Traffic speed prediction techniques in urban environments

The present study developed Multiple Linear Regression (MLR) and machine learning (ML) models, including Artificial Neural Network (ANN), Support Vector Machine (SVM), and Random Forest (RF), to predict the mean free-flow speed (FFS) using several geometric, traffic, and pavement condition variables. The traffic features group includes spot speed, speed limit, average speed, 85th percentile speed, traffic and crossing pedestrian volumes, volume of exiting vehicles, percentage of elderly crossing pedestrians (Elderly%), percentage of heavy vehicles (HV%), and traffic calming measures (TCMs). The geometric characteristics include lateral clearance, number of effective lanes, number of access points (including median openings), road grade, effective lane width, and median width. The pavement condition category includes pavement roughness in the International Roughness Index (IRI). A total of 11 urban arterials were used to develop the MLR model and train the ML models. Test data were collected from two randomly selected roads to evaluate the performance of each model, investigate the differences between conventional linear regression and ML approaches, and determine the best prediction models based on the results of the two techniques. Results showed that the proposed ML algorithms outperformed linear regression models. They are believed to be valuable and strong tools to predict the mean FFS that adapts to sudden changes in traffic flow caused by exogenous conditions on urban arterials and can be employed in determining the most influential factors and building reliable prediction models where spot study is not feasible due to time and resource limitations.