On-Device Adaptive Battery Power Prediction for Electric Vehicles

This paper presents an on-device adaptation method for battery power forecasting in resource-constrained electric vehicle systems. It transforms pretrained forecasting models into adaptable versions while retaining critical hyperparameter knowledge from initial training. The study evaluates online and offline adaptation strategies across multiple models and forecasting horizons. According to the abstract, online adaptation reduces mean absolute error by up to 7.49%, while offline adaptation achieves reductions of up to 14.88% compared with unadapted deployments. The work targets distribution shifts that commonly degrade battery prediction performance in real-world EV operation.