UBC Theses and Dissertations
A novel algorithm for model plant mismatch detection for model predictive controllers Tsai, Yiting
For Model Predictive Controlled (MPC) plants, the quality of the plant model determines the quality of performance of the controller. Model Plant Mismatch (MPM), the discrepancies between the plant model and actual plant transfer matrix, can both improve or degrade controller “performance”, depending on the context in which “performance” is defined. Instead of using simple, “yes-no” type of performance metrics to diagnose whether MPM is present, this thesis achieves the further goal of pinpointing the exact plant inputs causing MPM. The detection method consists of a two-step identification procedure. The first experiment identifies the MPM-affected rows in the plant matrix. The second experiment pinpoints the exact inputs causing said MPM, and these MPM-affected elements are further ascertained using a hypothesis test. Finally, using input design, an estimated optimal excitation sequence is generated based on the available closed-loop data, which helps the engineer determine whether the original excitation was sufficient for the aforementioned sys-ID experiments. The most important underlying assumptions are the linear time-invariance of the plant and noise dynamics. The proposed algorithm is exercised on artificial 3x3 and 5x5 plants, then on real data from an industrial lime slaker suffering from sparse MPM, to demonstrate its ability of accurately pinpointing MPM-affected elements.
Item Citations and Data
Attribution-NonCommercial-NoDerivatives 4.0 International