Physics informed machine learning based predictive control for intelligent operation of edge datacenters

Edge data centers (EDCs) are critical for supporting latency-sensitive applications; however, their high energy consumption, especially in cooling systems, presents significant sustainability challenges that are further compounded by recent advancements in Artificial Intelligence (AI). In recent years, integrating AI with Model Predictive Control (MPC) has shown promise in reducing cooling energy consumption while maintaining operational safety, but most AI models remain impractical for real-world deployment. This limitation arises from their inherent black-box nature and heavy dependence on the quantity and quality of training data, which hampers their ability to handle unseen conditions effectively. To address these challenges, we propose a physics-informed machine learning-based MPC (PIML-MPC) framework that integrates physical principles into neural networks to improve prediction accuracy. By embedding established physical knowledge, our approach enhances model generalization and reduces its dependence on extensive training data. We validated the framework using a simulation model derived from real data center measurements. Experimental results show that the PIML-MPC strategy reduces energy consumption by 12.6 % compared to a baseline, without compromising thermal regulation. Under unseen high heat load conditions, the framework further reduces energy consumption by 9.1 % while maintaining thermal performance, demonstrating robust generalization. Moreover, across ten randomized trials, PIML-MPC consistently outperforms three advanced time-series and machine-learning-based MPC schemes—achieving superior energy efficiency, tighter thermal control, and reduced sensitivity to stochastic variations. This work paves the way for the practical adoption of AI-driven MPC in data center cooling by substantially improving the generalization performance of AI controllers.