Uncertainty Representation in Stochastic Reservoir Optimization

Abstract:

Water resources managers attempt to operate reservoir and hydropower systems to maximize system objectives, subject to a host of physical and policy constraints, and in light of uncertainty about future conditions. Optimization models are widely used to advise the decision making process. An important aspect of such models is how uncertainties related to future hydrologic and economic conditions are represented, and the extent to which different uncertainty representations affect the quality of recommended decisions. This study explores the consequences of different uncertainty representations in stochastic optimization models of hydropower systems by comparing simulated system performance using different stochastic optimization models. An important question is whether the added computational burden from greater uncertainty resolution (which can be prohibitive for operational models in many cases) actually improves model recommendations. This is particularly relevant as more complex, ensemble forecasts are incorporated into short- and mid-term planning models. Another important consideration is how watershed hydrology (both seasonal and episodic characteristics), system size, economic context, and the temporal resolution of the model influence how uncertainty should be represented. These topics are explored through several US examples including a sampling stochastic dynamic programming (SSDP) model of a small single-reservoir system on the Kennebec River in Maine, and a stochastic programming model of the large multi-reservoir Federal Columbia River system in the Pacific Northwest. These studies highlight the importance of flexible model frameworks which allow exploration of different representations of a system and of uncertainties before locking operational decision support system development into a specific representation.