應用格位傳送模式建構預測型高速公路動態起迄矩陣推估演算法

Abstract

正確和有效率的動態起迄矩陣推估對運輸規劃、交通管理及交通策略研擬是很重要的資訊，近年來隨著智慧型運輸系統之快速發展，動態起迄矩陣之推估及即時性交通控制之執行如路徑導引及號誌控制等績效均有顯著提昇。過去許多文獻利用所偵測到的高速公路主線與上、下匝道流量，進行高速公路動態起訖矩陣之推估，或以增加額外的假設或是外生的資訊，如路徑選擇行為或已知的依時性交通流量資料等提昇績效，然而，在此課題中，所面臨的挑戰是欲推估之參數數量遠大於資料所提供之訊息及如何處理旅行時間變異影響之問題。 基此，本研究提出一個預測型高速公路動態起迄矩陣推估方法，包括一個二階段交通量預測模式及一個整合型動態起迄矩陣推估演算法，交通量預測模式係以滾動式自我建構交通模型來預測中、長期交通特性，其中，包含以成長型階層式自我組織放射圖(GHSOM)將所有交通量樣本分成若干群，並針對每一群以基因規劃法(GP)建構相對應之非線性交通量預測模式去預測交通特性。而整合型動態起迄矩陣推估演算法係為了能在不同的交通情況下，有效率且準確的獲取車輛到達型態，藉由結合格位傳送模式(Cell Transmission Model，CTM)及卡門濾波 (Extended Kalman Filtering，EKF)，來建構遞迴式動態起迄矩陣推估演算法，由CTM模擬車輛運行行為，預測各依時起迄對之到達率，再以EKF推估動態起迄矩陣。 為驗證此演算法之績效及實用性，本文以所設計之6個起迄對之小型路網推估90分鐘的起迄矩陣作為驗證範例。同時為比較本模式於旅行時間預測之績效，以Greenshields巨觀模式假設在進入路網之車輛會於兩時階範圍內到達迄點之條件下，預測車輛旅行時間。結果顯示本模式推估結果之RMSE為0.069遠較Greenshields之0.145為優。在實例應用上，本文先以國道1號泰山收費站至楊梅交流道計6個交流道36公里長之3車道高速公路中型路網進行實驗驗證，結果顯示格位傳送模式在自由流至擁擠流等不同情境交通狀況下，有效模擬車輛到達率，而本模式並以相當低的RMSE績效，精確推估動態起迄矩陣。 最後，本研究再以國道1號頭份交流道至北斗交流道計15個交流道長110公里之3車道高速公路大型路網作為實例驗證，結果也證明本文所提二階段交通量預測模式的績效及整合型動態起迄矩陣推估演算法的實用性。另外，本文亦比較所提二階段交通量預測模式之績效優於傳統ARIMA模式。並進一步針對交通量樣本長度進行敏感度分析，發現5分鐘為一時階之交通量樣本長度在120個時階（10小時），即能達到相當之預測績效。

Accurate and effective dynamic origin-destination (O-D) matrices estimation is important for transport planning, traffic management and strategic planning. Recently, the rapid development of intelligent transport systems has enhanced accurate dynamic O-D information and the implementation of real-time traffic control, such as real-time route guidance and signal control. Numerous studies have devoted to developing estimation algorithms based on observable mainline and ramp flow rates, with constraints dependent on time series traffic flow and extra system equations and using recursive or non-recursive system solution techniques to estimate O-D matrices. However, this dynamic O-D matrices estimation issue remains challenging in that the number of parameters to be estimated is always far greater than the available information, and the impact of travel time variability on the time-varying O-D matrices. In light of this, the study proposes a novel approach to estimate medium-to long- term freeway dynamic O-D matrices. The proposed approach includes a two-stage prediction model with an integrated algorithm. The traffic prediction model predicts medium- to long-term traffic features based on rolling self-structured traffic patterns. The rationales include using the growing hierarchical self-organizing map model (GHSOM) to partition unlabeled traffic patterns into clusters and then developing an associated genetic programming (GP) model to predict the traffic features in each cluster. And then, the integration algorithm, which combined cell transmission model (CTM) with extended Kalman filtering (EKF) to respectively and iteratively estimate the arrival distributions and O-D proportions. To demonstrate the performance and applicability of the proposed approach, a seminal example with six O-D pairs of 90 minutes estimation is designed. The performance of this mode in terms of travel time prediction is compared to the Greenshields macroscopic model prediction. The results showed that the propose approach is better than the Greenshields model. In the field study, a medium-scale networks and a large-scale network of on-ramp traffic patterns on a freeway are examined. The medium-scale network covers a section of Taiwan No.1 Freeway (Taishan toll station to Yangmei toll station), a 36 km three-lane freeway section with 6 interchanges, and the results showed that the CTM can accurately capture the degree of traffic dispersion under traffic scenarios ranging from free-flow to congested-flow conditions and that the proposed EKF algorithm can accurately estimate the O-D proportions with rather low RMSE. For the large-scale network, 15 interchanges from Toufen interchange to Beidou interchange, a 110-kilometer stretch of Taiwan No.1 Freeway, were tested and the results indicated the practical applicability of the proposed algorithm. In addition, the proposed method has performed much better than the conventional ARIMA model. The sensitive analysis has also revealed the necessity of acquiring five-minute traffic patterns longer than 120 time intervals (10 hours) in order to achieve sufficient high prediction accuracy.