Improving model prediction of PM_2.5 and PM₁₀ using combination of features selection and machine learning approach

This study was carried out to develop multi-input-single-output (MISO) models using an artificial neural network to predict the concentrations of PM_2.5 and PM₁₀ respectively based on meteorological parameters. Data pre-processing using variable importance in projection (VIP) scores managed to select significant input toward output for model development. Based on the feature selection, model development was built with and without input selection using a nonlinear autoregressive network with exogenous inputs (NARX) neural network model which made used of 2 number of delays, implementing Levenberg-Marquardt as training algorithm. The performance of the prediction model was evaluated by measuring mean sum square error (MSE) and coefficient of determination (R²) values. Model prediction of PM_2.5 and PM₁₀ concentration using machine learning is achieved and useful not only to improve public awareness but the air quality management in Malaysia as well. The model prediction using features selection is comparable with using all input variables in the model prediction of PM₁₀ and PM_2.5 with the range of coefficient determination, R² within 0.8700 and 0.9100 and mean square error (MSE) of 0.28 to 0.33 respectively.