Echo state network and long short-term memory encoder-decoder approaches for time series forecasting

Abstract

In this dissertation, we propose and evaluate neural network approaches for time series prediction and resources allocation in cloud computing. These include a novel Echo State Network (ESN) model with inputs reconstruction, Bayesian Ridge Regression, and Independent Component Analysis for chaotic time series prediction, an Echo State Network based ensemble for workload prediction and resources allocation of Web applications in the cloud, and a Long Short-Term Memory Encoder-Decoder (LSTM-ED) model for host load prediction in cloud computing. By experimenting on synthetic and real world datasets including one month trace of a Google data center, we show that our approaches outperform other ESN variants, component models of the ESN-based ensemble and other ensemble models, and current state-of-the-art host load prediction methods in term of prediction accuracy. Furthermore, we show that our ESN-based ensemble model has higher resource allocation capability than the component models and other ensemble models when applying to real-world Web server request logs.