Global Parameter Sensitivity in Forecast‐Informed Reservoir Operations Using Model Predictive Control

Abstract Hydrologic forecasts have demonstrated considerable benefits for flood risk reduction and water supply operations in specific basins. However, it remains unclear how these benefits depend on key parameters of the reservoir infrastructure and forecasts. In this study, we propose a global sensitivity and threshold analysis framework using Model Predictive Control (MPC) to evaluate the ability of forecast‐informed operations to safely manage large floods while also improving water supply storage. The approach is demonstrated for six major reservoirs in California using daily ensemble hydrologic forecasts. We first benchmark MPC performance with ensemble forecasts against a baseline policy (no forecast) and a perfect forecast policy. We then perform global sensitivity analysis using Decision Tree (DT) classifiers to assess the impact of maximum release, ramping rate, and available forecast lead time on the required flood pool for the event of record across all sites. This large parameter sample reflects many potential reservoir configurations and scaled events. We find that the ensemble forecast release sequence captures most of the value of a perfect forecast release sequence, but this value is highly sensitive to parameters. The classifiers accurately predict the extent of flood pool reduction and identify maximum release as the most important feature for both ensemble and perfect forecast release sequences. Finally, we extract generalized parameter thresholds from the DT classifier across all reservoirs, identifying conditions required to achieve varying levels of flood pool reduction. Overall, this framework represents an efficient screening tool to identify system properties where forecast‐informed policies likely provide the greatest benefits.