TrafficBERT: Pre-trained model with large-scale data for long-range traffic flow forecasting

Abstract

Traffic flow prediction has various applications such as in traffic systems and autonomous driving. Road conditions have become increasingly complex, and this, in turn, has increased the demand for effective traffic volume predictions. Statistical models and conventional machine-learning models have been employed for this purpose more recently, deep learning has been widely used. However, most deep learning-based models require data additional to traffic information, such as information on adjacent roads or road weather conditions. Therefore, the effectiveness of these models is typically restricted to certain roads. Even if such information were available, there is a possibility of bias toward a specific road. To overcome this limitation, based on the bidirectional encoder representations from transformers (BERT), we propose trafficBERT, a model that is suitable for use on various roads because it is pre-trained with large-scale traffic data. Our model captures time-series information by employing multi-head self-attention in place of the commonly used recurrent neural network. In addition, the autocorrelation between the states before and after each time step is determined more efficiently via factorized embedding parameterization. Our results indicate that trafficBERT outperforms models trained using data for specific roads, as well as commonly used statistical and deep learning models, such as Stacked Autoencoder, and models based on long short-term memory, in terms of accuracy.