A study of model predictive control for spark ignition engine management and testing

Pressure to improve spark-ignition (SI) engine fuel economy has driven the development and integration of many control actuators, creating complex control systems. Integration of a high number of control actuators into traditional map based controllers creates tremendous challenges since each actuator exponentially increases calibration time and investment. Model Predictive Control (MPC) strategies have the potential to better manage this high complexity since they provide near-optimal control actions based on system models. This research work focuses on investigating some practical issues of applying MPC with SI engine control and testing.;Starting from one dimensional combustion phasing control using spark timing (SPKT), this dissertation discusses challenges of computing the optimal control actions with complex engine models. A nonlinear optimization is formulated to compute the desired spark timing in real time, while considering knock and combustion variation constraints. Three numerical approaches are proposed to directly utilize complex high-fidelity combustion models to find the optimal SPKT. A model based combustion phasing estimator that considers the influence of cycle-by-cycle combustion variations is also integrated into the control system, making feedback and adaption functions possible.;An MPC based engine management system with a higher number of control dimensions is also investigated. The control objective is manipulating throttle, external EGR valve and SPKT to provide demanded torque (IMEP) output with minimum fuel consumption. A cascaded control structure is introduced to simplify the formulation and solution of the MPC problem that solves for desired control actions. Sequential quadratic programming (SQP) MPC is applied to solve the nonlinear optimization problem in real time. A real-time linearization technique is used to formulate the sub-QP problems with the complex high dimensional engine system. Techniques to simplify the formulation of SQP and improve its convergence performance are also discussed in the context of tracking MPC.;Strategies to accelerate online quadratic programming (QP) are explored. It is proposed to use pattern recognition techniques to "warm-start" active set QP algorithms for general linear MPC applications. The proposed linear time varying (LTV) MPC is used in Engine-in-Loop (EIL) testing to mimic the pedal actuations of human drivers who foresee the incoming traffic conditions. For SQP applications, the MPC is initialized with optimal control actions predicted by an ANN. Both proposed MPC methods significantly reduce execution time with minimal additional memory requirement.