以交通狀態為基礎之遺漏值補正策略

A Traffic State-based Missing Data Imputation Strategy

交控中心通常藉由線圈偵測器或影像偵測器蒐集交通流量資料與交通狀況，以對旅行時間進行預測與推估；然而，任何一個即時的交通資料預測系統在實際運作時，遺漏值的處理是無可避免，當面臨遺漏值現象時，過去的補值策略往往並未詳細考慮車流續進與延滯之特性，僅以偵測器本身的歷史均值或是以移動平均方式填補遺漏值。為了考慮車流續進的過程以及彰顯路段相鄰偵測器間之交通狀態關係，本文除了採用基本歷史均值與移動平均模式進行遺漏值填補方法外，更引進資料採集分析技術設計出一套能考量相鄰偵測器之交通狀態之混合補值模式。首先，以集群分析方法對每個線圈偵測器之歷史資料作交通狀態分類處理，根據這些偵測器所代表之次路段交通狀態再結合整體路段之ETC旅行時間構建整體迴歸關係；接著以CART演算法構建各偵測點與其相鄰偵測器及路段ETC旅行時間所關聯之分類決策樹；最後，當某偵測點發生遺漏值時，則以該點對應之CART決策樹作為補值之預測依據。經過以實際有效樣本資料驗證結果顯示，透過交通狀態分類後之CART演算法可以有效提供長時窗遺漏值情況下的補值作業；另外，本文也發現在不同遺漏時窗數情境下，應以不同的補值策略進行補值，才能符合即時多變的偵測器遺漏值補正之需。

While loop or image detectors have been frequently adopted to collect traffic flow data as a basis for predicting and estimating travel time, missing values is an inevitable issue in real operations. Mean and moving average values based on historical data are common choices to replace missing values in past studies, which do not consider the features of vehicle flow continuation and lagging. To resolve this issue, this study proposes a novel approach which is based on data mining technique by combining the traffic information of traffic detector itself and its adjacent detectors. First, a regression model representing all road sections was developed based on the original historical traffic data of each loop detector. A decision tree was then established using Classification And Regression Tree (CART) to connect each detection point to the adjacent detectors and the Electronic Toll Collection (ETC) travel time on the associated road section. Finally, missing data were imputed based on the developed CART model. The empirical study showed that the CART imputation method based on traffic state works effectively to impute data with missing values, especially under the circumstance of long-period data missing. Moreover, it was found that under circumstances with different number of missing time-windows, hybrid imputation strategies fit better in meeting varying real-time needs.