|Abstract: ||高屏地區之光化學污染問題值得長期關注，如何預測及降低臭氧濃度為目前空氣品質污染防制的重要課題。本研究以高雄市臭氧污染較嚴重之左營地區為例，分別採用判別分析(discriminant analysis, DA)、邏輯迴歸(logistic regression, LR)、類神經網路(artificial neural network, ANN)，以及較新之支持向量機(support vector machine, SVM)等四種模式作臭氧事件日之預測分析，各模式之預測成果雖互有差異，但相較於直接預測臭氧濃度，本研究展示了探討臭氧問題之另一可行途徑。|
It’s a problem worthy of much attention in the long term for the photochemical air pollution in Kaohsiung/Ping-Dong area. How to predict and reduce the ozone concentration is an important issue for the air quality pollution control. In this research, a case study of ozone pollution is performed for the Zuoying area in Kaohsiung. In order to predict the ozone episode days, four models, i.e., discriminant analysis (DA), logistic regression (LR), artificial neural network (ANN), and the support vector machine (SVM), are used to analysis and comparison, respectively. Rather than predicting the ozone concentration directly, the occurrence of ozone episode days is forecast instead in the analysis which demonstrates another way to explore the ozone issue.
In this study, five indicators including probability of detection (POD), false alarm rate (FAR), false positive rate (FPR), false negative rate (FNR) and accuracy (ACC) are used as a basis to assess the model performance. If based on the biased data according to the U.S. standard for ozone episodes, DA has the highest POD, but with the worst FAR and FPR, led to the poor performance of the prediction results. Whereas SVM has the better FAR, FPR, and the ACC, the performance of ANN prediction is the best overall. On the other hand, if based on the unbiased data according to the Taiwan standard for ozone episodes, the performance of DA and LR models is not satisfactory. ANN has the highest ACC and the lowest FPR and FAR. In contrast, although SVM has the higher FAR, it with the highest POD and the lowest FNR. The performance of SVM is not good than ANN in general, however, its POD and FNR values show good. If the influences of ozone episodes on the public health are considered, it will be a not bad choice by the SVM model to predict the ozone episode days.