Comparing the Robustness of Forecast-Informed Reservoir Operating Policies under Forecast Uncertainty and Hydrologic Extremes

Forecast-informed reservoir operations (FIRO) can enhance water storage while reducing flood risk, but it remains a challenge to design policies that are robust to forecast uncertainty and out-of-sample extreme events. In this study, we develop a framework to evaluate the robustness of FIRO policies using medium-range (1–14 day) synthetic forecast ensembles and scaled flood events. We compare three policy approaches: ensemble forecast operations (EFO) with an optimized risk curve; model predictive control (MPC); and a guide curve based on the cumulative inflow forecast. Different EFO policies are trained using all available hindcasts, with specific extreme events withheld, and with synthetic forecasts. These policies are benchmarked against a baseline policy with no forecast and a policy assuming perfect forecast information. We demonstrate the framework using a case study of Oroville Reservoir, CA, to determine the required flood pool for scaled 100-, 200-, and 300-year events under each policy for both hindcasts and synthetic ensembles. Results show that EFO and MPC require smaller flood pools than the cumulative method, with limited impact on water supply. Synthetic-trained EFO policies show the highest robustness to forecast uncertainty, which suggests that training on synthetic forecasts can help avoid overfitting to a single forecast ensemble. Furthermore, excluding the largest historical flood from training does not diminish EFO policy performance given a sufficiently low priority weight. This study offers a template for comparing the effectiveness of different policy formulations in reducing flood risk, potentially allowing for greater water storage and supporting the broader adoption of forecast-informed policies.