Impact of a stochastic and dynamic setting on the stability of railway dispatching solutions

Automatic decision support systems for the real-time dispatching of railway traffic have not reached yet real operations. The main concern with these systems is that they must guarantee a stable control strategy that prevents nervous behaviors of continuously changing solutions which might confuse human dispatchers. In this paper we investigate the stability of optimal dispatching plans against the dynamic evolution of randomly disturbed traffic conditions. A framework is developed that couples the state-of-the-art dispatching system ROMA, with the microscopic simulation environment EGTRAIN. Optimal plans are computed in a rolling horizon setup on the basis of updated traffic information and evaluated over multiple disturbed scenarios for different prediction horizons. Experiments on the Dutch railway corridor Utrecht-Den Bosch show that the instability increases as random perturbations propagate. Shorter prediction horizons give a more stable but less effective control strategy since mostly retiming is suggested by optimal plans. Larger horizons consent to manage traffic more effectively by reordering trains but lead to more unstable solutions. Enlarging the prediction horizon over a given threshold does not alter the structure of the control strategy also in terms of stability.