Multivariate Long-Term Forecasting Using Multilinear Trend Fuzzy Information Granules for Traffic Time Series

Abstract

Long-term forecasting for time series is gaining significant attention in many emerging fields, such as machine learning and artificial intelligence. Linear fuzzy information granulation is a well-recognized and powerful tool for long-term forecasting. However, existing studies on this topic have been restricted to univariate time series, falling short of meeting the real-world demands of complex forecasting issues. In this regard, this article proposes a multivariate long-term time-series forecasting model using fuzzy information granules and demonstrates its application to a transportation forecasting task. Our model begins by designing a multilinear trend fuzzy information granulation scheme within a justifiable granularity framework. This scheme effectively granulates and delineates time-series information for each feature. To ensure the normalization of the granular time series within each time window, we propose a scale equalization strategy that leverages the dynamic time warping algorithm. Building upon this foundation, our model introduces a long-term forecasting mechanism for multivariate time series. Notably, this mechanism fosters interconnections, bridging the temporal gap between past and future granular time series while concurrently linking the granular series of influential features and the target feature. Empirical validation on six real-world traffic flow forecasting datasets demonstrates the effectiveness and competitiveness of our proposed model.