Dynamic reconfigurable battery systems via graph-based deep reinforcement learning

Existing large-scale batteries, such as those used in electric vehicles, electric planes, and electric boats/ferries, are built from hundreds or thousands of battery cells connected in fixed series-parallel configurations. These rigid topologies limit efficient power management, making it difficult to respond to dynamic cell imbalance and thereby reducing overall battery operating time. To address these limitations, reconfigurable battery designs have been proposed where the interconnection topology among cells can be adjusted in real time to reflect evolving cell dynamics and uncertainties in the operating environment. In this thesis, we model reconfigurable batteries as dynamic graphs and investigate graph-based deep reinforcement learning approaches for adaptively optimizing cell-to-cell topology under practical operational constraints. We evaluate the proposed methods by using the open-source battery simulation platform PyBaMM, measuring performance in terms of voltage balancing and State of Charge (SOC) uniformity. Overall, this work establishes a foundation and offers insights for the development of future intelligent reconfigurable battery systems.