A generative model for vehicular travel time distribution prediction considering spatial and temporal correlations

Vehicular travel time distributions (TTDs) are of great importance for traffic management and control, and various probability distributions have been used for TTD prediction in previous studies. However, it is difficult to determine a generalized probability distribution of vehicular travel times on urban roads that is applicable to all traffic conditions in real situations. To solve this problem, this paper develops a machine learning-based generative model, named the travel time distribution prediction-generative adversarial network (TTDP-GAN) model, that uses license plate recognition data for TTD prediction. The TTDP-GAN model generates samples of predicted travel time to account for its probability distribution, and these samples are not based on any assumed distribution. In addition, the TTDP-GAN model considers the spatial and temporal correlations of the TTD predictions by applying the multi-head spatial and temporal self-attentions, structural similarity index measure (SSIM), and long short-term memory (LSTM) neural networks. The performance of the TTDP-GAN model is demonstrated in a case study of an urban road network in a medium-sized city in China. The results show that the TTDP-GAN model outperforms several state-of-art machine learning models (e.g., an LSTM neural network model, a GAN model, a Wasserstein GAN model, and an LSTM-GAN model) in the measurement of Jensen–Shannon (JS) divergence and in terms of mean, standard deviation, skewness, and kurtosis. In addition, the TTDP-GAN model with the SSIM has 21.43% better predictive accuracy for JS divergence than the TTDP-GAN model without SSIM. These results demonstrate that the adoption of SSIM is efficient in capturing the probability distribution for TTD prediction. A sensitivity analysis is also carried out to showcase the performance of the TTDP-GAN model in applications.