發展資料探勘為基礎之高速公路長距離且含非重現性旅行時間預測模式

Abstract

由於公路面積擴增之幅度遠不及於旅運需求之成長速度，因此，先進交通管理系統(Advanced Traffic Management System, ATMS)與先進旅行者資訊系統(Advance Traveler Information System, ATIS) 之規劃與興建，即為今後提升運輸系統效率並滿足旅運需求之重要手段。並且，唯有長距離旅行時間 預測，才能有效提供高速公路替代路徑資訊，進而發揮 ATMS 與 ATIS 之功能。再者，依據Oak Ridge National Laboratory (Chin et al., 2004)之研究指出，駕駛人於美國高速公路上所面臨的延滯(delay)，有 55%是非重現性擁擠所造成的，其中高速公路事故佔72% (Skabardonis et al., 2003)。由此可知，為提 供ㄧ穩健與連續型之預測模式，在面對長距離且非重現性擁擠影響下之複雜車流環境，如何尋找關鍵 影響變數以確實反映車流特性，即為高速公路旅行時間預測研究中，重要且必須克服之議題。 以往研究為提升旅行時間模式之預測能力，於變數選取與預測方法之應用皆有充分探討。然而， 卻鮮少在長距離且含非重現性擁擠下，針對變數之編碼與組合方式及變數縮減等議題進行探討。變數 之編碼與組合方式確實影響預測模式之準確度。再者，以最精簡變數提供準確之預測模式，一直為相 關單位所重視之問題。有鑑於此，本三年期計畫首先第一年計畫將藉由倒傳遞類神經(Back-Propagation Neural Network, BPN)方法，針對重要變數之組合與編碼方式進行探索，透過此必要且繁瑣的探索過程 後，獲得解析國內車流之重要變數特性。第二年計畫即在第一年計畫之研究基礎上，以資料探勘技術 中分群、分類與混合專家(Mixture of Experts)之概念，提高對長距離與非重現性擁擠下車流變異情形之 掌握度。最後，第三年計畫以資料探勘之決策樹(decision tree)模式之分類邏輯與第二年期兩階段分群 結果，挑選重要變數並達到縮減維度之目的，以期提供ㄧ穩健與高準確度之連續型預測模式。

Because the increase of roads can not meet the growth of transportation requirement, the planning and construction of Advanced Traffic Management System (ATMS) and Advance Traveler Information System (ATIS) have become the essential approaches for improving the efficiency of transportation systems as well as satisfying the transportation requirement. The long distance travel time prediction can effectively provide freeway alternative route information to facilitate ATMS and ATIS more successful. The study of Oak Ridge National Laboratory (Chin et al., 2004) pointed out that, 55% of the delays drivers encountering in American freeways are caused by non-recurrent events, 72% of which are freeway accidents (Skabardonis et al., 2003). Therefore, for providing a robust and continuous prediction model of travel time in the complex traffic flow with long distance and non-recurrent congestion, finding out the critical factors reflecting the characteristics of traffic flow is an important issue and needing to be addressed in the research area of freeway travel time prediction. For improving the performance of travel time prediction models, previous studies have substantially investigated the variable selection and applications of prediction models. However, relatively few studies have addressed the issues of variable encoding, variable combination and dimension reduction in the area of travel time prediction with long distance and non-recurrent congestion. Variable encoding and variable combination actually have an impact on the accuracy of prediction models. In addition, providing accurate prediction models with a least amount of variables is an important issue to the related organizations. Therefore, in this three-year project, the first-year project will investigate the important variable combination and variable encoding by using the Back-Propagation Neural Network (BPN). Through this sophisticated and necessary investigation process, the characteristics of Taiwan’s traffic flow can be captured. Basing on the results of the first-year project, the second-year project will apply the data mining techniques of clustering, classification and the concept of Mixture of Experts to further capture the characteristics of traffic flow with long distance and non-recurrent congestion. Finally, in order to create a robust and highly accurate continuous prediction model, the third-year project will select the important variables for reducing the dimensionality by using the decision tree and the results of two-stage clustering obtained from the second-year project.