Pattern recognition technique based active set QP strategy applied to MPC for a driving cycle test

Application of constrained Model Predictive Control (MPC) to systems with fast dynamics is limited by the time consuming iterative optimization solvers. This paper proposes a fast and reliable Quadratic Programming (QP) strategy to solve MPC problems. While the optimal control action is calculated with a fast online dual QP algorithm, a “warm start” technique is adopted to reduce iterations of the online search process. The warm start solution is calculated from a predicted active constraint set generated by a pattern recognition function (Artificial Neural Network, ANN, is discussed). This function is calibrated with data from Monte Carlo simulation of the MPC controller over finite sampling points of the state-space. The proposed MPC strategy can adapt to applications with long prediction/control horizons, Linear Parameter Varying (LPV) dynamics and time varying constraints with balance between computation time, memory requirement and calibration effort. This MPC approach is applied to control vehicle speed for a HIL driving cycle test on an engine dynamometer. Simulation results demonstrate the speed profile tracking error of the MPC “driver” can be 67% less than a PID “driver”. Furthermore, smooth throttle/brake actuations, similar to human drivers are achieved with the MPC controller.