High-Precision Air Quality Prediction via Attention-Driven Hybrid Neural Networks and Adaptive Feature Optimization

Rapid urbanization and industrialization have intensified air pollution, posing severe challenges to sustainable development and public health. As a core economic zone in China, the Beijing–Tianjin–Hebei (BTH) region faces persistent air quality deterioration, highlighting the urgent need for accurate and intelligent prediction models. However, existing studies often suffer from limited adaptability of single models and subjective feature selection thresholds, constraining predictive performance and generalization capability. To address these challenges, this study proposes a feature-optimized hybrid deep learning framework for AQI prediction across Beijing, Tianjin, and Shijiazhuang. An adaptive feature selection strategy is first developed by integrating the Relief_F algorithm with the Bat Optimization Algorithm (BOA), which adaptively determines feature importance, thereby enhancing objectivity and effectiveness in identifying key pollutant and meteorological indicators. Subsequently, an attention-enhanced CNN–BiLSTM–GRU hybrid network is constructed, where the attention mechanism emphasizes critical temporal information that most influences prediction results. Experiments show that the proposed model achieves MAPE values of 1.00%, 1.15%, and 1.09% for Beijing, Tianjin, and Shijiazhuang, outperforming benchmark models by 18.43–45.05%. These results confirm the framework’s reliability for practical application with strong robustness and statistical validity.