Short-term traffic state estimation and prediction based on spatiotemporal neural networks

Intelligent transportation systems use forecasted traffic states as basic information. This study aims to improve the performance of short-term traffic prediction using neural networks (NN). As a data-driven approach, NN models are heavily impacted by data samples. Under general data availability settings, we develop enhanced approaches in input data, model structure, and forecasting output. For data shortage and data overstock settings, several novel approaches regarding transfer learning, feature visualization, domain shifting are proposed and discussed. Overall, we present a more systematic and general methodology framework in short-term traffic prediction that positively impacts transportation research and industry.