This study presents a generic data-driven reservoir operation model (GDROM). The hidden Markov-decision tree (HM-DT) is applied to deriving representative operation modules for a reservoir; a classification and regression tree (CART) algorithm is used to identify the application and transition conditions for the operation modules. These two procedures result in the GDROM that is featured by 1) using a few input variables (inflow, storage, DOY, and PDSI); 2) inheriting merits of decision trees but dramatically reducing model complexity; 3) adopting a consistent and transparent structure (i.e., better interpretability than other machine learning models); and 4) showing a better performance than traditional decision tree models, especially in storage simulation. GDROM is developed for 467 reservoirs with diverse operation purposes in different regions of the Contiguous United States (CONUS), and the testing procedure shows comparable accuracy in release simulation to other ML models; among these reservoirs, 15 are selected for detailed analysis with diverse operational purposes and regulation capacities, from different USGS Water Regions. GDROM presents a ready-to-use reservoir operation model that can be incorporated into a watershed hydrological simulation model.

本研究提出了一个通用的数据驱动的油藏作业模型（GDROM）。隐马尔可夫决策树（HM-DT）被应用于推导水库的代表性操作模块；分类回归树（CART）算法用于识别操作模块的应用和转移条件。这两个过程产生了具有以下特点的 GDROM 1) 使用一些输入变量（流入、存储、DOY 和 PDSI）；2）继承了决策树的优点，但大大降低了模型的复杂性；3）采用一致且透明的结构（即比其他机器学习模型更好的可解释性）；4) 表现出比传统决策树模型更好的性能，尤其是在存储模拟方面。GDROM 是针对美国本土 (CONUS) 不同地区的 467 个具有不同作业目的的油藏开发的，测试过程在释放模拟中显示出与其他 ML 模型相当的准确性；在这些水库中，选择了来自不同 USGS 水域的 15 个具有不同运营目的和调节能力的水库进行详细分析。GDROM 提供了一个现成的水库运行模型，可以将其合并到流域水文模拟模型中。