A comparative analysis of Statistical and Computational Intelligence methodologies for the prediction of traffic-induced fine particulate matter and NO2

Abstract

With the urbanization increase, urban mobility and transportation induce higher traffic volumes causing environmental, economic and social impacts. This is due to continuous usage of fossil fuel energy resources generating air pollutants, such as nitrogen oxides (NOx), sulfur dioxide (SO2), carbon monoxide (CO), ozone (O3) and particulate matter (PM10 and PM2.5), which impact on climate and air quality and adversely affect the human health. The present paper aims at training an ensemble of forecasting methodologies for traffic-induced pollutant emissions and implementing it for predicting PM10, PM2.5 and NO2 for the case study of Cambridge, UK inner-city region. Such an ensemble enables decision makers to evaluate the impact of various transportation policies and measures on human health and the ecosystem, and subsequently contribute towards urban resilience and sustainability. Since the chemical synthesis of air pollution is triggered by meteorological factors, the forecasting incorporates them along with the traffic volumes. We opted to combine Statistical and Computational Intelligence learning methods including Adaptive Neuro Fuzzy Inference Systems (ANFIS), Long Short-Term Memory (LSTM) recurrent neural networks and Extreme Learning Machines (ELM). Initially, Multivariate Imputation by Chained Equation (MICE) and trend and seasonality removal was performed at data preprocessing and then Principal Component Analysis (PCA) highlighted the principal parameters for ANFIS to predict next day's PM10, PM2.5 and NO2 values. LSTM and ELM methods estimated next day values and compared with the ANFIS model results for hourly time series data of length 2703. The performance of the embedded models was quantified by the Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Mean Square Error (MSE), and Coefficient of Determination (R2) indices. The ensemble was found to be superior in predicting PM10, PM2.5 and NO2 emissions when compared with existing traditional models. © Elsevier Ltd