DATA-DRIVEN ADAPTIVE MODEL PREDICTIVE CONTROL FOR CONSTRAINED NON LINEAR SYSTEMS

Abstract

Electronic throttle systems (ETS) combine a DC motor, a gearbox, a limp-home return spring, Coulomb friction and airflow torque, and must track the commanded plate angle within a 0–24 V supply. The base paper of this thesis (Youssef, Wafa and Shalaby, 2025) addresses this with an Adaptive Relaxed Sliding Mode Controller (ARSMC) in which a Deep Deterministic Policy Gradient (DDPG) actor–critic agent adapts the lumped-uncertainty bound, shrinking the boundary layer that removes chattering. This thesis reproduces the ETS model and the SMC, RSMC and ARSMC controllers of the base paper and extends the work with an offset-free Model Predictive Controller (MPC). The MPC feedbacklinearises the throttle, estimates angle, speed and the lumped disturbance with a third-order extended state observer (ESO), and solves a 40-variable box-constrained quadratic program every 5 ms, so the 0– 24 V limit is enforced inside the optimisation rather than by clipping. All controllers are compared in MATLAB on the same staircase reference (0.4, 0.8 and 1.5 rad), parametric uncertainty and airflow disturbance, in three scenarios: uncertainty inside the design bound, uncertainty 50% above it, and 2 mrad sensor noise. Compared with ARSMC, the proposed MPC reduces the integral absolute error by 39%, 36% and 78%, the integral squared error by 37%, 34% and 73%, and the rise time by 41%, 38% and 78% in the three scenarios, with zero steady-state error and a chatter-free voltage. Its cost is a 4–6 mrad overshoot, which ARSMC avoids. The MPC needs no training and solves its QP in about 0.6 ms.