Learning-Based Predictive Controller for Connected and Automated Hybrid Electric Vehicles

This article aims at presenting a learning-based predictive control strategy for hybrid electric vehicles (HEVs) in the presence of uncertainty, where the controller structure and energy efficiency of the HEV is simultaneously optimized. The proposed approach includes development of a Bayesian optimization (BO)–based control structure optimization method, followed by an eco-driving–based hierarchical robust energy management strategy (EMS) development for connected and automated HEVs. To apply the learning-based strategy online, we also introduce an approach with approximate cost function for the BO to reduce training and computation time and improve energy in a given trip. The control structure is described by a parameter vector, which is updated, using BO, in an episodic fashion with the performance of the EMS and the computation time. With the current control structure, the hierarchical EMS includes a high-level powertrain energy manager that takes long-term decisions, and a low-level velocity optimized and a robust powertrain controller that takes decisions over a short time horizon. We develop a pseudo-spectral optimal control method to solve the long-term energy management problem, and robust tube–based model predictive controller (MPC) strategy to solve the short-term problem. Jointly, the control structure design and EMS development focuses on improving energy efficiency of the vehicle and reduce computation time. Simulation results show our proposed method with exact and approximate BO cost outperform existing methods by ≈ 12.6% and ≈ 9.71%, respectively.