Predicting Long‐Term PM 2.5 Concentrations Through Hybrid Attention Mechanism and Embedding Method

Air pollution, particularly PM 2.5 pollution, remains a serious environmental and public health concern, and accurate long‐horizon forecasting of PM 2.5 concentrations is crucial for effective PM 2.5 emission control strategies and public health management. In this context, a forecasting model uses historical air pollutant observations and meteorological conditions to estimate future PM 2.5 concentrations. Deep learning methods have demonstrated their effectiveness in air quality prediction in recent years. However, long‐horizon PM 2.5 forecasting remains challenging because long‐range temporal information may attenuate or be lost during sequence compression, and multivariate air quality series are often affected by noise and missing values. To address this gap, we propose a deep learning model that incorporates a hybrid attention mechanism and embedding method to enhance the accuracy of PM 2.5 concentration prediction. Our model integrates various types of attention mechanisms into both input and output sequences to address the challenge of long‐term memory retention, enabling better capture of long‐range dependencies. Additionally, an embedding method is employed to process input features, allowing end‐to‐end training for each feature value and yielding higher dimensional representations that are more comprehensive and faithful. To validate the proposed model, we conducted a 72‐h multi‐step PM 2.5 concentration prediction experiment using the Beijing dataset with an optimal input window of 180 h. The final experimental results show that, compared with the best performing models among statistical methods and machine learning methods, respectively, our model achieves relative MAE reductions of 35.41% and 22.48% when forecasting 72 h ahead. Furthermore, our best performing model attains a relative MAE reduction of 9.21% versus Transformer when forecasting 72 h ahead. This study provides a valuable tool for reducing pollution exposure risks and offers scientific support for air quality management and public health protection.