A New Fuzzy Inference System Applied to Time Series Forecasting

Abstract

Fuzzy systems are a class of machine learning introduced by Zadeh that combine accuracy and interpretability. This class of models consists of two main parts, the antecedent and the consequent. While the antecedent is responsible for modeling the inputs, the consequent concerns modeling the output. The literature reports two main types of fuzzy systems: Mamadani and Takagi-Sugeno. While Mamdani uses fuzzy sets in the consequent part, Takagi-Sugeno uses polynomial functions. Consequently, Mamdani provides better understandable models and Takagi-Sugeno more accurate ones. In this paper, we propose a new Takagi-Sugeno model. Still, instead of defining the rules based on the input, the proposed model designs the rules based on the output variation to capture linearities in the output and clusters them in the same rule. The model is applied in the regression problems of benchmark series and real datasets of power transformers. The performance of the proposed model is compared with the performance of classical models and evolving Fuzzy Systems. The results are evaluated using error metrics, the number of final rules, and runtime.