Ensemble Time Series Modeling for High Precision Cloud Memory Usage Prediction

Abstract

This study presents a novel approach to predicting cloud resource usage, focusing on memory allocation in a Google cluster environment. By combining traditional time series analy- sis with advanced machine learning techniques, we developed a highly accurate predictive model that significantly outperforms existing methods. Our research began with ARIMA and SARI- MAX models, providing insights into temporal patterns, and progressed to more sophisticated Prophet and Random Forest models, which greatly improved predictive accuracy. The key contribution of this study is the development of an ensemble model that combines predictions from Prophet and Random Forest. This innovative approach consistently outperformed individual models, achieving a remarkable mean squared error (MSE) of 1.678e-05 and a mean absolute error (MAE) of 0.002418. These results represent a 3.5-fold improvement over the best individual model, with predictions deviating on average by less than 0.25% from actual values. Our research demonstrates the potential to revolutionize cloud resource management through highly accurate predictions. The ensemble model’s exceptional performance suggests strong potential for real-world applications, potentially leading to significant enhancements in capacity planning, resource allocation, and overall efficiency in cloud computing environments. Furthermore, the methodological framework developed in this study, which combines statistical rigor with machine learning flexibility, offers a promising approach for addressing complex forecasting chal- lenges across various domains in cloud computing and beyond.