Graph-based spatial-temporal networks for traffic speed prediction in intelligent transport systems

Traffic speed forecasting is among the key challenges in improving traffic management and urban mobility in the framework of the Intelligent Transport Systems. The complexity and nonlinear nature of traffic data coupled with learning dynamic spatial-temporal dependencies present a significant challenge that the traditional forecasting methods are unable to address, due to their inability to model dynamic changes in the traffic network. Our paper presents two new deep learning-based models, ST-GIncep (Spatial-Temporal Gated Inception Network) as well as ST-GTransLSTM (Spatial-Temporal Graph Transformer LSTM). The ST-Incep uses a multi-scale convolution architecture along with gating mechanisms that are able to capture spatial features at various scales along with Matrix LSTM to model temporal features. The ST-GTransLSTM combines the ability to the graph transformer to capture spatial information efficiently coupled with the capacity of the LSTM to process sequences data to create a hybrid Encoder-Decoder architecture allowing effective modeling long-term temporal dependencies. The frameworks presented in this paper are tested on the real-traffic dataset PemSD7. The experimental results depict both proposed models outperform existing baseline frameworks with respect to predictive accuracy and robustness. The performance of the approaches presented on noisy data and efficiency on low-traffic networks demonstrate their utility in real-world ITS applications.