An Integrated Adaptive Neuro-Fuzzy Inference with Attention Mechanism for Enhancing Performance of Time-Series Weather Forecasting Through Transformer

In recent years, significant advancements in machine learning (ML) and deep learning (DL) have greatly propelled progress in time-series data analysis and forecasting. Among these, the emergence of sequence-to-sequence (seq2seq) as well as transformer-based architectures has revolutionized the field by providing powerful mechanisms—for modeling sequential data and capturing long-term temporal dependencies. However, the inherent complexity of transformer-based predictive models often leads to a lack of interpretability; therefore, posing challenges for practical applications—that require a clear understanding of the dynamic representations underlying time-dependent weather data. To address this issue, this paper introduces FAttn4TS, a novel fuzzy-driven enhanced attention mechanism integrated—within a transformer-based architecture to improve the accuracy and interpretability of weather forecasting. The proposed FAttn4TS model uniquely combines an adaptive neuro-fuzzy inference system (ANFIS) with the multi-headed attention mechanism; as a result, enabling the model to effectively manage uncertainty and provide transparent insights into temporal feature representation through fuzzy reasoning. This integration complements the transformer’s strength in capturing complex temporal dependencies while enhancing robustness as well as interpretability. Extensive experiments conducted on real-world weather datasets demonstrate that our proposed FAttn4TS model consistently outperforms existing baseline models in both predictive accuracy and explainability.