A novel attention-based deep learning model for accurate PM2.5 concentration prediction and health impact assessment

Abstract

Air pollution is a significant global health hazard, especially in developing, low-income countries with limited resources to address its impacts. Among pollutants, PM2.5 is particularly concerning due to its challenging containment and severe health implications. This study proposes a novel attention augmented hybrid deep learning (DL) model in multi-directed mode to predict the PM2.5 level accurately. The attention mechanism taps the long-term temporal dependencies in the latent vector space. Moreover, convolutional neural network and long short-term memory-based hybrid DL model focuses on short-term temporal dependencies in the feature space. The proposed model dynamically adjusts the focus with alignment score for efficient representation of the dataset, thereby outperforming standard deep learning benchmarks by 4.28 % compared to RNN, 10.5 % compared to LSTM, and 5.7 % compared to GRU. The utilization of ensemble technique in multi-directed mode enables the model to address the complex data dependencies. Subsequently, Bayesian hyperparameter optimization revealed that lower learning rates (1.60 × 10⁻⁶) combined with tanh activation functions and increased dense nodes yielded optimal performance. Additionally, quantitative healthcare impact assessment indicates that improved prediction accuracy potentially reduces direct healthcare economic burden by $82.4 million USD. This research provides a robust framework for PM2.5 forecasting that supports enhanced public health risk management and policy implementation.