Comparing graph neural network-based traffic speed prediction for static sensor and floating vehicle data sets

Abstract

Traffic speed prediction using deep learning neural networks has been the topic of many studies, mostly using data sets that were collected from static sensors. Floating vehicle data offers a more flexible alternative, as it can be obtained for any roads travelled by GPS tracked vehicles. In this paper we compare the performance of leading traffic speed prediction techniques when applied to both static sensor and floating vehicle data sets. Data sets were collected for the road networks serving Johannesburg, representing South Africa’s most congested roads. Based on prediction accuracy, training time and robustness the Graph WaveNet method produced the best results. We found that the static sensor and floating vehicle data sets, representing traffic movements on the same sets of roads, produced comparable results, providing evidence that static sensor data can be complemented and, in some cases, replaced by floating vehicle data. This will enable traffic speed prediction for roads where no static sensors are installed, resulting in significant cost savings. Extending traffic speed prediction to all major roads will result in improved traffic management strategies for the overall road network, leading to less congestion and an improved road user experience.