基於資料驅動的交通現象搜索與車流預測

Abstract

交通控制、運輸規劃與政策及路況服務是很重要的課題，交通管理者需要擬訂決策保持行車道路順暢，因此他們經常利用歷史資料搜索出相似交通現象來研究複雜的交通現象。我們使用台灣國道上的車輛偵測器資料來做為研究探討，但交通資料每分鐘便可取得一筆資料，隨著時間發展資料量越來越龐大，因此我們需要一個有效率的搜索方法。我們使用由粗至細的隨機搜索(CFRS)做為我們的搜索方法，幫助我們在龐大的歷史資料下快速搜尋相似交通現象。另外，由於搜索出來的是相似交通現象，未來的交通現象發展可能也會有相似的交通趨勢，所以我們還可以利用相似交通現象的歷史資料來預測未來車流發展，幫助交通管理者做出即時道路控制的決策。在最後的實驗結果裡，搜索功能部分我們探討了一些交通現象，並從其中發現了一些有趣的結果。在預測車流部分我們和k-nearest neighbors(KNN)方法做比較，我們的預測結果相較於KNN方法的誤差並沒有明顯提高，但搜索效能上可以快上約14倍到89倍。

Traffic control, transportation planning, and road service are important issues in transportation management. As traffic management officers need make good decisions to keep traffic flow smooth, they often gain experiences by studying similar traffic events in historical traffic data. Among different types of traffic data, vehicle detector (VD) data are commonly used by traffic management experts for analyzing traffic events in large scale and long term traffic database. As VD data are streaming data refreshed every minute and the number of VDs on highways is large, the historic VD data are huge. An efficient search approach is thus very critical for exploring VD databases. In this thesis, we propose a Coarse-to-Fine Random Search (CFRS) method to help users retrieve similar traffic patterns in massive historical traffic data. Our CFRS approach can also be applied to traffic prediction by searching a database using the current traffic as a query pattern. The traffic data following the retrieved pattern can be used as predicted traffic due to the continuity of traffic data. In our experiment, we test our approach on the VD data collected from nearly 3000 VDs on the nine freeways in Taiwan in two years. The experimental results show that our approach can use to retrieve traffic patterns efficiently. In addition, we compared our traffic prediction method with k-nearest neighbors (KNN). Our method achieves better performance which speeds up for 14 to 89 times while maintaining the same level of accuracy of KNN.