Advancing the value and use of forecasts in multi-objective reservoir control: a case study on the Lower Susquehanna River

Recent advances in Evolutionary Multi-Objective Direct Policy Search (EMODPS) have enabled the identification of a Pareto approximate set of operational policies in complex multi-purpose reservoir systems. EMODPS optimizes a set of global approximators (e.g. radial basis functions, RBFs) that determine release policies based on state variables of the system, and more recently, forecast information. RBFs and other global approximators are flexible machine learning-based tools capable of incorporating a broad range of information sources into control policies (e.g., reservoir storages, sensing-based observations, forecasts, etc.), but this flexibility may come at the cost of over-fitting. The implications of the tradeoff between policy complexity and generalizability in the context of emerging forecast-informed reservoir operations have been underexplored. This study investigates how parameterizations of the RBFs in EMODPS interact with features of forecast information to determine the robustness of forecast-informed policies. We test out-of-sample policy performance with different numbers of RBFs and forecast lead times using synthetically generated forecasts of varying skill, and measure performance degradation from the training period. These tests are conducted in a case study of the Conowingo Dam in the Lower Susquehanna River Basin. As the largest privately owned reservoir in the US, operating policies at the Conowingo Dam must balance trade-offs between multiple objectives, including urban water allocation, nuclear power plant cooling, hydropower production, federally regulated environmental flow requirements, flood protection, and recreation. These challenges will increase with the demand for ancillary hydropower services and shifting energy price dynamics as non-dispatchable renewables continue to penetrate into the PJM Interconnection. Results from this analysis will be used to determine how forecasts can help the Conowingo hydropower system better integrate with a rapidly shifting energy market. This study contributes to the growing suite of methodological advances available to link hydrologic forecasts with water resources management decisions to improve multi-objective tradeoffs in complex systems.