以深度學習方法評估國道交管措施時空影響

Abstract

國道為聯絡縣市間重要之公路系統，其中國道1號、國道3號、國道5號等縱向國道尤其為南北往來的重要公路。連假期間大量車流湧上國道，造成國道多路段、長時間的壅塞，高速公路局於國道1號、國道3號及國道5號實施多項交通管理措施，包含入口匝道封閉、入口高乘載管制及收費措施調整等措施。國道交管措施會產生國道上多路段長時間的車流影響，然而其帶來的影響難以評估，主要係因為國道交管措施實施的時間、空間、形式皆不相同，對於用路人產生的效果亦不相同。 本研究提出以深度學習方法建構之圖片基礎速度預測模型，模型根據所欲預測日期本身的屬性、預定交管措施以及百萬延車公里（Million Vehicle-Kilometers, MVK） 便能夠預測出該日的時空圖，實際以真實資料訓練，並且針對驗證資料進行預測後，各模型絕對百分誤差（Mean Absolute Perntage Error, MAPE） 介於4~7%，顯示出模型優良的預測能力。此外，為了實際應用於交管措施評估，本研究提出擾動基礎的模型解釋方法，透過改變輸入觀察輸出的方式檢視不同交管措施帶來的影響，結果顯示多項交管措施符合預期的變化，模型能夠合理學習並輸出交管措施帶來的影響，可作為管理機關用以評估交管措施之工具。

National freeways are important highway systems connecting counties and cities, and the vertical freeways such as Freeway 1, Freeway 3, and Freeway 5 are especially important national freeways for north-south travel. During the vacation, lots of vehicles poured onto the national freeways, causing long-term congestion in many sections of the national freeways. The freeway bureau implemented several traffic control measures on Freeway 1, Freeway 3, and Freeway 5, including closing entrance ramps, HOV control of entrance ramps, and adjusting toll collection measures. Traffic control measures on national freeways will have a long-term impact on traffic flow in multiple sections of national freeways, but the impact is difficult to assess, mainly because the time, space and form of implementation of traffic control measures on national freeways are all different, and the effects on road users are also different. This study proposes an image-based speed prediction model using a deep learning approach. The model can predict the time-space diagram of the day based on the properties of the predicted date itself, the predetermined traffic control measures, and the Million Vehicle-Kilometers（MVK）. After training with real data and predictions on validation data, the Mean Absolute Percentage Error（MAPE）of each model is between 4% and 7%, showing the excellent prediction ability of the model. In addition, in order to apply the assessment of traffic control measures, this study proposes a perturbation-based model explanation method, which examines the impact of different traffic control measures by changing the input and observing the output. The results show that multiple traffic control measures are in line with the expected changes, and the model can reasonably learn and output the impact of traffic control measures. This model can be a supporting tool for the authority to evaluate traffic control measures.