The Sequence-to-Sequence Architecture with An Embedded Module for Long-Term Traffic Speed Forecasting with Missing Data

Traffic forecasting plays a crucial role in Intelligent Transportation Systems (ITSs), which is proposed to provide traffic status in advance for road users to avoid traffic congestion or other traffic incidents and for authorities to optimise the strategies of traffic management. In this paper, we develop a novel deep learning framework, based on the Sequence-to-Sequence architecture with an embedded module, for long-term traffic speed forecasting with missing data and providing high forecasting accuracy. The embedded module uses Graph Convolution Neural Network for the local spatial dependency analysis by conducting convolutional operation on the k−hop neighbourhood matrix, while utilises Transformer for the global spatial dependency analysis by implementing the attention mechanism that assigns individual weights to neighbour detectors for contributing to the targeted detector. The sequence-to-sequence architecture is built to analyse temporal dependencies of the spatially-fused time series from the embedded module. To evaluate the proposed model against existing well-known ones, the real traffic speed dataset with missing data and frequent traffic incidents is used to train and test the models. The experimental results indicate that our proposed framework achieves the most accuracy forecasting, even obtaining more than 80% accuracy for forecasting two hours in advance.