Adaptation Triggers and Indicator Interpretability for Dynamic Reoptimization of Reservoir Control Policies Under Climate Change

Abstract Recent studies have shown the potential for reservoir control policies to adapt to uncertain future climate and demand by reoptimizing on a fixed time interval. However, this strategy is independent of the system evolution and might implement late or unnecessary adaptation. This study develops a framework to identify dynamic decisions on two levels: an “outer loop” adaptation policy that establishes indicator thresholds for reoptimization based on recently observed data, and an “inner loop” control policy that undergoes reoptimization according to these thresholds. We demonstrate this method for a case study of Oroville Reservoir, California, using an ensemble of climate model projections split into training and testing sets. The control policy uses inputs of storage, day of year, and a 5‐day inflow forecast, while the adaptation policy indicators include long‐term statistics of climate and demand as well as the recent system performance. Both policy levels are optimized simultaneously using heuristic policy search and analyzed with policy interpretation methods, including Shapley Additive Explanations (SHAP) and global sensitivity analysis. Results show that the adaptation solutions provide equal or better performance compared to the historical benchmark and are robust to out‐of‐sample scenarios. Additionally, the decision to reoptimize is primarily driven by demand, flood cost and mean annual flow indicators on different timescales. The proposed methodology identifies how control policy reoptimization can be initiated using observed thresholds of climate, demand, and system performance to improve adaptation under future uncertainty.