Identification of best machine learning model for the real-time vehicular data based prediction of PM2.5 and PM10

Abstract

In fast-developing urban regions such as the Guwahati City, the particulate matter (PM₁₀ and PM_2.5) concentration prediction is vital to ascertain air quality and public health. Utilizing a large dataset that constitutes historical real-time pollution data, vehicular population count (petrol and diesel), and meteorological characteristics (temperature, wind direction, solar radiation, relative humidity, wind speed) data, the article applies alternate machine-learning algorithms for the prediction of PM_2.5 and PM₁₀ levels in the Guwahati city. The intricate temporal patterns and seasonality inclines of the air pollution data were captured with the alternate ML models namely Extreme Gradient Boosting, Decision Tree, Random Forest, Support Vector Regression, K nearest neighbour and Multilayer Perceptron. The models were assessed for their efficacy with important metrics such as the coefficient of determination, root mean square error and mean absolute error. The algorithmic performance based data analysis was undertaken to analyze upon the sensitive influence of lag features, rolling statistics, seasonal decomposition components, temporal features and seasonality-specific issues on the model performance. Accordingly, they highlight the efficacy of machine learning models for their ability and effectiveness to predict air quality parameters. The explorations convey that ensemble techniques such as the Extreme Gradient Boosting outperform other models in terms of the lowest RMSE values of 0.024 μg/m³ and 0.041 μg/m³ for PM_2.5 and PM₁₀ respectively; MAE values of 0.017 and 0.027 for PM_2.5 and PM₁₀ respectively and coefficient of determination values of 0.96 for PM_2.5 and values of 0.92 for PM₁₀. Accordingly, the conducted investigations can foster the implementation of pragmatic policies that are to be meticulously followed to safeguard the air quality of the city.