Establishment of a Box-Jenkins Multivariate Time-Series Model to Simulate Ground-Level Peak Daily One-Hour Ozone Concentrations at Ta-Liao in Taiwan

Abstract

Box-Jenkins univariate autoregressive integrated moving average (ARIMA) and regression with time-series error (RTSE) models were established to simulate historical peak daily 1-hr ozone concentrations at Ta-Liao, Taiwan, 1997– 2001. During 1995–2003, the 600 days of Pollution Standard Index (PSI) more than 100 (peak daily 1-hr ozone concentrations detected by greater than 120 ppm) at Tao-Liao showed the highest ozone exceedances among the six monitoring stations in Kaohsiung County. To improve the predictability of extremely high ozone, two different principal components, PC1 and PC(1 + 2), were introduced in the RTSE model. Four typical predictors (particular matter with an aerodynamic diameter less than or equal to 10 μm, temperature, wind speed, and wind direction) plus a PC trigger remained significant in the RTSE model. The model performance statistics concluded that the RTSE model with PC1 was optimal, compared with the univariate ARIMA, the RTSE model without PC, and RTSE model with PC(1 + 2). The contingency table shows that the successful predictions of the univariate model were only 12.9% of that of the RTSE model with PC1. Also, the POD value was improved approximately 5-fold when the univariate model was replaced by the RTSE model, and almost 8-fold when it was replaced by the RTSE model with PC1. Moreover, introducing the PC trigger indeed enhanced the ozone predictability. After the PC trigger was introduced in the RTSE model, the POD was increased 69.9%, and the FAR was reduced 8.3%. The overall correlation between the observed and simulated ozone was improved 9.6%. Also, the first principal component was more useful than the first two components in playing the “trigger” role, though it counted only for<br/>58.62% of the environmental variance during the high ozone days.