Federated Computing in Electric Vehicles to Predict Coolant Temperature

Abstract

Reducing greenhouse gas emissions in mobility is paramount to achieving a carbon-neutral society. However, battery-electrical vehicles (BEV) introduce unique engineering challenges to protect expensive electrical components from overheating. A centralized architecture for model-driven predictions of coolant temperatures poses privacy and legal issues. Additionally, the applications in a vehicle compete for the available resources and must use them as sparingly as possible. Therefore, we introduce a new federated computing (FC) use case to help transform the mobility sector. We evaluate the performance of two FC approaches (linear regression and machine learning) on hardware and privacy metrics by leveraging a real-world dataset from BEVs. Our findings show trade-offs between hardware utilization and model accuracy. The linear regression model yields the best performance and prediction metrics. FC with ML shows up to 761 % variances when comparing vehicle-specific models with models trained with the entire fleet and clustering the data into velocity profiles partly improves prediction performance.