Transformer-based Generative AI for HVAC Control with Fragmented Data Center Histories of IT Workload and Cooling Infrastructure

Data centers often face fragmented operational histories due to sensor upgrades or logging failures, hindering the training of robust HVAC control policies. This paper introduces the Generative Thermodynamic Graph Transformer (GTGT), a novel framework designed to reconstruct missing temporal data in Cyber-Physical Systems. By integrating a thermodynamic graph with a physics-informed Transformer, GTGT synthesizes high-fidelity operational profiles that bridge multi-year data gaps. We utilize real-world datasets from a tropical data center testbed (2018-2023) to validate our approach. Experimental results demonstrate that GTGT achieves a 42\% reduction in reconstruction error compared to standard Transformers and maintains 99.8\% thermodynamic consistency. Furthermore, Deep Reinforcement Learning agents trained on this augmented history achieve a Power Usage Effectiveness (PUE) of 1.08 and 15.6% energy savings in volatile transition scenarios, significantly outperforming agents trained on fragmented data.