Multi-Step Ahead Time Series Prediction Process Based on CEEMDAN Decomposition and Recursive Update Strategy

Multi-step ahead time series prediction can be seen as one of the most complicated tasks in data mining due to the nature of strong uncertainties inside the time series. Deep learning models are the prominent approach to deal with this problem due to their advantages and accurate prediction results. This paper proposes a deep learning model based on CEEMDAN decomposition and recursive update strategy. The crucial idea of our proposed model is to group subsequences generated by CEEMDAN decomposition into three groups that are characterized by the stationery and the frequency of time series. The evaluation study shows a better performance of our results than other benchmark models.