Predict-then-optimise interchange control for rail logistics with combined passenger-freight chains

Abstract

Interchange decisions at rail transfer stations directly affect end-to-end service reliability and cost-to-serve; yet, existing delay-management approaches either forecast delays without prescribing actions or optimise dispatch with coarse delay representations. We propose a predict-then-optimise framework that integrates multi-task LSTM-based delay prediction with a feasibility-constrained Markov Decision Process (MDP) to determine whether to proceed, hold, reschedule, or reroute services at transfer stations. The LSTM models short-horizon delay propagation using only information available at the prediction epoch, while the MDP converts the predicted delay class, capacity regime, infrastructure condition, and bounded dwell feasibility into auditable dispatch recommendations. We evaluate the framework on a three-leg regional rail case study in central Germany comprising lines RB5 (Fulda–Bebra), RB87 (Bebra–Eichenberg), and RB83 (Eichenberg–Kassel), with two physical interchange stations (Bebra and Eichenberg). The results show that the sequence models provide sub-minute ETA accuracy across all three legs and that the Bebra interchange policy reduces the cost-to-serve index by 7.0% relative to the earliest-feasible heuristic while maintaining the same structural miss rate under a 60-minute admissible dwell window. At Eichenberg, the MDP and earliest-feasible policies perform almost identically among evaluable journeys, while the low matched-departure coverage highlights a data and timetable-coordination limitation for full multi-leg evaluation. The framework contributes a deployment-oriented closed-loop architecture in which policy-induced connection gaps are propagated into downstream ETA estimation, supported by per-timestep station embeddings, two-stage cross-leg fine-tuning, and replay-buffer stabilisation. The paper demonstrates how established prediction and control components can be operationally coupled to provide interpretable, manager-facing rail logistics decision support with journey-level KPIs and cost-to-serve reporting.