Abstract
Battery Management Systems (BMS) play a crucial role in electric vehicle operation, with State of Charge (SOC) estimation being one of their most critical functions. This paper presents a novel approach to SOC estimation inspired by the Mixture of Experts (MoE) architectures. We adapted the MoE paradigm to the BMS domain, demonstrating that an architecture utilizing two specialized simple feedforward neural networks can significantly improve the accuracy of SOC estimation compared to single-specialized models. Our experimental results show that when trained and tested on driving cycle data, the MoE architecture achieves a Mean Absolute Error of 0.0259 and an R 2 value of 0.9711, substantially outperforming the single expert, trained on the Coulomb counting method. The improved accuracy and robustness of the MoE approach have the potential to enable better battery utilization, extended range, and enhanced safety in electric vehicles.