Uncertainty in modelling and optimising operations of reservoir systems

Abstract

Reservoir operation optimisation aims to determine release and transfer decisions that maximise water management objectives such as the reliability of water supply, the hydropower production or the mitigation of downstream floods. This thesis studies two key issues in reservoir operation optimisation. Firstly, despite being an active field of research, the state of uptake of optimisation techniques by practitioners is largely unknown. Secondly, there are sources of uncertainty in the simulation models that underpin optimisation and the impact of these uncertainties on operation optimisation results has not yet been considered. We present a literature review that classifies different optimisation techniques based on what types of problem they are applicable to rather than the mathematical workings behind them, as previous reviews have done. This review is contrasted with a practitioner survey that reaches water managers and consultants around the world. We find that practitioners do not typically use operation optimisation tools, instead following decision-making procedures that are more informal than the formulaic operating policies presented in research. The survey suggests that a key reason for hesitation in the uptake of optimisation techniques is the limited fidelity of simulation models that underpin optimisation results. We discuss sources of uncertainty in these models and find that no work has yet considered the impact of structural uncertainty (i.e. arising from how interrelationships within the system model are defined) or contextual uncertainty (i.e. around definition of the model boundaries) on reservoir operation optimisation. Consequently, we formulate ‘rival framings’ of a real-world reservoir operation problem, each making different assumptions about structural/contextual uncertainties affecting the model of the system. We then test how the estimated performance of optimised decisions changes when evaluated under different framings. We find that contextual uncertainty in particular has a significant impact on estimates of performance. Finally, we investigate the applicability of ‘robust optimisation’, i.e. an approach where operations are directly optimised under multiple model formulations at once. In our case study, robust optimisation is effective because it produces a set of solutions that have greater robustness than would be achievable using conventional optimisation.

Date of Award	19 Mar 2019
Original language	English
Awarding Institution	University of Bristol
Supervisor	Francesca Pianosi (Supervisor) & Thorsten Wagener (Supervisor)