Short-term traffic prediction under disruptions using deep learning

In this chapter, we have proposed a novel graph -based model with TS-TGAT to predict short-term traffic speed under both normal and abnormal traffic fl ow conditions. The novelty of the proposed prediction model is that it can learn both spatial and temporal propagation rules for traffic on a network. Important concepts and improvements are introduced to the model, for example node -level attention weights, multi -head attention and depth -wise separable CNN module to take account of the unique and complex interactions between traffic fl ows and traffic network characteristics. The proposed prediction model was trained and tested using ILDs on a section of the M25 motorway network just before the Dartford Crossing (between Dartford Tunnel and M25 J2 with all slip roads). In order to make the model generic and reusable, the model was trained using generic data (including both normal and abnormal traffic fl ow data) and was tested under mixed conditions and disrupted conditions. A selection of baseline methods was used to benchmark the proposed model performance, including HA, kNN, GBDTs and LSTM, some of which are state-of-the-art methods in the problem of short-term traffic prediction. The results have shown that the proposed TS-TGAT method outperforms other benchmarking methods under both normal and abnormal traffic conditions.