2021 IEEE International Conference on Fuzzy Systems (FUZZ-IEEE) | 2021
High-dimensional Multivariate Time Series Forecasting using Self-Organizing Maps and Fuzzy Time Series
Abstract
Machine learning models that follow the FTS (Fuzzy Time Series) approach stand out as data-driven non-parametric models of easy implementation and high accuracy, which can be applied to uni-variate and multivariate time series. However, this approach encounters difficulties when dealing with databases of many variables, given the explosion of rules that are generated for the construction of models. Usually filter and wrapper techniques (e.g. Boruta test) and data projection techniques (e.g. Principal Component Analysis) are used. The present work proposes a methodology for tackling this issue by projecting the original high-dimensional data into a low dimensional embedding space using self-organizing Kohonnen maps and later using the Weighted Multivariate FTS method (WMVFTS) for rule discovery and forecasting. The results obtained showed good values of RMSE and MAPE, illustrating the validity and potential of the method.