Enhancing Intelligent Battery Management on Automotive Microcontrollers Using LightGBM

State-of-Health (SoH) is a core metric of any Battery Management System (BMS), impacting operational safety and performance optimization. However, its on-board estimation remains challenging due to the dynamic nature of battery data and the tight memory and energy constraints of embedded systems, requiring frequent yet impractical model updates. We present a framework for training and deploying ensembles of decision trees on BMS hardware, delivering accurate SoH estimates across the entire battery lifespan without model updates, while also being resource-inexpensive on on-board Microcontrollers (MCUs). On two benchmark datasets, our method achieves up to 39% lower MAE than state-of-the-art deep-learning baselines. Moreover, when deployed on an industry-grade automotive MCU, it delivers inference speedups of 11–30x while occupying under 3% of the available memory.