FD-LSTM: A Fuzzy LSTM Model for Chaotic Time-Series Prediction

The main issue of time-series prediction is to determine the grade of uncertainty in knowledge, with its essential vagueness and haziness in complex problems. In this study, a deep fuzzy long short-term memory (LSTM) architecture has been proposed to handle the high-order uncertainty associated with time-series applications. The LSTM and type-2 fuzzy logic combination aim to make a more transparent, interpretable, and accurate predictive system. The experiments of this study with real data contain global standard and real-valued benchmarks, including Mackey–Glass (MG), sunspot, and English Premier League seasonal datasets. The obtained performance shows the superiority of the proposed fuzzy–deep model in predicting time series with an average AUC = 0.96 in sunspot, 0.93 for football match time series, and 0.95 on a chaotic equation of MG.