A comparative study of using Random Forests (RF), Extreme Learning Machine (ELM) and Deep Learning (DL) algorithms in modelling Roadside Particulate Matter (PM10 & PM2.5)

Abstract

The effects of traffic-derived air pollution can be controlled through the provision of adequate and effective air quality control and mitigation measures with the aid of air quality models. This paper examines the application of Machine Learning (ML) methods (Random Forests (RF), Extreme Learning Machine (ELM) and Deep Learning (DL)) algorithms in air quality modelling. Data collected from continuous monitoring stations in London comprising air pollutants, traffic and meteorological variables were used for training the models. The selected ML methods were trained to predict roadside PM10 and PM2.5 concentrations. The results obtained showed that all the methods can be used for training models for the prediction of the PM10 and PM2.5 concentrations. The RF, ELM and DL algorithms were found to be suitable for this purpose due to their predictive accuracy and faster training speed. The models performed slightly better in predicting PM2.5 than the PM10 concentrations. The average performance of the machine learning models was found to be 0.90 and 0.94 (R); 99% and 98% (FAC2); 9.2 and 4.5 (RMSE) and 0.83 and 0.84(IOA) for the prediction of PM10, PM2.5. The advantages of using the Deep Learning, ELM and RF algorithms over traditional ANN are the faster training speed and scalability especially when using high-performance computing.