A Data-Driven Network Model for Traffic Volume Prediction at Signalized Intersections

Network-wide traffic prediction at the level of an intersection can benefit transportation systems management and operations. However, traditional traffic modeling approaches relying on mathematical or simulation-based models are either less useful or require higher computational time in predicting high fidelity traffic volumes. In addition, these frameworks need to be modified to ingest large-scale data (such as automated traffic signal performance measures) available from intersections. To overcome these challenges, in this study, a data-driven method based on a deep learning architecture has been developed for network-wide intersection-level traffic prediction. The study has tested two deep learning architectures: Graph Convolutional LSTM (GCN-LSTM) and Graph Convolutional Encoder–Decoder LSTM (GCN-Encoder–Decoder) model to predict intersection-level hourly traffic movement volumes over multiple time steps (e.g., 4-h sequence). Such deep learning architectures capture the spatiotemporal cross correlation among network-wide traffic features while learning the patterns in traffic movement volumes. To test the model performances, we have fused data from multiple sources such as travel demand data, built environment characteristics, etc. We have extracted 1 year (2016) of traffic movement volume data from Seminole County’s automated traffic signal performance measure (ATSPM) database. Experiment results show that the developed GCN-LSTM model outperforms all the other baseline models. The absolute difference between actual and predicted volumes are quite low (GEH < 5); for right turn, through and left turn movement RMSE values are 4.02, 59.37, and 2.47, respectively. The R2 score of the model is 0.98, which indicates that the model can capture the spatiotemporal variations of traffic movement volumes very well.