Delay prediction using machine learning algorithms for connected autonomous traffic flow in uninterrupted facilities

Abstract

Road incidents and breakdowns in freeways create excess delays in road sections for a duration of time. Amount of this delay plays an important role in the route planning of the people on the road. In this thesis, we propose a delay prediction methodology using machine learning techniques on feature engineered detector data. Our method uses incident characteristics, lagged detector data, adjacent detector data & lagged adjacent detector data as features. Created features are selected using mutual information criteria, correlation analysis and regularized & standard random forest feature importance values. The final model successfully predicts the delay for next timestep with mean squared error of 224.89 for training and 247.77 for testing data sets. Model performance further improves for the simulation conditions with less uncertainty such as incidents with short duration, accidents on right or left lane and detectors further away from the incident location. .