Scalable Multiobjective Control for Large-Scale Water Resources Systems Under Uncertainty

Advances in modeling and control have always played an important role in supporting water resources systems planning and management. Changes in climate and society are now introducing additional challenges for controlling these systems, motivating the emergence of complex, integrated simulation models to explore key causal relationships and dependences related to uncontrolled sources of variability. In this brief, we contribute a massively parallel implementation of the evolutionary multiobjective direct policy search method for controlling large-scale water resources systems under uncertainty. The method combines direct policy search with nonlinear approximating networks and a hierarchical parallelization of the Borg multiobjective evolutionary algorithm. This computational framework successfully identifies control policies that address both the presence of multidimensional tradeoffs and severe uncertainties in the system dynamics and policy performance. We demonstrate the approach on a challenging real-world application, represented by the optimal control of a network of four multipurpose water reservoirs in the Red River basin in Northern Vietnam, under observed and synthetically generated hydrologic conditions. Results show that the reliability of the computational framework in finding near-optimal solutions increases with the number of islands in the adopted hierarchical parallelization scheme. This setting reduces the vulnerabilities of the designed solutions to the system's uncertainty and improves the discovery of robust control policies addressing key system performance tradeoffs.