부산대학교 도서관

The use of artificial intelligence (AI) enables the optimization of driving profiles so that energy efficiency and punctuality in the rail system can be increased. We show that with our comprehensive modelling approach of Deep Reinforcement Learning (DRL) agents, we can determine driving profiles that reduce energy demand by an average of 11.1 % and schedule deviation by 65 seconds compared to average drivers from regular passenger operations. For the first time, the trained agents are then examined for their transfer learning capabilities. In scenarios with highly divergent driving time reserves, the DRL agents reduce energy demand by an average of 12.5 % and schedule deviation by 61 seconds. We also show that the trained DRL agents are able to increase energy efficiency and punctuality on unknown route sections. Thus, we demonstrate that the system is suitable for use as a driver assistance system and is able to adapt to the variable operational conditions and possible detours in light rail operations.