<p>In recent years, transformer-based architectures have recently revolutionized time-series modeling, delivering major advancements in multivariate time-series forecasting by leveraging powerful spatial–temporal feature extraction capabilities. However, despite their success, state-of-the-art (SOTA) transformer models built upon time decomposition approaches still encounter significant limitations, including challenges in handling noisy and uncertain data, limited interpretability of attention mechanisms, and suboptimal temporal feature decomposition for complex forecasting tasks. To overcome these challenges, in this paper, we propose a novel model, naming: FLDT, which is a fuzzy-enhanced learnable decomposition transformer. Our FLDT model can integrate neuro-fuzzy learning principles into the transformer framework. Our proposed FLDT model can effectively incorporate Gaussian membership functions as well as multiple fuzzy rules within its multi-headed attention mechanism; as a result, directly enhancing the model's robustness to noise and improving interpretability. This integration enables the FLDT model to dynamically capture complex spatial–temporal dependencies while effectively preserving critical patterns in multivariate time-series data. Extensive experiments on different real-world time-series stock datasets demonstrate that our proposed FLDT model significantly outperforms existing transformer-based methods in both forecasting accuracy and model robustness. These results confirm the effectiveness of embedding neuro-fuzzy mechanisms into transformer-based decomposition frameworks; as a result, better providing a more resilient and interpretable solution for time-series forecasting. Overall, our work establishes a novel paradigm that bridges transformer architectures with fuzzy logic systems, offering an advanced approach for handling noise, uncertainty, as well as complex dependencies in multivariate forecasting problems.</p>