Scholarlink

<p>The traditional power consumption prediction algorithms are no longer suitable for deep learning-based workloads owing to the long-term dependencies and highly volatile nature of power usage in such applications. To address these challenges, this study proposes ENNP, an enhanced neural network-based power consumption prediction algorithm for deep learning-based workloads in AI-enabled data centers. In general, ENNP inherits the advantages of the TimeMixer framework in mitigating model overfitting. However, direct application of TimeMixer does not provide the flexibility needed to handle complex power consumption patterns. To overcome this limitation, ENNP integrates three key techniques: Bayesian optimization, KAN, and LSTM, to maximize the predictive performance of TimeMixer. Specifically, (1) Bayesian optimization is employed for parameter selection and extraction of multiscale time series; (2) KAN is utilized to reconstruct multiscale time series; and (3) an improved LSTM is adopted to predict on multiscale time series and fusing these predictions to obtain the final prediction. Finally, experiments were conducted on real CPU/GPU servers using trace data collected from various deep learning workloads, including natural language processing, image recognition, multivariable prediction, and anomaly detection jobs. The experimental results show that ENNP achieves the best performance compared with other well-known prediction algorithms such as BP, LSTM, and MLR, as well as recent general time-series prediction models, with a maximum improvement of 38.29% in accuracy.</p>

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

ENNP: an enhanced neural networks-based power consumption prediction algorithm to deep learning-based workloads on AI-enabled data centers

  • 期刊论文

  • 发表日期：

    2025年11月19日

查看全文
  • Hengdong Wang,
  • Chao Jing

摘要

The traditional power consumption prediction algorithms are no longer suitable for deep learning-based workloads owing to the long-term dependencies and highly volatile nature of power usage in such applications. To address these challenges, this study proposes ENNP, an enhanced neural network-based power consumption prediction algorithm for deep learning-based workloads in AI-enabled data centers. In general, ENNP inherits the advantages of the TimeMixer framework in mitigating model overfitting. However, direct application of TimeMixer does not provide the flexibility needed to handle complex power consumption patterns. To overcome this limitation, ENNP integrates three key techniques: Bayesian optimization, KAN, and LSTM, to maximize the predictive performance of TimeMixer. Specifically, (1) Bayesian optimization is employed for parameter selection and extraction of multiscale time series; (2) KAN is utilized to reconstruct multiscale time series; and (3) an improved LSTM is adopted to predict on multiscale time series and fusing these predictions to obtain the final prediction. Finally, experiments were conducted on real CPU/GPU servers using trace data collected from various deep learning workloads, including natural language processing, image recognition, multivariable prediction, and anomaly detection jobs. The experimental results show that ENNP achieves the best performance compared with other well-known prediction algorithms such as BP, LSTM, and MLR, as well as recent general time-series prediction models, with a maximum improvement of 38.29% in accuracy.