Forecasting Daily Ambient PM_2.5 Concentrations in Qingdao City Using Deep Learning and Hybrid Interpretable Models...

Area of Science:

• Environmental Science
• Data Science
• Public Health

Background:

• Urbanization in China exacerbates air pollution, particularly fine particulate matter (PM2.5), posing significant public health risks.
• Accurate forecasting of PM2.5 concentrations is crucial for implementing effective mitigation strategies and protecting human health.

Purpose of the Study:

• To develop and evaluate advanced deep learning models for precise daily forecasting of PM2.5 concentrations.
• To identify key meteorological and atmospheric factors influencing PM2.5 levels in Qingdao City.

Main Methods:

• Utilized a dataset of meteorological factors and atmospheric pollutants in Qingdao City.
• Compared the predictive performance of various deep learning models: RNN, ANN, CNN, BiLSTM, Transformer, and a novel hybrid CNN-BiLSTM-Transformer architecture.
• Employed Shapley Additive Explanations (SHAP) to determine the influence of different factors on PM2.5.

Main Results:

• The hybrid CNN-BiLSTM-Transformer model demonstrated superior prediction accuracy, achieving a low RMSE of 5.4236 μg/m³, MAE of 4.0220 μg/m³, MAPE of 22.7791%, and a high R of 0.9743.
• SHAP analysis identified PM10, CO, mean atmospheric temperature, O3, and SO2 as critical factors influencing PM2.5 concentrations.
• The hybrid model effectively extracts local features, captures temporal dependencies, and enhances global pattern recognition.

Conclusions:

• The hybrid CNN-BiLSTM-Transformer model offers a robust and interpretable approach for PM2.5 forecasting.
• Findings provide valuable insights for public health management and environmental policy-making in response to air pollution.
• This study advances the multidimensional prediction of air pollution using integrated deep learning techniques.