Federated deep learning for enhanced prediction of electric vehicle charging station availability

Douaidi L., Senouci S., El Korbi I., Harrou F., YAZICI A.

Cluster Computing, cilt.28, sa.6, 2025 (SCI-Expanded) identifier identifier

  • Yayın Türü: Makale / Tam Makale
  • Cilt numarası: 28 Sayı: 6
  • Basım Tarihi: 2025
  • Doi Numarası: 10.1007/s10586-025-05152-z
  • Dergi Adı: Cluster Computing
  • Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Academic Search Premier, PASCAL, Applied Science & Technology Source, Compendex, Computer & Applied Sciences, INSPEC
  • Anahtar Kelimeler: Charging station, Data privacy, Electrical vehicle, Federated learning, Non-IID data, Occupancy prediction
  • Eskişehir Osmangazi Üniversitesi Adresli: Evet

Özet

The growing demand for public Electric Vehicle (EV) Charging Stations (CSs) is vital for promoting wider EV adoption but frequently results in congestion due to limited availability and high usage, leading to increased wait times for drivers. In response, this study proposes a predictive occupancy model utilizing a privacy-preserving Federated Learning (FL) approach, which enables multiple Charging Station Operators (CSOs) to collaborate without sharing sensitive user data. Unlike traditional centralized models, our FL framework allows CSOs to contribute to a central server that aggregates their local models, maintaining privacy and ensuring data security. A key challenge in this setup is managing non-Independently and non-Identically Distributed (non-IID) and heterogeneous data, common in real-world scenarios with diverse user behaviors and charging patterns. To address this, we evaluate various aggregation algorithms, including FedAvg, FedProx, SCAFFOLD, and FedPer, to determine their effectiveness under different conditions. Using 10-min interval data from the Dundee City CS dataset, we predict station occupancy 1 h ahead. The results show that FedPer and SCAFFOLD perform exceptionally well when handling unbalanced data, while FedAvg proves to be more effective in situations with skewed feature distributions. This FL approach not only improves the accuracy of EV charging station occupancy predictions but also lays the groundwork for securely scaling EV infrastructure, ensuring that privacy concerns do not hinder the development of intelligent, data-driven services.