Long-Term Time Series Forecasting With Multilinear Trend Fuzzy Information Granules for LSTM in a Periodic Framework

Considerable research achievements have been made in utilizing information granulation as an effective technique for addressing long-term time-series forecasting. However, existing studies suffer from limitations in their failure to account for the impact of periodicity on information granulation. As a result, the predictive outcomes of the model are significantly restricted, including issues, such as the accumulation of original errors and low temporal correlation. To address the aforementioned issues, this article presents a novel approach for long-term time series forecasting by employing a multilinear trend fuzzy information granule (FIG) within a periodic framework for LSTM. The proposed model tackles the problems through several key steps. First, the time window is divided during the information granulation process to enhance the periodic information of the time series. Second, a dynamic time warping (DTW) information granule segmentation and merging algorithm is utilized to adjust the division of information granules, ensuring periodic consistency and alignment with the time series' change pattern. Lastly, Gaussian linear FIGs and LSTM are employed to achieve consistent forecasting of data points with consistent trend characteristics, including trend changes, fluctuation magnitudes, and trend persistence. This approach effectively prevents the accumulation of errors over time and enhances the temporal correlation of the data. Experimental results on open time series demonstrate the superior performance of the proposed forecasting model.