Traffic flow prediction with graph convolutional networks

Abstract

Traffic flow prediction is a complex task utilising both spatial and temporal information in order to predict the expected speed of traffic at different points in a transportation network. We study the use of graph convolutional networks for traffic speed prediction. Different models are investigated utilising internationally benchmarked data before the most promising approach (Graph WaveNet) is applied to a new database of South African road network data. This task is relevant, since traffic forecasting enables better road network management, reducing congestion. This in turn reduces travel time and pollution. Different traffic forecasting models were considered before selecting Graph WaveNet as the best available traffic forecasting tool in terms of performance, availability of code, and training time. Graph WaveNet was first analysed using well-known datasets containing fixed sensor readings, while the South African datasets were being constructed. It was found that the models produced perform well when it comes to handling missing data and utilising the historic trend of data when making predictions. While Graph WaveNet performs well overall, we found that there is still room for improvement during times of congestion in road networks. Graph WaveNet was successfully implemented on new Floating Car Data (FCD) (not fixed sensor data) recorded on South African road networks utilising only self-learned adjacency matrices. The South African models produced shared many similarities with the earlier models: the models successfully utilised the historic trend of the data, and performance degraded during periods of congestion in the road network. While the final error is not directly comparable, the new models achieved an MAE of 4.84, 4.99 and 5.10 over a 15-minute, 30-minute and 60-minute prediction horizon on the Johannesburg dataset, compared to an MAE of 2.69, 3.05 and 3.49 over the same prediction horizons on the METR-LA dataset.