Big data analytics and artificial intelligence in air pollution studies for the prediction of particulate matter concentration

Abstract

Statistical modeling has found not suitable to be used when predicting the particulate matter (PM10) as it is non-linear in nature. The complexity and nonlinearity of PM10 concentration in the atmosphere are known best captured by the nonlinear model which emerges nowadays such as Multi-Layer Perceptron Neural Network (MLP-NN). In order to assess the capability of MLP-NN model in predicting the PM10 concentration, a statistical or traditional model known as Multiple Linear Regression (MLR) was also developed as a reference model. The daily air quality data and meteorological variables from the year 2010-2014 were assembled in developing the models. The MLP-NN model with the combination of logsig and purelin activation function revealed 75.5% of the variance in data with 6.59 μg/m3 (RMSE) and 88.0% of the variance in data with 6.30 μg/m3 (RMSE), during training and testing phase, respectively. The MLP-NN model improves by 61.5% and reducing the 62.2% error as compared to the MLR model. This model is appropriate for operational used by respected authorities in managing air quality in maintaining sustainability and as an early warning during an unhealthy level of air quality.