Accurate and Decentralized Approach for Traffic Flow Prediction Using Federated Learning

Traffic prediction involves forecasting the density and flow of traffic, primarily aimed at efficiently managing vehicular movements in a highly congested urban environment. Its purpose is to offer valuable insights for reducing traffic congestion and devising optimal routes that minimize energy consumption and travel time. Nevertheless, conventional traffic prediction faces a significant challenge: the need to store data when making predictions, resulting in increased latency, pro-longed training time, and the consumption of expensive storage resources. Moreover, sharing such data for training may raise privacy concerns due to the potential inclusion of sensitive information. To address these issues, this paper adopts Federated Learning (FL) to train the prediction model closer to the data sources. To assess the effectiveness of the proposed approach, real-world traffic data from Arhus City has been employed. The proposed method's prediction accuracy has been evaluated compared to a centralized approach.