Data-Driven Dam Outflow Prediction Using Deep Learning with Simultaneous Selection of Input Predictors and Hyperparameters Using the Bayesian Optimization Algorithm

Reservoirs and dams are critical infrastructures that play essential roles in flood control, hydropower generation, water supply, and navigation. Accurate and reliable dam outflow prediction models are important for managing water resources effectively. In this study, we explore the application of three deep learning (DL) algorithms, i.e., gated recurrent unit (GRU), long short-term memory (LSTM), and bidirectional LSTM (BiLSTM), to predict outflows for the Buon Tua Srah and Hua Na reservoirs located in Vietnam. An advanced optimization framework, named the Bayesian optimization algorithm with a Gaussian process, is introduced to simultaneously select the input predictors and hyperparameters of DLs. A comprehensive investigation into the performance of three DLs in multistep-ahead prediction of outflow of two dams shows that all three models can predict the reservoir outflow accurately, especially for short lead-time predictions. The analysis results based on the root mean square error, Nash–Sutcliffe efficiency, and Kling–Gupta efficiency indicate that BiLSTM and GRU are the most suitable models to diagnose the outflow of Buon Tua Srah and Hua Na reservoirs, respectively. Conversely, the results of the similarity assessment of 11 hydrological signatures show that LSTM outperforms BiLSTM and GRU in both case studies. This result emphasizes the importance of determining the purpose and objective function when choosing the best model for each case study. Ultimately, these results strengthen the potential of DL for efficient and effective reservoir outflow predictions to help policymakers and operators manage their water resource system operations better.