Adaptive and safe fast charging of lithium-ion batteries via hybrid model learning and control barrier functions

Prolonged charging times remain a critical challenge for Lithium-ion batteries (LiBs), which are the leading energy storage solution for many applications. This work proposes a framework for generating a safe and optimal charging policy by leveraging in-situ operational data streams from LiBs. The operational data is used to continuously learn a virtual hybrid model decomposed into physics-based and neural residual components in a manner that maintains maximum explainability. Then the fast-charging policy is learned via deep reinforcement learning (DRL) with latent imagination using the hybrid model. Since the DRL policy is based on a continually adapted hybrid model, the resulting charging profile is expected to be optimal for the prevailing LiB age and operational conditions. By conducting DRL exploration on the virtual hybrid model, the framework mitigates safety risks and operational inconveniences associated with on-battery DRL training. Additionally, the framework incorporates formal safety guarantees using a control barrier function (CBF) filter that includes state of charge (SOC), voltage and temperature safety and operational constraints. The DRL policy is trained to promote rapid SOC increases while penalizing CBF constraint violations. After training the DRL+CBF policy on the hybrid model, the fast charging policy is implemented on the physical LiB. Both experimental and simulation results demonstrate that the framework delivers adaptive charging policies tailored to various operational conditions, such as ambient and battery temperature, while adhering to safety constraints. In experimental tests, the proposed framework delivers up to 8X faster charging compared to the manufacturer-recommended constant current constant voltage (CCCV) profile. • Adaptive Charging policy for lithium-ion batteries with deep reinforcement learning. • Deep reinforcement learning trained with latent imagination of virtual hybrid model. • Control barrier function filters guarantee formal safety constraints.