A Physics-Aware Framework for Short-Term GPU Power Forecasting of AI Data Centers

Researchers at arXiv (2605.04074) introduced PI-DLinear, the first physics-informed DLinear time-series model for GPU power forecasting in AI data centers, targeting 5-80 minute horizons. The architecture embeds a multi-node lumped thermal RC network governed by time-dependent ODEs coupling GPU compute, memory utilization, and temperature via Newton's law of cooling. Grid operators and data center infrastructure teams should prioritize this, as LLM training versus inference power divergence directly threatens grid stability. PI-DLinear outperforms standard DLinear by grounding predictions in thermodynamic physics rather than purely statistical patterns.