Development of a Hybrid Attention Transformer for Daily PM2.5 Predictions in Seoul

A hybrid attention transformer (HAT) was developed for accurate daily PM2.5 predictions in Seoul. The performance of the HAT was evaluated through a comparative analysis of its predictions against ground-based observations and those from a three-dimensional chemical transport model (3-D CTM). The results demonstrated that the HAT outperformed the 3-D CTM, achieving a 4.60% higher index of agreement (IOA). Additionally, the HAT exhibited 22.09% fewer errors and 82.59% lower bias compared to the 3-D CTM. Diurnal variations in PM2.5 predictions from both models were also analyzed to explore the characteristics of the proposed model further. The HAT predictions closely aligned with observed PM2.5 throughout the day, whereas the 3-D CTM exhibited significant diurnal variability. The importance of the input features was evaluated using the permutation method, which revealed that the previous day's PM2.5 was the most influential feature. The robustness of the HAT was further validated through a comparison with the long short-term memory (LSTM) model, which showed 18.50% lower errors and 95.91% smaller biases, even during El Ni & ntilde;o events. These promising findings highlight the significant potential of the HAT as a cost-effective and highly accurate tool for air quality prediction.