Multivariate event hypergraph diffusion model for train delay prediction

Train delay prediction is a key technology for train scheduling and timetable optimization, and constitutes a critical component of intelligent transportation systems. We present the first study on regional-level multi-train delay prediction problem, and focus on modeling the regional-level delay propagation and evolution process, and capturing coordinated operation status among multiple train clusters in the complex operation network. First, we propose a brand-new Multivariate Event Hypergraph Diffusion (MEHD) model, and introduce a novel data structure, the mixed hypergraph, which accurately models the spatio-temporal high-order correlations between the regional-level multi-train arrival events. Then, we propose a mixed hypergraph convolution method to characterize complex train operation network, which improves the ability to capture the spatio-temporal high-order correlations and non-Euclidean characteristics between events. Finally, we propose an event hypergraph diffusion process, and design a prior operational schedule-conditioned attention denoising module to enhance the ability to learn all train arrival event generation mechanisms. Extensive experiments demonstrate that our MEHD achieves superior performance compared to current state-of-the-art models on actual high-speed rail performance datasets, with an average improvement of 20%-30% on multiple metrics, and performs good robustness and efficiency. Subsequent experiments and analyses demonstrate the unique advantages of MEHD over single-train prediction methods. To the best of our knowledge, this is the first end-to-end model for regional-level multi-train delay prediction. The dataset and source code are available online: https://github.com/bjtuxuyi/MEHD.