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

Researchers have developed PI-DLinear, a physics-informed machine learning model that forecasts GPU power consumption in AI data centers 5-80 minutes ahead with significantly higher accuracy than existing methods. The model integrates thermal physics principles with deep learning to predict power fluctuations caused by different AI workloads, addressing grid stability challenges from volatile LLM inference and training operations.

AI data centers face unprecedented operational challenges as large language models and other compute-intensive workloads create unpredictable power demand spikes. Traditional forecasting models struggle because they lack understanding of the underlying physical relationships between GPU utilization, heat generation, and power throttling. This research bridges that gap by embedding Newton's law of cooling and thermal resistance-capacitance networks directly into a machine learning framework, enabling the model to capture non-linear dynamics that purely data-driven approaches miss.

The physics-informed approach demonstrates substantial practical advantages. Across multiple accuracy metrics, PI-DLinear outperforms state-of-the-art transformer and non-transformer baselines by 0.8% to 51.8%, depending on the evaluation window. Critically, the model respects physical constraints during power throttling and load transient events, meaning forecasts remain physically plausible rather than producing unrealistic predictions that violate thermal principles.

For infrastructure operators and grid managers, this advancement addresses a growing pain point. AI data centers now consume comparable power to entire cities, and their volatile demand patterns threaten grid stability. Accurate short-term forecasting enables better load scheduling, demand response coordination, and prevents cascading failures. Energy providers can optimize reserve capacity, while data center operators can improve cooling efficiency and reduce costs through predictive thermal management.

The framework's success suggests physics-informed machine learning represents a viable path for solving real-world infrastructure problems where domain knowledge matters. Future work likely extends this approach to larger prediction windows, multi-facility networks, and renewable energy integration scenarios.

• →PI-DLinear forecasts GPU power consumption 5-80 minutes ahead with 0.8-51.8% greater accuracy than existing models
• →Physics-based constraints prevent unrealistic predictions and improve performance during power throttling events
• →Accurate power forecasting enables better grid management and reduces energy costs for data centers
• →The model bridges pure machine learning with thermal physics to capture GPU utilization-temperature-power relationships
• →This advancement addresses critical infrastructure challenges as AI workloads create volatile and unpredictable power demands