Reservoir control policies provide a flexible option to adapt to the uncertain hydrologic impacts of climate change. This challenge requires robust policies capable of navigating scenarios that are wetter, drier, or more variable than anticipated. While a number of prior studies have trained robust policies using large scenario ensembles, there remains a need to understand how the properties of training scenarios impact policy robustness. Specifically, this study investigates scenario properties including annual runoff, snowpack, and baseline regret—the difference between baseline policy and perfect foresight performance in an individual scenario. Results indicate that policies trained to scenario subsets with high baseline regret outperform those generated with other training sets in both wetter and drier futures, largely by adopting an intra-annual hedging strategy. The approach highlights the potential to improve the efficiency and robustness of policy training by considering both the hydrologic properties and baseline regret of the training ensemble.

水库控制政策提供了一个灵活的选择，以适应气候变化带来的不确定的水文影响。这项挑战需要强大的策略，这些策略必须能够处理比预期更潮湿，更干燥或更可变的场景。尽管许多先前的研究已经使用大型场景集合训练了健壮的策略，但是仍然需要了解训练场景的属性如​​何影响策略的健壮性。具体而言，本研究调查了情景属性，包括年度径流，积雪和基准后悔—基准策略与单个情景中完美的预见绩效之间的差异。结果表明，针对基线后悔程度较高的情景子集进行训练的策略要优于在较干旱和较干旱的期货中通过其他训练集生成的策略，主要是通过采用年度内对冲策略。该方法通过同时考虑水文特性和培训组合的基线遗憾，突出了提高政策培训效率和稳健性的潜力。