Machine Learning and Motion Coordination for Battery Electric Vehicles : Safe, Efficient and Smart Framework for Online Learning

Sammanfattning

The integration of electric trailers with battery electric trucks offers a promising solution for extending driving range and improving energy efficiency in long-haul transport. A central challenge is optimizing the torque split between truck and trailer when the trailer’s internal power-loss characteristics are unknown and proprietary. This thesis addresses this challenge by developing a safe online learning and control framework that enables the truck to identify the trailer’s loss behavior during operation and leverage this knowledge for energy-efficient coordination. The framework is modular, consisting of four interconnected blocks: learning, exploration, safety simulation, and control allocation. Learning is performed using Recursive Least Squares (RLS) as a lightweight parametric baseline and Gaussian Process (GP) regression as a non-parametric, uncertainty-aware alternative. Exploration is guided by heuristic coverage and Bayesian Optimization- inspired strategies, while safety is enforced via short-horizon dynamic simulations and Control Barrier Functions (CBFs). The learned trailer power-loss model is then embedded in optimization-based control allocation, solved via Quadratic Programming for RLS and Sequential Quadratic Programming for GP-based models. Simulation studies on drive cycles show that GP models achieve up to 97% learning accuracy, reduce energy consumption by 2-4% compared to baseline strategies, and operate within 1% of an oracle with full trailer knowledge, all while preserving safety against jackknifing and trailer swing. These findings highlight the feasibility and benefits of safe online learning for articulated electric vehicles and provide insight into future deployment of data-driven control in heavy-duty transport.