A shock wave diagram based deep learning model for early alerting an upcoming public event

Abstract

Even though extensive efforts worked on traffic anomaly detection, this study noticed that little attention is given to predict public events before they start, like celebrations, sports games, and so on. Given a public event produces complicated spatiotemporal traffic impacts, existing data-driven approaches without involving traffic flow analysis show limited effectiveness to predict a public event. Motivated by this view, this study seeks to develop a domain-knowledge-based learning approach, which integrates shock wave analysis into deep learning models to predict the occurring of an upcoming public event (i.e., the SW-DLM approach). This integration raises new research challenging and calls for new approaches. Specifically, we develop an efficient algorithm to generate the shock wave diagrams to present the evolvement of the traffic anomaly, which expands whenever new traffic data is collected. Next, the shock wave diagram itself is not a well-coded input for feature extraction and learning. This study thus developed an innovative encoding approach, which transforms a shock wave diagram into an optimal pixel grid. Considering the features extracted from the encoded shock wave diagrams are fed into the long-short term memory (LSTM) model for the event prediction. The numerical experiments based on the field data indicate that the SW-DLM is able to predict a public event with 87% accuracy around 84 min before it starts in a day. It outperforms all data-driven machine learning models using point traffic data as inputs. Thus, we claim that using the shock wave diagrams can significantly improve the accuracy and efficiency of the learning approaches for predicting a public event. The SW-DLM will help develop preventive traffic control or route plans to avoid traffic congestion induced by public events.