Search

To reduce energy consumption in data center cooling systems and ensure the healthy and rapid development of data centers, this study investigated on the cooling system of a data center in Guangzhou. Aimed at the efficiency variations of chillers at different Partial Load Rates (PLR), a novel model predictive control method was proposed. The study employed Gaussian process regression to establish a chillers' operational efficiency model and utilized Long Short-Term Memory (LSTM) neural networks to create an outdoor wet bulb temperature time series prediction model. The cooling system's chilled water storage and discharge modes, storage and discharge capacity, the number of chillers, and the PLR were adjusted based on the data center cooling load and the prediction of outdoor wet bulb temperature. This adjustment aims to partially decouple the demand and supply sides, resulting in the chillers being operated in high efficiency or turned off, thereby reducing the operational energy consumption of the cooling system. Compared to the traditional rule-based control strategy, the model predictive control strategy reduced the overall energy consumption by 3.13% and operating costs by 3.33%.

Thermal issue has been regarded as one of the bottlenecks for GaN high-electron-mobility transistor (HEMT) performance and reliability, which highlights the importance of accurate thermal modeling. In the present work, we propose a GP (Gaussian process)-resistor–capacitor compact thermal model integrated with the ensemble Kalman filter (EnKF) to handle the nonlinear problems attributed to the temperature-dependent properties of GaN HEMTs under large-signal working conditions. The GP predictor is employed for the nonlinear correction term, with strong ability and extendibility to characterize various temperature-dependent relations resulting from different design configurations and materials. The model is identified via the EnKFs by inputting a sequence of channel temperature oscillations induced by imposing a large-signal continuous wave heating source to the device. Furthermore, an adaptation mode is devised for the in situ and timely update of the model parameters to adapt to the thermal variability of GaN devices, avoiding storing a large amount of historical data and repeated offline regressions. The validation of our modeling scheme is conducted through the case study on GaN-on-SiC HEMT's detailed 3D finite element method simulations. [ABSTRACT FROM AUTHOR]