Plag-Llama for IoT-Enabled Data Centers: A Multivariate Time Series Forecasting Approach
Yu Sun, Bo Hao Cheng, Jinan Li, Gaoxiang Jiang, Tianqi Zhang, Haibo Zhou · 2024
By harnessing advanced Internet-of-Things (IoT) technologies, deep-learning methods have garnered significant attention for accurately predicting data center status, serving as the cornerstone for mitigating the exponential growth of energy consumption in data centers. However, these methods encounter data scarcity issues during practical deployment. While the proliferation of large models holds promise for addressing this challenge, their application in the context of data centers remains largely unexplored. Furthermore, these models encounter diverse obstacles, including multivariate tasks, immersive computation, and so on. In this paper, we investigate multivariate time series forecasting in IoT-enabled data centers by harnessing large models. Specifically, we introduce a multivariate time series forecasting framework tailored for IoT-enabled data centers. We propose the point Lag (Plag)-Llama model, which transfers univariate forecasting knowledge from the Lag-Llama for multivariate forecasting. The Plag-Llama benefits from zero-shot ability and fine-tuning facilitated by our proposed transfer block. To mitigate computational intensity and enhance the Plag-Llama’s performance, we introduce a novel Joint Channel-Time (JCT)-adapter fine-tuning technique. Extensive experiments demonstrate that the transferred Plag-Llama exhibits superior zero-shot ability, while the proposed JCT-adapter fine-tuning achieves state-of-the-art performance and remarkable few-shot ability on real-world datasets collected from data centers.