<p>Accurate forecasting of the Air Quality Index (AQI) is essential for mitigating the adverse environmental and public health impacts associated with air pollution. However, existing predictive models often exhibit limited robustness to extreme outliers and struggle to effectively capture multi-scale temporal dependencies and associated uncertainties. To address these challenges, this study proposes a novel Multi-scale Quantile Neural Network (MQNN) framework that integrates Fourier expansion for temporal feature extraction and the pinball loss function for probabilistic quantile estimation. Specifically, the AQI time series is first decomposed via Fourier expansion to capture multi-scale temporal dynamics. These transformed features are then processed through multiple parallel subnetworks, each trained to predict a distinct quantile using the pinball loss. The overall training objective aggregates the loss across all time steps and quantiles, with optimization proceeding until convergence. The proposed MQNN model is empirically evaluated using AQI datasets from five major cities in China. Experimental results demonstrate that MQNN consistently outperforms conventional machine learning and deep learning baselines, achieving higher accuracy and more reliable uncertainty estimation. Furthermore, the MQNN framework facilitates the identification of dominant pollutants, offering actionable insights for targeted air quality management and policy interventions. While the MQNN model has strong predictive capabilities, its computational cost remains high for high-dimensional inputs. Moreover, integrating external information may further improve prediction accuracy and enhance the model’s applicability in real-world scenarios.</p>