Multi-train trajectory planning

Although different parts of the rail industry may have different primary concerns, all are under increasing pressure to minimise their operational energy consumption. Advances in single-train trajectory optimisation have allowed punctuality and traction energy efficiency to be maximised for isolated trains. However, on a railway network safe separation of trains is ensured by signalling and interlocking systems, so the movement of one train will impact the movement of others. This thesis considers methodologies for multi-train trajectory planning. First, a genetic algorithm is implemented and two bespoke genetic operators proposed to improve specific aspects of the optimisation. Compared with published results, the new optimisation is shown to increase the quality of solutions found by an average of 27.6% and increase consistency by a factor of 28. This allows detailed investigation into the effect of the relative priority given to achieving time targets or increasing energy efficiency. Secondly, the performance of optimised control strategies is investigated in a system containing uncertainty. Solutions optimised for a system without uncertainty perform well in those conditions but their performance quickly degrades as the level of uncertainty increases. To address this, a new genetic algorithm-based optimisation procedure is introduced and shown to find robust solutions in a system with multiple different types of uncertainty. Trade-offs are explored between highly optimised trajectories that are unlikely to be achieved, and slightly less optimal trajectories that are robust to real world disturbances. Finally, a massively parallel multi-train simulator is developed to accelerate population-based heuristic optimisations using a graphical processing unit (GPU). Execution time is minimised by implementing all parts of the simulation and optimisation on the GPU, and by designing data structure and algorithms to work efficiently together. This yields a three orders of magnitude increase in rate at which candidate control strategies can be evaluated