Estimating PM_2.5 utilizing multiple linear regression and ANN techniques.

Area of Science:

• Environmental Science
• Data Science
• Atmospheric Chemistry

Background:

• Accurate prediction of air pollutants, especially Particulate Matter (PM), is essential for effective air quality management.
• Identifying and combining the most relevant input variables is crucial for dependable PM prediction models.

Purpose of the Study:

• To utilize correlation coefficients for selecting pertinent input and output variables for an air pollution model.
• To estimate PM2.5 concentration using Artificial Neural Networks (ANNs) and various input variables.

Main Methods:

• Employed correlation coefficients to select key input variables (sulfur dioxide, nitrogen dioxide, PM10) and output variable (PM2.5).
• Developed and compared three Artificial Neural Network (ANN) models: Levenberg-Marquardt (LM-ANN), Bayesian Regularization (BR-ANN), and Scaled Conjugate Gradient (SCG-ANN).
• Evaluated model performance against the Multiple Linear Regression (MLR) method.

Main Results:

• The LM-ANN model demonstrated superior performance compared to BR-ANN and SCG-ANN.
• The LM-ANN model outperformed the Multiple Linear Regression method.
• LM-ANN achieved a high R² value of 0.8164 and a low Root Mean Square Error (RMSE) of 9.5223.

Conclusions:

• The Levenberg-Marquardt algorithm-trained ANN is an effective method for predicting PM2.5 concentrations.
• The study successfully identified key variables and a robust model for air pollution forecasting.
• The findings support the use of ANNs for improving air quality management strategies.