, thereby demonstrating robust quantitative predictive performance. Furthermore, qualitative evaluation indicates that the model effectively reproduces temporal storage variability patterns, including those observed during severe drought conditions.