Scholarlink

<p>The rapid growth of cloud computing has intensified the thermal and energy demands of modern data centres, highlighting an urgent need for accurate temperature prediction to support efficient thermal management. Existing approaches, which largely rely on traditional machine learning models, often fail to capture the nonlinear and dynamic behaviour of cloud workloads. To address this limitation, this study proposes a Multi-Stage Stacked Ensemble Machine Learning (MSEML) framework that integrates two layers of heterogeneous base learners with a ridge regression meta-learner. Each layer contributes to improved prediction accuracy, as demonstrated through an ablation study. The effectiveness of the framework is evaluated against ten benchmark models using multiple regression metrics. Results show that MSEML reduces <i>RMSE</i> by 15.81% and 25.14% relative to RF and KNN, and improves the <i>R</i>² score by 1.18% and 2.27%. Qualitative and uncertainty analyses further confirm its consistent generalisation and reduced uncertainty. In dynamic scheduling experiments, the framework achieves up to 35.28% fewer <i>SLA</i><sub><i>violations</i></sub> and a 5.18% reduction in peak temperature. SHAP-based interpretability also offers meaningful insights into feature influence. Overall, MSEML provides an efficient, accurate, and interpretable solution for thermal and resource management in cloud data centres.</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

Thermal prediction for efficient resource management in cloud data centres using a multi-stage stack ensemble machine learning model

  • 期刊论文

  • 发表日期：

    2026年3月3日

查看全文
  • Hirdesh Varshney,
  • Avtar Singh

摘要

The rapid growth of cloud computing has intensified the thermal and energy demands of modern data centres, highlighting an urgent need for accurate temperature prediction to support efficient thermal management. Existing approaches, which largely rely on traditional machine learning models, often fail to capture the nonlinear and dynamic behaviour of cloud workloads. To address this limitation, this study proposes a Multi-Stage Stacked Ensemble Machine Learning (MSEML) framework that integrates two layers of heterogeneous base learners with a ridge regression meta-learner. Each layer contributes to improved prediction accuracy, as demonstrated through an ablation study. The effectiveness of the framework is evaluated against ten benchmark models using multiple regression metrics. Results show that MSEML reduces RMSE by 15.81% and 25.14% relative to RF and KNN, and improves the R² score by 1.18% and 2.27%. Qualitative and uncertainty analyses further confirm its consistent generalisation and reduced uncertainty. In dynamic scheduling experiments, the framework achieves up to 35.28% fewer SLAviolations and a 5.18% reduction in peak temperature. SHAP-based interpretability also offers meaningful insights into feature influence. Overall, MSEML provides an efficient, accurate, and interpretable solution for thermal and resource management in cloud data centres.