利用手機信令推估旅運起迄矩陣

Abstract

傳統在都市中進行運輸規劃時，主要透過大量家訪問卷調查蒐集家戶成員旅運資料，做為推估模式的資料來源。然透過家訪問卷蒐集資料成本高昂、曠日廢時且更新頻率緩慢。近年來隨著行動裝置及基地台建置日益普及，透過手機信令資料所提供的使用者時空訊息開發運輸規劃模型具有經濟效益，相較於傳統資料蒐集而言也帶來許多便利。與傳統家訪問卷相比，手機基地台數據具有高覆蓋率、更新頻率迅速、效率高以及幾乎零成本的優點。雖然有諸多優勢，但訊號資料中缺乏一些關鍵訊息，如人口分布狀況、旅次資訊(旅次目的、運具選擇)。為了透過手機信令資料發展都市交通模型，許多的模型及相關演算法需要進一步著手開發。此外，要發展都市交通模型，值得關注的因素在於，由於手機信令資料僅佔道路使用者的一部分，因此必須進一步探討擴增係數，才能透過加乘結果回推實際旅次起迄點分布矩陣。本研究即發展出一個擴增係數推估模式，期透過求解擴增係數，進一步探討影響擴增係數因素。 為測試模式的適用性，本研究對台北市進行實例應用。本研究設定同一起點交通分區(以里為單元)至其他迄點交通分區之手機信令起迄量均乘以同一套擴增係數。至於擴增係數是否因時而異，本研究設定三種型態：擴增係數不因時而異、因時而異、因晨昏峰而異。此外，本研究係以位於屏柵線上之車輛偵測器所蒐集之小時交通量資料為基礎進行擴增係數之推估。至於屏柵線之設定，本研究採取三種方式：(1)所有車輛偵測器、(2)4條屏柵線、(3)2條屏柵線。因此，依據擴增係數及屏柵線之設定方式共計有9種推估結果。 為驗證本模式推估之交通分區起迄交通量，本研究以臺北都會區整體運輸需求預測模式建立與應用調查(TRTSIV)所推估之交通分區起迄交通量做為比較基礎。推估結果顯示9種推估結果中，在晨峰方面，以擴增係數因時而異且設定2條屏柵線者表現最佳，其整體準確率(MAPE)值為117.2%；表現最差的為擴增係數因時而異且設定所有車輛偵測器，其整體準確率(MAPE)值為363.4%。而在昏峰方面，以擴增係數因時而異且設定2條屏柵線者表現最佳，其整體準確率(MAPE)值為124.6%；表現最差的為擴增係數不因時而異且設定所有車輛偵測器，其整體準確率(MAPE)值為407.8%。進一步分析擴增係數數值高低之影響因素則以不同產業及業人口數、19~64歲人口比例、65以上人口比例、家戶數及平均家戶收入(元)較為顯著。

The quality of traditional urban transportation planning models mainly rely on a tremendous household survey which collects detailed travel itinerary of each of family members. However household survey has long been criticized as high cost, time consuming and low updating frequency. With the rapid growing popularity of cellular phones and densely distributed base transceiver (BT) stations, to develop transportation planning models based on the position and track of cellular probe data become economically feasible and comparatively advantageous. Comparing to traditional household travel survey, cellular probe data have advantages of high coverage, high updating frequency, high efficient and nearly zero additional cost, but lack of some key information, such as the demographics and trip information (trip purpose, mode choice) of mobile users. To develop urban transportation models based on cellular probe data, several models and algorithms still have to be developed. Most importantly, since cellular probe vehicles only account for a part of road users, adjustment factors (cellular probe per vehicle equivalents) have to be determined so as to “scale-up” cellular probe origin-destination (OD) matrices to traffic OD matrices. To show the applicability of the proposed model, a case study of Taipei City has been conducted. This study presumes that each traffic analysis zone (TAZ) has its corresponding adjustment factor which is applied to all OD pairs originated from the TAZ. Three types time variant of adjustment factors are assumed: the adjustment factor remains unchanged across hours, the adjustment factor varies across hours, and the adjustment factor remains unchanged within morning or afternoon peak hours, but varies from morning to afternoon peak hours. Additionally, this study also sets three types of screen lines: to select all VD as screen points, to select those VD on four predetermined screen lines, and to select those VD on two predetermined screen lines. With three types of time variates and three types of screen line section, a total of nine models are estimated and compared. To verify the proposed model, this study also compared the estimated OD matrices with those estimated by traditional transportation planning of Taipei metropolitan (i.e., Transportation Planning for Taipei Metropolitan-IV, TRTS-IV). The results show that at morning peak, the best performing model the VD on two screen lines and the adjustment factor varies across hours with an average MAPE (mean absolute percentage error) of 117.2%. The worst performing model is to select all VD as screen points and the adjustment factor varies across hours with an average MAPE of 363.4%. As to afternoon peak hours, the best performing model is to select VD on two screen lines and the adjustment factor varies across hours with an average MAPE of 124.6%. The worst performing model is to select all VD as screen points and the adjustment factor varies across hours with an average MAPE of 407.8%. Furthermore, this study further investigates the key factors affecting adjustment factors. The results show that the significant variables are employment of different industries, 19-64 years old population ratio, over 65 years old population ratio, number of households, and average household income.