Enhanced Workload Prediction in Data Centers Using Two-Stage Decomposition and Hybrid Parallel Deep Learning

Workload prediction is one of the most basic requirements in developing cost and energy-efficient Cloud Data Centers (CDCs). Most traditional approaches have suffered from noise and failed to capture the complex dynamic patterns in workload data, reducing their accuracy. To improve this, we introduce CVCBM which blends signal processing techniques Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) and Variational Mode Decomposition (VMD) with advanced deep learning models like Convolutional Neural Networks (CNN) and Bidirectional Long Short-Term Memory (Bi-LSTM) networks. CVCBM utilizes a hierarchical two-stage decomposition process, beginning with CEEMDAN. We use the capability of CEEMDAN for denoising by additive white noise to clean the noisy workload data and further decompose it into several Intrinsic Mode Functions (IMFs), ranging from high to low frequencies. Then, we propose a partitional clustering approach based on Sample Entropy (SE) to select components of similar complexity to increase the effectiveness of the second-stage denoising. After that, VMD is a method based on the center frequency applied to decompose further the high-frequency components, which may contain fluctuations that obscure underlying trends, thus enhancing the model’s overall accuracy. A novel hybrid model is utilized to forecast future workloads, incorporating two sets of three distinct parallel Conv1D layers with varying kernel sizes. These layers extract patterns from the input data, capturing short-term, medium-term, and long-term workload information, allowing the model to learn variations at different scales. Following this, Bi-LSTM layers capture the temporal dependencies within the patterns identified by the Conv1D layers. Extensive experiments on various real-world datasets from Google and Alibaba show that CVCBM significantly outperforms the other methods, which makes it a solution for future workload prediction in cloud data centers.