Adaptive federated learning framework for predicting EV charging stations occupancy

Forecasting the occupancy of electric vehicle (EV) charging stations is crucial for addressing key challenges in e-mobility, including charging inefficiency, traffic congestion, and drivers’ difficulty in locating available stations. Recent studies on predicting charging station occupancy have predominantly employed conventional centralized machine learning models. However, these models struggle to adapt to evolving data streams, limiting their accuracy and practical applicability. Furthermore, centralized approaches exacerbate these limitations by conflicting with privacy requirements, as they necessitate sharing sensitive data from charging stations and users. To address these limitations, this study introduces a novel federated learning framework designed to collaboratively predict EV charging station occupancy by leveraging data from all participating stations without sharing any raw data. The framework addresses critical challenges, including concept drift, data heterogeneity among clients, and model generalization. It clusters stations with similar drift patterns and leverages personalized models, updated using incremental learning, to ensure adaptive and precise forecasting. This adaptability allows the system to respond to dynamic environments characterized by changing traffic patterns, user behaviors, and station utilization trends. Three datasets with diverse characteristics were used to validate the proposed framework, demonstrating its robustness, adaptability, and scalability in improving forecasting accuracy. This work provides a robust, privacy-preserving, and scalable framework to enhance the reliability and efficiency of EV infrastructures, promoting broader adoption of sustainable transportation.