Large scale real-time traffic flow prediction using SCATS volume data

© 2017 Rabindra PandaRoad traffic congestion is a global issue that results in significant wastage of time and resources. Rising population, urbanisation, growing economies and affordable personal vehicles aggravate the issue. Many urban cities have been trying to mitigate this by expanding and modernising the transportation infrastructure. Even though increasing the road capacity accommodates the travel demands, studies have shown this does not eliminate the congestion problem. Hence, since 1970’s advanced traffic management systems have been used to address the issue of congestion. But for these systems to increase their operational efficiencies and fully realise their effectiveness, they need to have the predictive capabilities in the short term, usually ranging between few seconds to few hours. The research in short term traffic prediction has been active since the 1970’s. Numerous models have been proposed to use the traffic data collected by in- ductive loop detectors for short term traffic prediction. Most of the works have shown promising results through experiments at particular locations, however we are still to find a robust and globally adaptable solution. In last decade the attention have shifted from theoretically well established parametric methods to non parametric data driven algorithms. This work is an extension to that. Neural networks have always been one of the most capable mathematical models that can model complex non-linear relations. Up to 2006, their use have been hindered by practical issues related to the training. But recent breakthroughs in new ways of training deep neural architectures have made them reemerged as victors by realising the capabilities they had promised. In this thesis we study and extend their applications to short term traffic predictions. We applied three deep recurrent neural networks (Simple RNN, LSTM and GRU) in predicting the short term traffic volumes. The goal was to use both the temporal and spatial relationships that are present in the traffic flow data. We used these networks at univariate and multivariate settings to make predictions at single location and multiple locations respectively. For this work we used the volume data collected by VicRoads in Melbourne. We compared the results of our work with several existing methods and found promising results