Optimal Control and Reinforcement-Learning Strategies for Advanced Management of Lithium-ion Battery Packs

The use of mathematical models and predictive control in advanced battery management is fundamental in order to achieve high performance, especially in the case of large battery packs in which several cells are arranged both in series and parallel connections. A basic requirement for the development of predictive control is the exploitation of an accurate model. Within this context, as a first contribution, this Thesis provides a thorough investigation of optimal design of experiments methodologies which are able to enhance the identifiability of electrochemical battery models, relying both on the concept of Fisher information, which exhibits a local nature, and global sensitivity analysis. A further contribution is provided in the battery control field, where dissipation and ageing-aware fast-charging strategies are developed for a single lithium-ion cell. Subsequently, model-predictive control is exploited for both the state-of-charge balancing of series-connected cells and the optimal charge of a whole battery pack. In the latter case cells connected in series and parallel arrangements are considered and a sensitivity-based linearization of the overall model is proposed in order to maintain the computational burden at a reasonable level. Finally, the use of model-free reinforcement learning is considered with application to battery fast charging and optimal velocity planning for autonomous hybrid electric vehicles in an urban context.