A time-varying state-space model for real-time temperature predictions in rack-based cooling data centers

Abstract

Fast-growing data centers (DCs) require efficient cooling systems (such as rack-based cooling architectures) and control strategies to reduce operating costs and guarantee desired indoor conditions. Thus, this study proposed a novel real-time temperature prediction model for rack-based cooling DCs, in order to facilitate advanced control regarding cooling management and workload assignment. Specifically, a data-driven technology was introduced to estimate time-invariant model parameters, in order to avoid the time-consuming physics-based parameters extracting process. The mass conservation relationships were employed to update time-varying flow parameters in real-time to capture the nonlinear behaviors in DCs. Moreover, the proposed control-oriented thermal modeling method can model hot air recirculation and cold air bypass occurring simultaneously for the first time. The performance of the developed time-varying state-space model was validated by CFD simulation data. Additionally, the timeliness of modeling and temperature prediction was also investigated. The results show that the developed model achieves sufficient accuracy with a mean absolute error (MAE) equal to 0.28 °C, even for long prediction horizons and dynamic IT workloads. Also, the developed model has outstanding timeliness for advanced control techniques, in terms of less than 30 min for parameter identification and less than 10 s for temperature prediction.