A Multi-Agent Reinforcement Learning Approach for Optimizing Shared Spare Parts Procurement Management

Power plant operations critically depend on effective spare parts management, where traditional shared procurement strategies often employ simplistic single-factor prioritization methods that fail to consider comprehensive allocation factors. These approaches lead to suboptimal resource utilization, increased costs, and potential safety risks. We address these limitations by proposing a novel Multi-Agent Deep Deterministic Policy Gradient framework integrated with Transformer Encoder architecture (TE-MADDPG) for multi-plant spare parts management. Our innovation lies in modeling each power plant as an independent decision-making agent within a Markov Decision Process that dynamically incorporates inter-plant distances, spare part values, criticality indexes, and real-time demand patterns. The Transformer Encoder component effectively solves the challenge of handling variable-length observation spaces across different plants—a significant limitation in previous approaches. Experimental validation with operational data from 8 power plants demonstrates that TE-MADDPG consistently outperforms traditional methods, achieving an average 2.3% reduction in procurement costs and substantial improvements in transfer efficiency, particularly during high-demand periods. Our framework provides a scalable, adaptable solution that enhances both operational safety and economic performance in power generation facilities.