ABSTRACT Existing data centers face significant challenges in meeting QoS requirements for resource scheduling while preserving energy efficiency under dynamic and heterogeneous workloads. To address these limitations, we introduce a two-stage prediction framework that integrates a set of complementary techniques, including data cleaning, digital filtering, optimization, and time-series forecasting, to enable proactive resource management and efficient scaling. The contribution of this research is twofold. On the one hand, we design an advanced preprocessing pipeline in which the Local Outlier Factor (LOF) removes abnormal workload spikes, the Savitzky–Golay (SG) filter smooths high-frequency noise, and Empirical Mode Decomposition (EMD) splits nonlinear and non-stationary signals into multi-scale components. On the other hand, we develop a hybrid model that leverages CNN to extract spatial feature correlations, BiLSTM to capture long-term temporal dependencies, and an attention mechanism to highlight the most informative time steps. The Bitbrains fastStorage dataset is used to evaluate the proposed model across multiple time intervals. Overall, the enhanced CNN–BiLSTM–Attention model delivers superior performance (MSE = 0.030, RMSE = 0.173, R 2 = 0.89, Time = 21.5 s), outperforming all baseline models. An ablation study is conducted to assess the contribution of each component to the overall system performance. Importantly, selecting an optimal subset of 7–9 IMFs (Intrinsic Mode Functions) reduces noise, extracts meaningful oscillatory patterns, and enables faster and more accurate modeling.
Marwan et al. (2026) studied this question.