A Multivariate Cloud Server Aging Forecasting Using PCA, EEMD, and MLP-GBInformer

Accurate aging prediction is critical for preventing failures and ensuring system reliability and availability. However, aging-related time series data often exhibit high fluctuations, strong nonlinearity, and complex interdependencies among multiple indicators, posing significant challenges for effective prediction. This study proposes a novel multivariate hybrid learning approach that fully exploits the interrelationships among heterogeneous aging indicators for cloud server aging prediction. In the data preprocessing stage, the Local Outlier Factor (LOF) algorithm is applied to remove anomalous data, while a Random Forest (RF)-based selection strategy, optimized by the Pelican Optimization Algorithm (POA), identifies key indicators to reduce redundancy and enhance representativeness. The selected multivariate time series is then decomposed into high- and low-frequency Intrinsic Mode Functions (IMFs) using an improved Ensemble Empirical Mode Decomposition (EEMD) method, effectively isolating frequency-specific patterns. For prediction, a Good Beginning Informer (GBInformer) model is designed to capture the rapid fluctuations in high-frequency components, while a Multilayer Perceptron (MLP) models the slow-varying nonlinear trends in low-frequency components. Final aging prediction is obtained by aggregating the outputs of both models across all IMF components. Comparative experiments on OpenStack cloud server datasets demonstrate that the proposed approach outperforms five baseline methods, achieving a MAPE of 0.01 for memory utilization and 0.9903 for response time prediction. These results underscore the effectiveness of the proposed multivariate approach and offer valuable insights for proactive aging management and maintenance of cloud servers.