Coal bulk loading and unloading operations constitute a critical component in the coal logistics supply chain, involving complex coordination among train scheduling, coal type matching, stacker-reclaimer collaboration, and yard management. This paper proposes a machine learning-based optimization method that integrates Large Language Models (LLMs) for identifying key efficiency factors and Deep Double Q-Networks (DDQN) for real-time decision-making. The system generates constrained yard decision topology graphs from operational data, employs spatiotemporal graph convolutional networks for state encoding, and implements multi-objective reward functions encompassing efficiency, compliance, energy consumption, and operational costs. Experimental results demonstrate that the proposed method achieves significant improvements in total operation time compared to manual scheduling, while ensuring strict adherence to operational constraints including coal type matching, equipment conflict avoidance, and temperature-based safety limits. The dynamic optimization mechanism enables continuous improvement through feedback-based adjustment of reward weights and model parameters.