DCITNet: A Temporal Forecasting Scheme for IT Power Demand of Data Center
Y. D. Wang, Wenyu Liu, Shijie Chen, Yuejun Yan, Zhaoyang Wang, Yimeng Sun, Xuan Wei, Jun Shu, Zhaohao Ding · IEEE Transactions on Industry Applications · 2025
Accurate forecasting of Information Technology (IT) power demand is essential for the operation of data centers. However, the complex coupling relationships between various factors and IT power demand pose a challenge to its prediction. The widespread application of free cooling technology further highlights this problem, which makes it difficult for traditional models based only on CPU utilization and historical load to accurately forecast IT power demand at different time scales. To address these challenges, we propose DCITNet, an innovative end-to-end framework designed for IT power demand forecasting under the water-side free-cooling system. Specifically, we first extract the primary coupling mechanisms applicable for forecasting under water-side free-cooling conditions by analyzing the relationships between IT factors (e.g., CPU utilization), non-IT factors (e.g., environmental factors, auxiliary equipment), and IT power demand. Secondly, we extend these coupling mechanisms into three neural network modules. We establish two separate modules to capture the periodic coupling characteristics of IT factors and the time-lagged coupling characteristics of non-IT factors. Then we introduce a fusion module based on crossattention to integrate the coupling characteristics between IT factors and non-IT factors, which enables the precise forecasting of IT power demand. Finally, our proposed method shows significant improvements in accuracy through experiments on a realworld dataset and comparisons with state-of-the-art algorithms. Specifically, it improves the accuracy of short-term forecasting by more than 20% and improves the accuracy of mid- to long-term forecasting by around 50%.