Vulnerability-aware reinforcement learning scheduling of AI data centers under typhoon-induced thermal coupling

Abstract

The rapid growth of AI data-center loads, amplified by typhoon-induced thermal coupling between workloads and cooling systems, poses new challenges to transmission grid vulnerability assessment and scheduling. This paper develops a vulnerability-aware reinforcement-learning (RL) scheduling framework that integrates a measured-data-driven AI load model, a typhoon-day temperature–PUE–cooling model, a seven-indicator vulnerability assessment covering physical, structural, and AI-specific dimensions, and RL scheduling via PPO, SAC, and TD3 with a PTDF-based top- K branch-selection mechanism that confines control to AI-responsive branches. Validation on the IEEE 39-bus and 118-bus systems shows that all three RL algorithms achieve statistically significant V DC reductions of 6.9–9.3% ( p < .001), with PPO outperforming an LP comparator by 3.4–8.4% on the 118- and 1354-bus systems. The reduction magnitude correlates with the vulnerability-gradient signal—the product of mean PTDF sensitivity and baseline vulnerability—rather than with system size per se: on the PEGASE 1354-bus system, where the AI load represents only 0.016% of total system demand and the transmission network is highly robust, no scheduling strategy produces a meaningful V DC response, confirming a physical optimization ceiling governed by AI load penetration rather than network size. Under extreme stress at 3.0 × AI load, TD3 reduces EENS by 44%, confirming V DC as a proactive leading indicator of curtailment risk. An ablation study identifies topological sensitivity as the most influential indicator. The proposed framework provides grid operators with ex-ante guidance on when vulnerability-aware scheduling becomes effective, based on the relationship between AI load penetration, vulnerability gradients, and system controllability.