Evaluation of Support Vector Machine Performance for Carbon Monoxide Prediction

In Carbon monoxide (CO) is one of the main air pollutant parameters in the atmosphere of Tehran, Iran. Generally, it is difficult to predict and control CO concentration because it is essentially nonlinear time-varying system. Recently, in particular, environmental control such as CO concentration level control is regarded as one of the most important factors in environmental protections. This paper describes forecasting and more specifically uncertainty determination of CO concentration during the modeling process using a support vector machine (SVM) technique. Uncertainty of the air pollution modeling studies highly affected the simulation results. In this regards, it is very important to determine the uncertainty of air pollution models due to consequences on health of people exposed to the pollution. Therefore, this research aims to calibrate, verify, and also determine the uncertainty of support vector machine (SVM) in the process of air pollution modeling in the atmosphere of Tehran. To achive this goal, the SVM model was selected to predict arithmetic average of daily measured CO concentration in the atmosphere of Tehran. In this regards, the SVM model was calibrated and verified using six daily air pollutants include particulate matter with diameter equal or less than 10 micrometer (PM10), total hydrocarbons (THC), nitrogen oxides (NOx), methane (CH4), sulfur dioxide (SO2) and ozone (O3) and also six daily meteorological variables include pressure (Press), temperature (Temp), wind direction (WD), wind speed (WS) and relative humidity (Hum). The data was collected from Gholhak station located in the north of Tehran, Iran, during 2004-2005. Thereafter, the best developed SVM model for predicting the CO concentration was chosen based on determination of coefficient (R2). Finally, to determine the SVM uncertainty, the model was run many times with different calibration data. It led to many different results because of the model sensitivity to the selected calibration data. Then, the model uncertainty in the CO prediction process was evaluated using the width of uncertainty band (d-factor) and the percentage of measured data bracketed by the 95 percent prediction uncertainties (95PPU). Generally, the results confirmed the strong performance of the SVM model in predicting CO concentration in the atmosphere of Tehran. The predicted average daily CO concentrations by SVM model had a good agreement with the measured ones in the Gholahak air quality monitoring station. It was found that the determination of coefficient for calibration and validation of SVM model were equal to 0.89 and 0.88, respectively. Furthermore, the results indicated that the SVM model has an acceptable level of uncertainty in prediction of CO concentration in which the level of d-factor and the percentage of measured data bracketed by the 95PPU in the validation step were 0.74 and 76, respectively. Therefore, The obtained results indicated that the SVM model had an acceptable level of uncertainty in prediction of CO concentration. Therefore, it can be concluded that the SVM model is able to predict the CO concentration in the atmosphere of Tehran while it resulted an acceptable level of uncertainty. Finally, due to the proposed methodology is general, the authors suggest to apply it for analyzing the uncertainty of SVM model in other fields of science and engineering.