Predicción mejorada del índice de calidad del aire a través de un marco integrado de aprendizaje profundo

Accurate air quality forecasting is vital for public health, urban planning, and environmental policy. Traditional models often fail to capture the complex temporal, spatial, and topological dependencies inherent in air quality data. This study presents a hybrid deep learning model that integrates long short-term memory (LSTM) networks for sequential prediction, convolutional neural networks (CNNs) for spatial feature extraction, and graph neural networks (GNNs) for modeling spatial correlations between monitoring stations. To further enhance prediction robustness, ensemble learning techniques-stacking and boosting-are employed to fuse the outputs of individual models. The model is evaluated using real-world datasets from air quality monitoring agencies, covering multiple pollutants and meteorological parameters. Experimental results show significant improvements in performance metrics, achieving R² scores exceeding 0.92 and reduced prediction errors compared to baseline models. Visualization of predicted vs. actual AQI values and residuals confirms the model's accuracy and consistency. Specifically, the hybrid model achieved an MAE of 11.30, RMSE of 15.60, and R² of 0.89 on the Beijing dataset, outperforming standalone LSTM (MAE 14.55, R² 0.80), CNN (MAE 15.12, R² 0.78), and GNN (MAE 13.67, R² 0.83). In cross-dataset evaluation, the model attained an MAE of 12.62 and R² of 0.85 when trained on Beijing and tested on Los Angeles, confirming its robustness and transferability. The proposed framework demonstrates strong generalizability and interpretability, making it a promising tool for environmental monitoring and decision-making in pollution-sensitive regions.