Traffic Prediction with Peak-Aware Temporal Graph Convolutional Networks

In this study, traffic speed prediction on a large-scale traffic network in Ankara City is performed using deep neural networks. For this purpose, a spatiotemporal deep learning model consisting of Graphical Convolutional Networks and Gated Recurrent Units used as the baseline, and (i) the input space is expanded by temporal embedding to better take into account temporal information, and (ii) to increase the performance for the peak hours of traffic, the loss function is extended with a novel weighting mechanism. Our comprehensive experiments have shown that the proposed method is significantly more successful in peak hours than ARIMA (Autoregressive Integrated Moving Average) and deep learning-based methods.