Short-Term Prediction of Carbon Monoxide Concentration Using Artificial Neural Network (NARX) Without Traffic Data: Case Study: Shiraz City

Abstract

Air pollution is one of the most widespread and important issues in mechanical civilization, and it has made supervision and control of air quality an ineluctable issue that has been introduced as a principal national problem. This study investigates the ability of dynamic neural networks, particularly the nonlinear autoregressive exogenous (NARX) network, in predicting air carbon monoxide concentration in Shiraz in the absence of traffic data since there are no accurate statistical data on traffic volume (as one of the primary sources for air pollution modeling). Dynamic networks have been utilized to model time-variable patterns as they have time memory through the history of concentration volume implicitly containing traffic characteristics. To begin this study, meteorological data including temperature, moisture content, rainfall amount, and wind velocity and direction at a 3-h mean basis were obtained from the Bureau of Meteorology at the Shiraz Airport. Moreover, air pollutant concentration data due to Setad Square’s measurement station between 2005 and 2008 were prepared from the Fars Department of Environmental Protection. According to the results obtained from the static neural network, the correlation coefficients (R) for the training, validation, and test datasets are estimated as 0.49, 0.37, and 0.41, respectively. Moreover, the R2 correlation coefficient, the root mean square error (RMSE), and mean absolute percentage error (MAPE) are 0.31, 0.43, and 51%, respectively. However, the correlation coefficients achieved from NARX model for the training, validation, and test datasets are estimated as 0.77, 0.76, and 0.80, respectively, while the R2 correlation coefficient, RMSE, and MAPE are 0.72, 0.05, and 7%, respectively. The results demonstrate the dynamic neural network’s high performance in modeling carbon monoxide concentration in the absence of traffic data. Moreover, sensitivity analysis indicates the stability of the model to the noisy data.