The General Autoregressive Conditional Heteroskedastic (GARCH) model and 10 ordinary air quality monitoring stations in the entire air quality control district in Kaohsiung-Pingtung were used in this study. First, the factor analysis results within multivariate statistics were employed to select the main factor that affects air pollution, namely, the photochemical pollution factor. The characteristics of the GARCH model were discussed in terms of asymmetric volatility among the three air pollutants (PM10, NO2, and O3) within the factor. In addition, this study also combined the multiple time series model VARMA to explore changes in the time series of the three air pollutants and to discuss their predictability.The results showed that, although the coefficient of the GARCH model was negative when estimating the variance equation, the conditional variance would always be positive after taking the logarithm. The results also suggested that the GARCH model was quite capable of capturing the asymmetric volatility. In other words, if the condition that pollution factors might be subject to seasonal changes or outliers generated by the human contamination is not considered, the GARCH model had very good ability to verify the results and make predictions, regardless of whether it adopted any of the three risk concepts: normal distribution, t-distribution, and generalized error distribution. For example, under the trend of time series temporal and spatial distribution in various pollution concentrations of photochemical factors, the optimal model VARMA(2,0,0)-GARCH(1,1) selected in this study was used to conduct time series predictability after the verification procedure. After capturing the last 50 entries of data on O3 concentrations in the sequence, the results showed that the predictability correlation (r) was 0.812, the predictability of NO2 was 0.783 and the predictability of PM10 was 0.759. It can be learned from the results that under the sequence of the GARCH model with strong asymmetric volatility, the residual values of these three sequences as white noise were quite evident, and there was also a high degree of correlation in predictability.