Under high load conditions, unreasonable scheduling strategies in distributed data centers will increase energy consumption and carbon emissions, which goes against the original intention of green data centers. Traditional static policies are difficult to adapt to dynamically changing environments. Even reinforcement learning methods struggle to infer the policies of other centers with limited observability, thus impeding cooperation. To address these challenges, we propose Belief and Attention-based Agent Modeling (BAAM), a novel multi-agent reinforcement learning framework for energy-optimized job scheduling in distributed data centers. BAAM integrates a fact-based belief reasoning module and self-attention mechanism, allowing each agent to infer the hidden strategy of its peers and predict its own future state changes using only local observations, actions, and rewards. We conducted extensive experiments using Alibaba’s real dataset cluster-trace-v2018 to validate BAAM in a distributed data center scheduling scenario that simulates job dynamic arrival and virtual machine execution. Compared with traditional scheduling methods and MARL baselines, BAAM provides better coordinated scheduling policies among centers and substantially reduces energy consumption, ​achieving about 20% lower energy usage than commonly used advanced MARL approaches. Moreover, BAAM exhibits strong robustness and adaptability in different environments, demonstrating its practical potential for achieving green data center operations.