| description abstract | Hydropower release decision making relies on multisource information, such as climate conditions, downstream water quality, inflow and storage, regulation and engineering constraints, and so on. The decision tree (DT) method is one of the commonly used techniques to simulate reservoir operation and release strategies because of its simplicity and effectiveness. However, the performances and simulation accuracy vary among different DT models due to many structures and splitting rules associated with each DT model. In this study, we propose a dynamic merge technique (DMerge), which adopts a concept from particle swarm optimization, to postprocess outputs from different DT models with the purpose of increasing the simulation accuracy and producing a model ensemble with dynamically changing weights throughout the validation phase. A case study of Shasta Lake in northern California is presented, where the daily hydropower releases are predicted and compared using the DMerge, AdaBoost DT, random forest, and extremely randomized trees methods. Results show that the DMerge method has the best statistics compared to other popular DT algorithms. Furthermore, scenario tests were carried out to analyze the sensitivity to model inputs (i.e., hydrological condition, reservoir storage and regulation, climate phenomenon indices, and water quality) with respect to explaining the variability of hydropower releases. According to the results, we found that the hydropower releases are a complex decision-making process and water quality and climate conditions could play an even more significant role than both hydrological forcing and system states in our case study. The proposed DMerge method is a robust and efficient technique in solving water-energy prediction and simulation problems, and it is suitable for joint use with other data-driven approaches. | |
