Abstract
Time series modeling and prediction are very attractive topics, which play an important role in many fields such as transportation prediction [4], power prediction [13, 18], and health care study [7]. The purpose of time series prediction is to forecast the values of data points ahead of time, where long-term time series prediction is to make the predictions multi-step ahead. The prediction process is commonly performed by observing and modeling the past values, and assuming that the future values will follow the same trend. When the prediction horizon increases, the uncertainty of the future trend also increases, rendering a more challenging prediction problem. Researchers have dedicated their effort to study how to extract as much knowledge as possible from the past values, and how to better utilize such knowledge for long-term time series prediction. There has been previous research work in order to tackle this challenge based on some classical time series prediction approaches, such as exponential smoothing [12], linear regression [14], autoregressive integrated moving average (ARIMA) [33], support vector machines (SVM) [25], artificial neural networks (ANN) [10, 33], and fuzzy logic [10].