Hybrid Deep Learning for Air Quality Prediction: A Multi-Output, Attention-Based Approach for Pollutant and AQI Classification

Abstract

Air pollution is a critical worldwide issue requiring precise forecasting for the deployment of efficient proactive interventions. The current study proposes a hybrid deep learning model based on multi-head attention mechanisms, bidirectional LSTMs, and dense layers to forecast the overall Air Quality Index (AQI) and identify specific thresholds of pollutant severity. The model utilizes deep learning algorithms for predictive accuracy to process primary air pollutants such as PM2.5, PM10, NO₂, CO, and O₃. Robustness tests were conducted by using different performance measures such as the F1-score, the Precision-Recall Curve (PRC), the Receiver Operating Characteristic (ROC) Curve, and training loss, accuracy patterns, and a confusion matrix. The outcome reveals that the proposed model performs exemplary classification with robust precision-recall scores and well-optimized ROC curves, thus proving to be effective in discriminating across different levels of pollution severity. F1-score analysis reveals tremendous success in the detection of cleaner air conditions but shows minor misclassifications in high AQI levels, signifying areas for improvement. Reliability of the model for real-world application is further established with training loss and accuracy curves revealing a smooth learning pattern with limited overfitting. Through attention-driven learning paradigms, the model proposes a scalable and adaptive solution for real-time monitoring of air quality, enabling efficient decision-making for pollution mitigation measures. The paper contributes to the emerging domain of deep learning in environmental science through the demonstration of hybrid AI-driven models for predictive air quality modeling. Future refinement will involve incorporation of other meteorological variables and spatiotemporal inputs to enhance the performance of the model.