Learning What Information Matters and When: Concurrent Input Selection and Policy Optimization in Multipurpose Reservoir Operations

Abstract

Managing multi‐purpose reservoirs requires balancing flood protection, water supply, and ecosystem needs under growing uncertainty. A critical challenge is deciding what information to use and when: forecasts exist across multiple variables and lead times, yet their operational value depends on both the management objectives and the policy structure. Here, we introduce a unified framework that learns optimal reservoir operating policies while simultaneously identifying the most valuable information to inform them. By extending neuro‐evolutionary multi‐objective optimization, our method evolves policy parameters, network topology, and input set in a single process. To interpret the resulting “black‐box” strategies, we leverage an adapted Time‐Varying Sensitivity Analysis (TVSA) to characterize the influence of each input throughout the annual cycle. We demonstrate the framework on Lake Como, a regulated lake in Northern Italy with competing objectives of flood control, irrigation supply, and ecological preservation. The framework automatically discovers policies with performance matching, and sometimes surpassing, state‐of‐the‐art methods that rely on expert‐selected inputs. Crucially, it does so without prior knowledge, within a single holistic optimization. TVSA reveals that policies adaptively shift their reliance on different forecast horizons: short‐term information dominates flood mitigation in spring, while medium‐range signals help balance irrigation and navigation in summer. When trained on real‐world forecasts, the policies prioritize accurate short‐term predictions while discarding uncertain long‐range signals, highlighting the gap between current forecasting capabilities and their operational potential. Our results demonstrate how concurrent input and policy optimization can transform Forecast‐Informed Reservoir Operations, offering a generalizable framework to discover what information matters and when.