Harnessing Ensemble Deep Learning for Scalable and Precise Cloud Workload Prediction

Abstract

Cloud computing platforms face a massive demand for resources, which is typically volatile and unpredictable, and require accurate and scalable prediction techniques to assure appropriate resource allocation and service continuity in such environments. Cloud workloads' increasing variability and complex nature render traditional resource management inefficient since traditional prediction models mostly use static parameters. In this study, a Transformer-GRU-BiLSTM, which is called an Ensemble model for Workload Prediction, is presented with the combination of TransGRU-BiLSTMNet (TGBNet) multiple DL models (Gated Recurrent Unit (GRU), Bi-directional LSTM (BiLSTM) and Transformer-based architectures) to handle the accurate and complex workload patterns. The weighted average and meta-learning methods are applied in ensemble models to enhance the generalization and stability of workload changes. Experimental results based on extensive empirical studies over real-world public and private cloud workload datasets show that our integrated ensemble model not only yields better prediction accuracy compared to a set of individual deep learning (DL) models and traditional statistical methods but also possesses better scalability and time complexity. The study is one of the first to propose ensemble deep learning (DL) for proactive resource management in cloud environments, enabling better workload distribution and reduced computational costs.