Modelling and Model Predictive Control of Power-Split Hybrid Powertrains for Self-Driving Vehicles

Abstract

Designing an autonomous vehicle system architecture requires extensive vehicle simulation prior to its implementation on a vehicle. Simulation provides a controlled environment to test the robustness of an autonomous architecture in a variety of driving scenarios. In any autonomous vehicle project, high-fidelity modelling of the vehicle platform is important for accurate simulations. For power-split hybrid electric vehicles, modelling the powertrain for autonomous applications is particularly difficult. The mapping from accelerator and brake pedal positions to torque at the wheels can be a function of many states. Due to this complex powertrain behavior, it is challenging to develop vehicle dynamics control algorithms for autonomous power-split hybrid vehicles. The 2015 Lincoln MKZ Hybrid is the selected vehicle platform of Autonomoose, the University of Waterloo’s autonomous vehicle project. Autonomoose required high-fidelity models of the vehicle’s power-split powertrain and braking systems, and a new longitudinal dynamics vehicle controller. In this thesis, a grey-box approach to modelling the Lincoln MKZ’s powertrain and braking systems is proposed. The modelling approach utilizes a combination of shallow neural networks and analytical methods to generate a mapping from accelerator and brake pedal positions to the torque at each wheel. Extensive road testing of the vehicle was performed to identify parameters of the powertrain and braking models. Experimental data was measured using a vehicle measurement system and CAN bus diagnostic signals. Model parameters were identified using optimization algorithms. The powertrain and braking models were combined with a vehicle dynamics model to form a complete high-fidelity model of the vehicle that was validated by open-loop simulation. The high-fidelity models of the powertrain and braking were simplified and combined with a longitudinal vehicle dynamics model to create a control-oriented model of the vehicle. The control-oriented model was used to design an instantaneously linearizing model predictive controller (MPC). The advantages of the MPC over a classical proportional-integral (PI) controller were proven in simulation, and a framework for implementing the MPC on the vehicle was developed. The MPC was implemented on the vehicle for track testing. Early track testing results of the MPC show superior performance to the existing PI that could improve with additional controller parameter tuning.