Search

The use of various advanced traffic data collection systems on one hand, and the development of Big Data technologies for the storage and processing of large amounts of data on the other hand, have enabled the application of various non-parametric methods for traffic volume prediction. In this research, the possibilities of application of supervised machine learning, as a method of Big Data analytics, with the aim to predict various indicators of the traffic volume were investigated. The research was conducted through two case studies. In both studies, for training and testing predictive models, traffic data generated by selected automatic traffic counters on the roads in the Republic of Serbia, in the period from 2011 to 2018, were used. Prediction models were trained, tested and applied using Weka software tool. The most basic data preparation was performed using macros for MS Excel written in VBA (Visual Basic for Applications). In the first case study, the goal was to predict the total volume of traffic by days, on selected sections of state roads in the Republic of Serbia. The datasets used for training and testing of machine learning models in the first case study were prepared using MS Access database, and the prediction results were presented using Excel Pivot Charts. In the second case study, we selected one counting point and performed prediction of the hourly vehicle flow, by directions and in total for both directions. The preparation of data sets, as well as the visualization of the results of the Big Data analysis in the second case study, was performed using programs written in the Python programming language. On the prepared data sets, using Weka software tool, different regression prediction models were trained and tested in both case studies. In the first case study, the best results were received by models based on regression decision trees, while in the second study, models based on Lazy IBk, Random Forest, Random Committee and Random Tree algorithms were among best. In each of the case studies, the best prediction model was selected by comparing model performance measures, such as: correlation coefficient, mean absolute error, and square root of mean square error. The model based on the M5P algorithm has shown the best performance in the first study, while the Lazy IBk algorithm gave the best results in the second study. Using the best predictive models, the prediction of daily or hourly traffic for 2020 was made at selected traffic counting points. Supervised machine learning has proven to be an effective method in predicting the volume of traffic flow. [ABSTRACT FROM AUTHOR]