Bridging the grid thermal computational nexus via physics inform 2027 Applie

Driven by the rapid growth of cloud computing and artificial intelligence, data centers (DCs) are becoming large, flexible loads that increasingly affect active distribution networks (ADNs). Coordinating these loads motivates joint assessment of thermal-model accuracy, operating performance, and online computation. This paper proposes a physics-informed hierarchical multi-agent reinforcement learning (H-MARL) framework for grid–thermal–computational scheduling of a single-operator DC park within an ADN. A physics-informed neural network (PINN) is first developed as a dynamic cooling surrogate that captures zone heat balance and envelope heat gains while promoting thermodynamic consistency through residual losses and a soft coefficient-of-performance (COP) plausibility penalty. The PINN achieves a test mean absolute percentage error of 1.40% in cooling-power prediction, averaged across four DCs and five training seeds, with an approximately 5000-fold speed-up over EnergyPlus based on mean per-DC evaluation times. A H-MARL scheduler coordinates spatiotemporal workload migration, dynamic voltage and frequency scaling (DVFS), and cooling, while alternating-current optimal power flow (AC-OPF) dispatches electrical assets under network constraints. Agents observe carbon-inclusive nodal prices and share a rew.