Enhancing Electric Vehicle Range Prediction: A Federated Learning Framework Addressing Non-IID Data Challenges

Range anxiety hinders EV adoption, emphasizing the need for precise range prediction to support sustainable transportation. Developing effective models is challenging due to the non-IID nature of data among diverse EV users. To address this, we propose UniFed, a federated learning (FL) framework integrating Deep Reinforcement Learning (DRL) for adaptive client selection and Differentially Private GANs (DP-GANs) for data augmentation, overcoming limitations of traditional methods like FedAvg. UniFed addresses data heterogeneity while introducing uniformity as an optimization objective, ensuring even accuracy distribution across clients via the Gini coefficient. This promotes uniform participation, mitigates bias, and strengthens model robustness. Extensive experiments on EV trip datasets and benchmarks show UniFed achieves faster convergence, higher accuracy, and fairer participation in FL settings.