A hybrid granular model with decomposition-integration strategy toward long-term prediction of time series

Time series data usually exhibit high dimensionality and multiple intricate features, making accurate prediction of future fluctuation range challenging. Accurate long-term prediction is particularly crucial for guiding management activities across various fields. To capture underlying fluctuation range patterns and improve long-term prediction accuracy, this study offers a novel decomposition-integration hybrid prediction model. First, numerical time series are converted into granular time series, realizing a feature representation of the original time series within a low-dimensional space. Then, three significant components of the granular version sequence are extracted for representing its frequency features using bivariate empirical mode decomposition and sample entropy reconstruction. Next, the individual frequency sequences are incorporated into a higher-order fuzzy cognition map to generate the causal relationships for capturing the internal relationships within each component sequence as well as the mutual impacts between distinct component sequences. The output represented by information granules effectively reflects future fluctuation range. Empirical results based on eight datasets demonstrate that the established hybrid prediction model yields ideal accuracy, verifying its effectiveness in minimizing error accumulation, and achieving sound long-term predictive performances.