Data-driven predictive modeling of PM_2.5 concentrations using machine learning and deep learning techniques: a case...

Area of Science:

• Environmental Science
• Data Science
• Atmospheric Chemistry

Background:

• Air pollution, particularly PM2.5, poses significant health risks globally.
• Accurate forecasting of PM2.5 concentrations is crucial for public health advisement and policy-making.
• Existing models often lack the granularity or accuracy needed for effective real-time prediction.

Purpose of the Study:

• To forecast PM2.5 concentration in Delhi using advanced machine learning (ML) and deep learning (DL) models.
• To evaluate the performance of various ML/DL models including MLFFNN, SVM, RF, and LSTM.
• To introduce the aerodynamic roughness coefficient as a novel input parameter for PM2.5 prediction.

Main Methods:

• Applied MLFFNN, SVM, RF, and LSTM models for PM2.5 forecasting.
• Utilized air pollutant data (CO, Ozone, PM10, NO, NO2, NOx, NH3, SO2, benzene, toluene) and meteorological parameters as inputs.
• Incorporated the aerodynamic roughness coefficient as a novel input feature.

Main Results:

• The LSTM model demonstrated superior performance with an Index of Agreement (IA) of 0.986, RMSE of 21.510, NSE of 0.945, R² of 0.945, and R of 0.972.
• LSTM outperformed MLFFNN, SVM, and RF models in PM2.5 prediction accuracy.
• Sensitivity analysis identified PM10, wind speed, NH3, and benzene as the most influential parameters for PM2.5 estimation.

Conclusions:

• The LSTM model is highly effective for PM2.5 forecasting in Delhi.
• The inclusion of the aerodynamic roughness coefficient shows promise for improving prediction accuracy.
• The findings support the development of advanced, fine-scale air pollution forecasting systems using DL techniques.