Real Mine Data-Driven Multi-Agent Reinforcement Learning for Coordinated Ventilation Fan Operational Scheduling

Ventilation fan operation scheduling is a core problem of mine-site operational management. This paper proposes a multi-agent reinforcement learning (MARL)-based scheduling framework that leverages historical operational data to improve decision-making and address key shortcomings of conventional evolutionary methods, including low efficiency, limited generalizability, and high operational complexity. Using a graph neural network architecture, a performance calculation model for combined multi-fan operating configurations is constructed by integrating historical observations with prior knowledge of fan characteristic curves. We formulate the scheduling of multiple fan ventilation units as a set of parallel decisions, constrained by airflow demand and operating rules. The objective is to reduce costs and increase efficiency while ensuring that the approach remains practical and simple to implement. To improve generalization, a MARL framework is developed that explicitly models variability in the operating conditions. A comparative evaluation against evolutionary algorithms shows that, among the reinforcement learning methods examined, the trained MARL approach adjusts reliably to fluctuating airflow delivery conditions and demonstrates strong generalization across scenarios. Relative to the observed operational schedule, the proposed approach yields substantial savings, reducing operating costs by more than CNY 43,900 and lowering energy consumption by over 47,000 kWh. Compared with evolutionary algorithms, the reinforcement learning-based approach produces decisions that better satisfy operational constraints and operational logic while substantially improving computational efficiency, achieving a planning speedup of more than 50 times.