a Stochastic Approach to Multiobjective Optimization of Large-Scale Water Reservoir Networks

A main challenge for the planning and management of water resources is the development of multiobjective strategies for operation of large-scale water reservoir networks. The optimal sequence of water releases from multiple reservoirs depends on the stochastic variability of correlated hydrologic inflows and on various processes that affect water demand and energy prices. Although several methods have been suggested, large-scale optimization problems arising in water resources management are still plagued by the high dimensional state space and by the stochastic nature of the hydrologic inflows. In this work, the optimization of reservoir operation is approached using approximate dynamic programming (ADP) with policy iteration and function approximators. The method is based on an off-line learning process in which operating policies are evaluated for a number of stochastic inflow scenarios, and the resulting value functions are used to design new, improved policies until convergence is attained. A case study is presented of a multi-reservoir system in the Dalälven River, Sweden, which includes 13 interconnected reservoirs and 36 power stations. Depending on the late spring and summer peak discharges, the lowlands adjacent to Dalälven can often be flooded during the summer period, and the presence of stagnating floodwater during the hottest months of the year is the cause of a large proliferation of mosquitos, which is a major problem for the people living in the surroundings. Chemical pesticides are currently being used as a preventive countermeasure, which do not provide an effective solution to the problem and have adverse environmental impacts. In this study, ADP was used to analyze the feasibility of alternative operating policies for reducing the flood risk at a reasonable economic cost for the hydropower companies. To this end, mid-term operating policies were derived by combining flood risk reduction with hydropower production objectives. The performance of the resulting policies was evaluated by simulating the online operating process for historical inflow scenarios and synthetic inflow forecasts. The simulations are based on a combined mid- and short-term planning model in which the value function derived in the mid-term planning phase provides the value of the policy at the end of the short-term operating horizon. While a purely deterministic linear analysis provided rather optimistic results, the stochastic model allowed for a more accurate evaluation of trade-offs and limitations of alternative operating strategies for the Dalälven reservoir network.