TWAFR-GRU: An Integrated Model for Real-time Charging Station Occupancy Prediction

The fast-growing number of electric vehicles has resulted in a variety of charging-related issues, such as the shortage of public charging piles, the inconvenience of finding available charging piles, the additional traffic congestion caused by unnecessary cruising to search for available charging piles, etc. Since only constructing new charging piles may not significantly improve the charging efficiency, methods of predicting charging station occupancy are widely discussed as an effective means to address these issues. However, several challenges are encountered in training an efficient and effective model by utilizing distributed data, improving convergence speed, and enhancing model generalization ability. To address these issues, this paper proposes a novel mechanism, named TWAFR-GRU, which integrates Temporally Weighted Asynchronous Federated Learning (TWAFL) with Reptile and Gated Recurrent Unit (GRU). As shown by the holistic evaluation based on the charging station occupancy dataset, compared with other state-of-the-art baselines, TWAFR-GRU can 1) decrease MAE, RMSE and RAE by 19%, 15% and 17% separately and improve R
²
, by 67%; 2) cut rounds for the model to converge by 75%; 3) save training time by 44%; and 4) reduce 17% in forecasting error after the personalization of the initial model to serve a specific charging station.