TACOS: A Topology-Aware Cooperative Multiagent Framework for Power-Computing Coscheduling in Data Center Microgrids

The rapid expansion of data center (DC) clusters has led to a significant increase in electricity consumption and carbon emissions, making power-computing coscheduling a critical enabler for sustainable DC operation. Large-scale AI services introduce massive and bursty workloads with complex task dependencies, which pose substantial challenges to minute-level scheduling in DC microgrids integrated with renewable energy and energy storage. These challenges are further aggravated by multidimensional uncertainties, including stochastic job arrivals, volatile electricity prices, and photovoltaic generation, as well as by stringent quality-of-service requirements and time-coupled computation-energy constraints. Existing analytical optimization and single-agent learning methods are often limited by poor scalability, weak topology awareness of job structures, and insufficient safety guarantees under temporally coupled constraints. To address these issues, this article proposes a topology-aware heterogeneous multiagent deep reinforcement learning framework for dynamic power-computing co-optimization in DC microgrids. Specifically, the problem is formulated as a constrained Markov Game to characterize interactions among scheduling and energy management entities. A Graphormer-based representation module is developed to encode directed acyclic graph job topologies and capture task dependencies. Moreover, constrained heterogeneous-agent proximal policy optimization with hybrid advantage decomposition and sequential updates is designed to handle temporal coupling constraints and improve the stability and safety of policy learning. Experimental results under realistic AI workload and energy scenarios demonstrate that the proposed method can effectively reduce energy costs while satisfying task deadline constraints, and exhibits strong scalability and robustness in large-scale DC microgrids.