Nonlinear Model Predictive Control based on Local Linearized Gaussian Process and its Application to a Quadruple-Tank System

This paper introduces a data-driven model predictive control(MPC) method based on Gaussian process(GP) and local linearized model. Firstly, the first principle model of a quadruple tank system is established derived by physical law and taken as a benchmark. Then the brief control block diagram is presented. Secondly, the Gaussian process model is introduced. As a non-parametric probabilistic model, it has a solid theoretic and inference basis to be combined with the model predictive control. One-step prediction and multi-step prediction are inferred as a basis of control. Thirdly, GP-MPC combines the offline identification of GP and the online optimization of MPC. Furthermore, the local linearization based on GP is proposed to promote the algorithm by converting the nonlinear optimization to a quadratic programming problem with less time and better performance. Finally, simulations are executed on the quadruple tank system to compare the GP-MPC algorithm and the one with local linearization.