An FMI-Based Fast Thermal Prediction FMU for Real-Time Digital Twins of Data Center Server Rooms

Abstract

Data center server rooms require thermal prediction components that can be called by digital twins and system simulation platforms under changing IT loads and cooling conditions. CFD provides detailed spatial information but is too expensive for step-wise online simulation. Machine-learning surrogates are efficient, but offline models often lack state management, rolling prediction, and physical safeguards for continuous simulation. This paper presents an FMI(Functional Mock-up Interface)- compatible fast thermal prediction FMU (Functional Mock-Up Unit) for data center server rooms. The FMU embeds a physics-guided graph neural thermal predictor and exposes standardized FMI inputs and outputs for supply-air temperature, fan state, IT load, historical thermal states, and predicted temperatures. Each simulation step performs input validation, historywindow update, model inference, output correction, and FMI variable write-back. A CRAH M1-08 shutdown replay case shows an overall RMSE of 0.284 °C and an average step time of 0.0656 s, supporting near-real-time use in thermal digital twins.