本篇研究針對臺灣2017年1月份與2月份的PM2.5資料進行預測，以Huang et al.（2018）所提出的模型為基礎，該模型利用多解析度樣條基底函數（multiresolution spline basis functions）作為迴歸模型的解釋變數，並使用穩健迴歸（robust regression）來對參數進行估計，空間相依性則假設為指數（exponential）函數，最後透過簡單克利金法（simple kriging）做空間預測。本文於該模型中增加基底函數，並考慮新增兩個氣候因子，改以最小絕對值壓縮和篩選算法（least absolute shrinkage and selection operator, Lasso）來對參數進行估計，試圖比較出哪種模型的預測效果較佳。根據研究結果顯示，我們提出的模型於預測能力上並無明顯的差異，但模型增加兩個氣候因子有助於加強模型的預測能力。
In this study, we try to predict Taiwan PM2.5 data in January and February, 2017. We modify the model proposed by Huang et al. (2018). Their model used the multiresolution spline basis functions as the explanatory variables, and robust regression to estimate the parameters. The spatial dependence is assumed to be exponential. Then the spatial prediction is obtained by using the simple kriging method. For improving prediction accuracy, we consider increasing the number of the basis functions and adding two climate factors. Besides, the least absolute shrinkage and selection operator (Lasso) is used to estimate the parameters instead. We compare the prediction performances of few different models. The results show there is no significant differences between these models but adding two climate factors can improve prediction accuracy.
