Time Series Forecasting Based on Improved Multi-Linear Trend Fuzzy Information Granules for Convolutional Neural Networks

Although the construction of multilinear trend fuzzy information granules (FIG) achieves a win-win situation in terms of interpretability and trend extraction, in its second stage of segmentation, the equal-length segmentation will result in the loss of local trend. The granulation effect will further affect the forecasting performance of the time series. To this end, this paper establishes a convolutional neural network (CNN) prediction method based on improved multilinear trend FIGs. Firstly, considering the natural cycle characteristics of the time series, this paper establishes a time series segmentation algorithm based on the valley points, which replaces the equal-length segmentation in the second stage of the construction of the multilinear trend FIGs, thus enhancing the interpretability of the granulation process. Later, an evaluation index of Gaussian fuzzy information granules (GLFIGs) is proposed for improving the trend extraction effect of each multilinear trend FIG. Since the multilinear trend FIGs are constructed in the natural period segment, in order to fully exploit the correlation of the corresponding positions of each granule to enhance the prediction accuracy, a GLFIG correspondence algorithm based on the segmentation and merging is introduced in this paper. Finally, CNN is selected as the prediction model based on the data characteristics. We conduct experiments on 6 datasets and 2 artificial cycle datasets, and compare the constructed model with commonly used prediction models and the latest granularity model. At last, the experiments reveal that our model performs better.