A decomposition-guided mechanism for nonstationary time series forecasting

Time series forecasting has been playing an important role in decision making, control, and monitoring across various fields. Specifically, the forecasting of nonstationarity time series remains a challenging problem where traditional time series modeling may not fully capture temporal dynamics. Recent studies of applying machine learning (ML) or more advanced hybrid models combining the ML and decomposition methods have shown their flexible nonstationary and nonlinear modeling capability. However, the end-effect problem introduced by the decomposition methods still introduces significant forecasting errors because of the unknown realm beyond the time series boundary. Therefore, a novel method applying a decomposition-guided mechanism is proposed in this work to eliminate the end effect problem while inheriting the knowledge learned from the decomposition state space to improve the prediction accuracy of such hybrid models in time series forecasting. Additionally, a domain adaptation model is integrated with the proposed mechanism to transfer knowledge from the source domain to the target domain regarding the decomposition state space. In this work, the intrinsic time-scale decomposition and Gaussian process are considered as examples of decomposition and ML methods to demonstrate the proposed mechanism's effectiveness. Both simulation experiments and real-world case studies are conducted to show that a hybrid model with the proposed mechanism outperforms the conventional time series forecasting model, pure ML, and the original hybrid model in terms of prediction accuracy.