A Bi-Layered Machine Learning Model for Travel-Time Prediction Along a Congested Section of I-495, USA

Many Machine Learning (ML)-based approaches have been developed for travel-time prediction, including ensemble-based, neural network, and others. However, these approaches often rely solely on historical data and lack the capability to adapt to real-time traffic conditions. They are unable to extract live traffic stream and develop a more realistic travel-time prediction between desired Origin–Destination (O-D) pairs in a dynamic manner. In this paper, we develop a bi-layered ML model to predict travel time along congested traffic networks using real-time traffic information from Google Maps. The first ML layer uses typical traffic conditions from Google Maps using historical data. The second ML layer uses a Random Forest Classifier to predict travel time at various peak times of the weekday at 5-min intervals. An example using a 7.7-mile congested section of I-495 outer loop in the USA is presented. We first use simulated data to gain insight into the effect of different variables on prediction. We then use the actual congestion level from the first layer using Google Maps to perform the prediction. The results show about 89% match with travel-time congestion under typical conditions. A plot of the resulting travel time shows up to a twofold increase in the travel time during some time periods. Additional test scenarios, including a further decomposition of the congestion segment, can be studied in future works. In addition, the first ML layer can be directly integrated into the second ML layer via a computer vision algorithm in future works.