Network-wide Lane-level Traffic Flow Prediction via Clustering and Deep Learning with Limited Data

This paper proposes a method based on the hierarchical clustering algorithm and deep learning for network-wide short-term traffic flow prediction with limited data. The proposed method treats each lane detector as an observation point. To address the challenges posed by the heterogeneity and complexity of various traffic patterns observed in the traffic network, a profile model and clustering method are proposed to divide all detectors into clusters. After clustering, the detectors within each cluster are homogeneous with similar traffic trends, making them easy to learn, even with limited data. Each cluster corresponds to learning a certain trend, equipped with a separate predictive model, the proposed CNN-based deep model. This paper uses MAE, RMSE, and MAPE to test the prediction accuracy. The proposed method is validated on the PeMS dataset. It has been proven to provide accurate lane-level traffic flow predictions considering detector reliability.