Decision intelligence for electric bus charging scheduling through LLM-driven reward shaping and reinforcement learning

The rapid expansion of electric bus fleets has made charging scheduling a critical decision-intelligence problem in urban fleet-energy logistics. Reinforcement learning (RL) is promising for this problem, but reward function design still relies on manual trial-and-error and transfers poorly across scenarios. This paper develops an artificial intelligence and operations research (AI-OR) decision intelligence framework that integrates large language model (LLM)-driven reward shaping with RL-based sequential decision-making. The framework represents reward functions as executable code and iteratively optimizes them through four modules: reward initializer, performance analysis, reward exploration, and rollback. An event-triggered two-stage charging reallocation mechanism enforces the action-level operational constraints through action masking, while the battery safety bound is controlled through readiness conditions, reward shaping, and the acceptance criterion. Experiments on a synthetic scenario and two real-world bus stations in Guangzhou compare three mainstream LLMs under different initialization conditions. Results show that LLM-driven reward iteration outperforms traditional methods across all scenarios, reducing costs by up to 32.6 % in nominal within-day charging cost and by up to 22.5 % after an approximate adjustment for the energy left on board. Ablation and sensitivity analyses validate the contribution of each module and the framework’s adaptability to multiple RL algorithms. Reward-evolution analysis further identifies design principles for improving charging efficiency, supporting valley-period energy shifting, and enhancing fleet-energy logistics resilience.