Improving the protection of aquatic ecosystems by dynamically constraining reservoir operation via direct policy conditioning

Water management problems generally involve conflicting and non-commensurable objectives. Assuming a centralized perspective at the system-level, the set of Pareto-optimal alternatives represents the ideal solution of most of the problems. Yet, in typical real-world applications, only a few primary objectives are explicitly considered, taking precedence over all other concerns. These remaining concerns are then internalized as static constraints within the problem's formulation. This approach yields to solutions that fail to explore the full set of objectives tradeoffs. In this paper, we propose a novel method, called direct policy conditioning (DPC), that combines direct policy search, multi-objective evolutionary algorithms, and input variable selection to design dynamic constraints that change according to the current system conditions. The method is demonstrated for the management problem of the Conowingo Dam, located within the Lower Susquehanna River, USA. The DPC method is used to identify environmental protection mechanisms and is contrasted with traditional static constraints defining minimum environmental flow requirements. Results show that the DPC method identifies a set of dynamically constrained control policies that overcome the current alternatives based on the minimum environmental flow constraint, in terms of environmental protection but also of the primary objectives.