Federated Learning Predicts EV Charging Demand Early

A new study from arXiv (2605.04993) explores federated learning for early prediction of electric vehicle charging demand. Using the Adaptive Charging Network (ACN) dataset, researchers at Caltech developed a session-level model that estimates total energy consumption from plug-in time and initial charging minutes. The approach enables real-time decisions for EV network operators by capturing user intent, temporal patterns, and initial behavior. Focusing on a single depot at Caltech, the study models intra-depot heterogeneity via station-level client partitions and evaluates multiple model families in a federated learning setting. This work aims to improve grid stability and infrastructure planning.