Selection of the Most Significant Variables of Air Pollutants Using Sensitivity Analysis

Abstract

Abstract This study was conducted to determine the most significant parameters for the air-pollutant index (API) prediction in Malaysia using data covering a 7-year period (2006–2012) obtained from the Malaysian Department of Environment (DOE). The sensitivity analysis method coupled with the artificial neural network (ANN) was applied. Nine models (ANN-API-AP, ANN-API-LCO, ANN-API-LO3, ANN-API-LPM10, ANN-API-LSO2, ANN-API-LNO2, ANN-API-LCH4, ANN-API-LNmHC and ANN-API-LTHC) were carried out in the sensitivity analysis test. From the findings, PM10 and CO were identified as the most significant parameters in Malaysia. Three artificial neural network models (ANN-API-AP, ANN-API-LO, and ANN-API-DOE) were compared based on the performance criterion [R2, root-mean-square error (RMSE), and squared sum of all errors (SSE)] for the best prediction model selection. The ANN-API-AP, ANN-API-LO, and ANN-API-DOE models have R2 values of 0.733, 0.578, and 0.742, respectively; RMSE values of 8.689, 10.858, and 8.357, respectively; SSE values of 762,767.22, 1,191,280.60, and 705,600.05, respectively. The findings exhibit the ANN-API-LO model has a lower value in R2 and higher values in RMSE and SSE than others. ANN-API-LO model was considered as the best model of prediction because of fewer variables was utilized as input and far less complex than others. Hence, the use of fewer parameters of the API prediction has been highly practicable for air resource management because of its time and cost efficiency.