Predictive modeling of PM2.5 using soft computing techniques: case study—Faridabad, Haryana, India.

Particulate matter has a detrimental consequence on the health of living organisms throughout the world and predicting their concentration is very imperative to assess their impact on human health. Faridabad is the most populated and largest city of Haryana, India, and the current study was designed to foresee the PM_2.5 content by different modeling techniques: (1) support vector machine (SVM), (ii) random forest (RF), (iii) artificial neural network (ANN), (iv) M5P model, and (v) Gaussian process regression (GP). Collected data (659 observations) from May 2015 to May 2018 were used to develop the models. Parameters such as temperature (T), ground-level ozone (O₃), sulfur dioxide (SO2), nitrogen dioxide (NO2), nitric oxide (NO), NOx, wind speed (WS), wind direction (WD), relative humidity (RH), bar pressure (BP), and solar radiation (SR) are used as input parameters for prediction of PM_2.5. The results of all the models suggested that RF model with testing correlation coefficient (CC) 0.8312, mean absolute error (MAE) 30.7757, R² (correlation of determination) 0.6909, and root mean square error (RMSE) 44.6947 is the best estimator for appraisal of PM_2.5 followed by SVM, GP, M5P, and ANN models. The sensitivity analysis results indicated that wind speed is the utmost influencing parameter for the estimation of PM_2.5. Abilities of different models were compared and RF was established as the best technique based on assessment criteria. We recommend more studies employing RF and other techniques as hybrids that lead to better models. [ABSTRACT FROM AUTHOR]