Stochastic MPC for a medium-term optimization Semester Thesis

ii This thesis focuses on stochastic model predictive control applied on a hy-dro power plant. The goal was to provide a tool which robustly analyzes the current and assumed future resources and then calculates the optimal secondary control power to advertise for one week. This allows an opera-tor of a plant to maximize his profits and being robust against the acting stochastic variables at hand like missing resources in the form of empty water reservoirs. To solve this problem first a discrete mixed-logic-dynamic model of the plant is generated. Then the three acting stochastic values which are the secondary control power signal and two stochastic water-inflows are modeled with quantile regression. Quantile regression is used because measured data indicate that varying distributions affect the stochastic variables. To achieve robustness against these stochastic variables a certain number of realizations of the random variables are drawn and the most restrictive of these is applied to the system. This practice called scenario approach would yield hard bounds on the percentage of the possible realizations of the stochastic variable but only if the stated problem would be convex. The formed model however is not convex. Therefore these theoretical bounds provided by the scenario approach are no longer valid. The simu-lations on the model done with numerous scenarios based on the quantile regression model showed that reliable bounds apply. Because of this a robust algorithm can be provided which reaches the set goals. The findings also indicate that the model of the secondary control power signal is most sensitive to the outcome of the developed algorithm and there-fore crucial. Key word