Long-term traffic flow predictions in a transformer-based framework

Abstract

Traffic flow predictions are an important component in the rising demand for solutions to cope with the increasing pressure on transportation networks. Especially on a long prediction horizon, traffic flow predictions remain challenging due to the complex, nonlinear nature of traffic flow and the influence of both temporal and external features. To incorporate sequential behavior of time series, without being subject to limitations inherent in recurrence-based models, the transformer is increasingly used. However, whether this model is advantageous on long horizons is still unknown. Therefore, in this paper, first, multiple correlation analyses are applied to identify important features in traffic flow. Next, these are incorporated into a generic transformer-based framework, and the adequacy of the transformer on long prediction horizons is investigated based on real data. To test the genericity of the proposed prediction model, all analyses are conducted for two locations, which are subject to different traffic behavior. The first is located on the ring road of Haarlem and is mainly affected by commuter traffic, whereas the second is located on the road to the coast and has more irregular behavior. Results show that the transformer outperforms baseline prediction models on both short and long horizons, especially when the location is subject to irregular behavior. In addition, the inclusion of external features, such as the day of the week, holidays, and temperature, improves the model performance. Moreover, the genericity of the model is highlighted by its applicability to multiple locations.