Keywords: railway departure delay prediction, retrieval-augmented time series diffusion model, probabilistic forecasting, generative models, train operation control, uncertainty quantification.

In high-frequency urban railway operations, departure delays at stations may propagate to subsequent trains and develop into chronic delays. Conventional delay prediction methods generally output a single predicted value, although actual delays are probabilistic phenomena affected by passenger flow, boarding and alighting time, and interactions among trains. This paper proposes a railway departure delay prediction method using a Retrieval-Augmented Time Series Diffusion Model (RATD). The proposed method retrieves past delay sequences similar to the observed historical sequence and uses them as reference information to guide the diffusion-based generation process. By generating multiple prediction sequences under the same input condition, the method represents departure delay prediction as a probability distribution. Using actual train operation records from a commuter railway line in the Tokyo metropolitan area, the generated distributions were compared with actual departure delays. The results indicate that the proposed method can qualitatively capture major delay levels and increasing, stable, and decreasing delay trends, suggesting its potential for uncertainty-aware train operation control support.

download the full-text of this paper (PDF, 9 pages, 888 Kb)

go to the previous paper
go to the next paper
return to the table of contents
return to the volume description