Scholarlink

Data centers have become crucial infrastructure, continuously providing services for large-scale distributed computing, computational networks, and large model training. Accurate prediction of future power consumption in data centers is essential for ensuring system efficiency and energy management. However, data centers often employ multi-layered architectures and consist of heterogeneous devices, leading to highly complex and chaotic time-series information, making overall power consumption prediction challenging. To address this issue, this paper analyzes the complexity of aggregated time series from both horizontal and vertical perspectives, tackling the challenges posed by data aggregation in multi-layered architectures. We introduce AnaNET, an innovative analytical network designed specifically to dissect and model the complex and chaotic nature of power consumption. AnaNET utilizes a novel frequency decomposition and inverse progressive decomposition method, isolating and analyzing different periodic patterns within the aggregated data. AnaNET demonstrated approximately 22% relative improvement across six real-world time-series datasets. Code is available at: https://github.com/hihiztc1/AnaNET_NPC .

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

AnaNET: Anatomical Network for Aggregated Time Series Forecasting in Multi-layered Architecture

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Tiancheng Zhang,
  • Cheng Zhang,
  • Shuren Liu,
  • Xiaofei Wang,
  • Shaoyuan Huang

摘要

Data centers have become crucial infrastructure, continuously providing services for large-scale distributed computing, computational networks, and large model training. Accurate prediction of future power consumption in data centers is essential for ensuring system efficiency and energy management. However, data centers often employ multi-layered architectures and consist of heterogeneous devices, leading to highly complex and chaotic time-series information, making overall power consumption prediction challenging. To address this issue, this paper analyzes the complexity of aggregated time series from both horizontal and vertical perspectives, tackling the challenges posed by data aggregation in multi-layered architectures. We introduce AnaNET, an innovative analytical network designed specifically to dissect and model the complex and chaotic nature of power consumption. AnaNET utilizes a novel frequency decomposition and inverse progressive decomposition method, isolating and analyzing different periodic patterns within the aggregated data. AnaNET demonstrated approximately 22% relative improvement across six real-world time-series datasets. Code is available at: https://github.com/hihiztc1/AnaNET_NPC .