Gantry2Vec: Embedding Heterogeneous Gantry Information into Vectors for Enhanced Traffic Flow Prediction

Abstract

Highway gantry sensors record real-time traffic data and uniquely demonstrate spatiotemporal and behavioral heterogeneity for traffic prediction. This study combines the modeling capabilities of representation learning and traffic domain knowledge to embed gantry information to vectors, termed Gantry2Vec. Three representations are defined: temporal embedding and spatial embedding at the macro level and behavioral embedding at the micro level. As the representation of behavior information for highway traffic flow prediction is underutilized in the current literature, we incorporated these macro–micro representations and subsequently proposed a gantry-aware framework for downstream traffic flow prediction. Herein, customized feature ablation experiments using our proposed Gantry2Vec have been conducted for validating the performance of model. We also provide insights into the embedding performance on distinguishing traffic flow patterns of heterogeneous gantries on a real-world highway data set. Experimental results demonstrated that our proposed model outperforms other baselines.