Fusing Interval Type-2 Fuzzy Logic With Bidirectional RNNs for Uncertainty-Aware Time Series Forecasting

With the explosive growth of multidimensional data, uncertainty in time series forecasting has become increasingly prominent. Although deep learning models demonstrate strong predictive capabilities, they remain sensitive to data uncertainty, while fuzzy logic systems (FLS) offer interpretability but often suffer from predefined parameter sensitivity and high computational overhead. To address these challenges, this paper proposes the IT2-FLS-BiRNN framework, which combines the temporal feature extraction capabilities of Bidirectional Recurrent Neural Networks (BiRNNs) with the uncertainty quantification strengths of Interval Type-2 FLS (IT2-FLS). To overcome limitations related to membership function sensitivity and constrained generalization, a fuzzy membership learning strategy is introduced to enable a data-aligned transformation of input features into IT2 fuzzy sets. Moreover, to alleviate the considerable computational overhead commonly associated with fuzzy systems, a parallelizable IT2-FLS inference mechanism is developed, which leverages matrix-based parallel computation for efficient reasoning. To evaluate the effectiveness and robustness of the IT2-FLS-BiRNN framework, experiments are conducted on ten publicly available benchmark datasets with varying levels of noise and nonlinear characteristics. Experimental results show that the proposed method consistently achieves superior performance across multiple evaluation metrics and outperforms most baseline models. In addition, noise analysis further reveals that IT2-FLS-BiRNN maintains significantly lower prediction errors compared to competing methods, particularly under high noise conditions, thereby highlighting its robustness in managing uncertainty for time series forecasting.