A Deep Learning Model for Predicting Work Zones Delay Costs

Highway infrastructures require frequent maintenance that involves in most cases lane closures. With the increase of vehicles-mile traveled, work zones became the second contributor in non-recurrent congestion. work Zone congestion accounts for 24 percent of non-recurrent congestion and 10 percent of the overall congestion. Work Zone congestion occurs on the upstream segments and depending on the characteristic of the work zone, traffic volume, and geographic conditions, work zone congestion can spillback to connected freeways. Transportation systems provide means for passengers and goods movement. However, work zones with lane closures are accounted for congestion.

Transportation system aims on forecasting the work zone congestion impact on transportation networks. Work zone capacity is the maximum number of vehicles entering work zone, which contributes to the reduction in speed in the upstream segments. Other factors include vertical gradient of work zone segments, speed during normal conditions, traffic volume approaching work zone, and distance of segments from work zone. While the effect of each of these factors over estimating upstream speed is not explicit, a non-parametric approach is more efficient for predicting speed with work zone conditions.

Deep artificial neural networks are a machine learning technique that are used to identify traffic patterns. Previous studies used artificial neural networks to predict speed upstream work zone on the mainline only. While previous models are not able to capture the complexity of larger scale networks, with a sufficient big data available, a new deep neural network is able to predict the speed on upstream mainline, interchanges, and connected freeways. Furthermore, previous studies suffered from overfitting problems in the artificial neural networks. Therefore, speed prediction has accuracy limitations when predicting work zone conditions that are not included in the training dataset.

This project aims to predict speed on mainline segments upstream work zone in addition to interchange and connected freeway segments. The prediction would capture the spatial-temporal impact of work zone on New Jersey freeways. Once a congestion approaches interchange segments, a spillback of congestion usually occurs to the interchange and connected freeways. The research focuses on the factors affecting the prediction of congestion on upstream interchanges and connectors in addition to upstream mainline freeway segments. Moreover, the study recognizes the problem of deep artificial neural networks and suggests a dropout technique to prevent the co-adaptation on training data. A numerical evaluation is conducted on Interstate-295 to compare the predicted speed results of the model with previous prediction models in terms of their consistency with actual work zone speed data. .

Skills: SQL, Python, Google API