A New Granular Approach for Multivariate Forecasting

The research for computationally cheaper, scalable and explainable machine learning methods for time series analysis and forecasting has grown in recent years. One of these developments is the Fuzzy Time Series (FTS), simple and fast methods to create readable and accurate forecasting models. However, as the number of variables increase the complexity of these models becomes impractical. This work proposes the \(\mathcal {FIG}\)-FTS, a new approach to enable multivariate time series to be tackled as univariate FTS methods using composite fuzzy sets to represent each Fuzzy Information Granule (FIG). \(\mathcal {FIG}\)-FTS is flexible and highly adaptable, allowing the creation of weighted high order forecasting models capable to perform multivariate forecasting for many steps ahead. The proposed method was tested with Lorentz Attractor chaotic time series and the GEFCom 2012 electric load forecasting contest data, considering different forecasting horizons. The results showed that the Mean Average Percentual Error of the models was at about 2% and 4% for one step ahead, and for a prediction horizon of 48 h, the MAPE is at about 10%.