Breaking the Curse of Stochastic Dynamic Programming: Advancements in Optimal Reservoir Control

Abstract

Reservoirs reduce flood risks, support regional development by increasing water availability for various economic sectors, contribute renewable electricity, and are increasingly important in the water cycle. Advancing their operations is only becoming more relevant given renewed interest in dam construction and the need to reoperate existing infrastructure. This thesis contributes a critical analysis of 103 recent studies attempting to overcome the limitations, or ‘curses’, of Stochastic Dynamic Programming, the current gold standard in optimal reservoir operations. The findings from this analysis suggest a possible road map of open research directions. First, the programming of a multi-scale cascading data dependency structure to facilitate the development of a composite hydroclimatic reservoir model that includes both upstream and downstream states in the problem formulation. Second, acknowledging then subsequently tackling the ’black box perception’ of approximation networks by reservoir operators, specifically looking at the interface between optimization algorithms and human supervisors.