請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/90810
標題: | 國道隨機事件影響時間推估模式之研究 Exploring Impact Patterns of Freeway Incidents: Impact Duration Modeling and Estimation |
作者: | 蕭仲綱 Chung-Kang Hsiao |
指導教授: | 許聿廷 Yu-Ting Hsu |
關鍵字: | 交通管理,交通事件影響時間,機器學習, Traffic duration analysis,Tobit model,Machine learning,Random forest, |
出版年 : | 2023 |
學位: | 碩士 |
摘要: | 隨著私有運具逐年成長的趨勢下,交通壅塞便成為許多城際公路面臨的重大議題。就交通壅塞而言,可以區分為重現性壅塞以及非重現性壅塞,重現性壅塞多發生於尖峰時段,發生的地點及時間較易預期;而非重現性壅塞發生的地點與時間呈隨機分布,較難進行預測。非重現性壅塞的發生原因通常來自交通事故、散落物、動物闖入等事件,根據事件發生的空間及時間異質性,使每起國道事件對於車流的影響時間推估變得十分複雜。在過往的文獻中,常將每起事件拆分成許多時間段,建立回歸模型預測個別時間段長度,並加總為整起事件的影響時間;但就實務上而言,個別時間段的長度會因調度班執行勤務的影響,而與原先時間段的定義有所差距。本研究結合國道事故班的調度資料和即時車流偵測器資料,以速度隨時間變化的熱力圖定義每起國道事件對車流的影響時間,顯示事件影響時間多數集中在10分鐘以內。根據事故班調度資料,其中又以交通事故及散落物為發生頻率最高的國道事件,因此本研究根據這兩種事件類別個別建立模型對其影響時間進行推估。由於影響時間為非負實數,因此以Tobit模型對個別國道進行估計,Tobit模型結果顯示顯著的變數多與即時車流資訊有關,而就平均絕對誤差及均方根誤差而言,散落物類模型的誤差皆較事故類模型小。然而,儘管Tobit模型得以推估影響時間及捕捉顯著的變數,但模型的適配指標較低,因此本研究將影響時間切割為離散變數,以機器學習的分類演算法架構預測影響時間。在隨機森林分類模型的架構中,整體模型的準確度會受到高度右傾的資料分布影響,因此整體準確度可達九成以上;然而就樣本數較少的類別來看,各個類別的準確度亦可達近八成。整體而言,本研究提出的模型可捕捉關於國道影響時間的顯著變數,機器學習分類模型架構亦在準確度足以信賴,期望能提供實務應用上有效率的國道隨機事件影響時間推估模型,能夠根據給定的事件條件預測影響延時,得以提供予交通管理單位及國道使用者更多資訊掌握目前的國道狀況。 In light of the significant increase in private vehicles, traffic congestion has become a critical issue in numerous cities as well as inter-city travel. Traffic congestion can be categorized into two types: recurrent congestion and non-recurrent congestion. Non-recurrent congestion arises from stochastic traffic incidents, which makes it challenging to predict the occurrences and locations of such incidents. Additionally, estimating the duration of traffic incidents is a complex task due to their temporal and spatial heterogeneity. In past research, it would divide the total incident duration into several periods and build regression models to estimate individual periods. The summation of each period equals the total incident duration. However, the length of separate periods would differ from the original period's definition due to the influence of the emergency response team's duties. This study applies heatmaps to analyze the impact duration of traffic incidents using real-time VD data, indicating that the impact duration of most traffic incidents is within 10 minutes. Since traffic accidents and scattered objects are the primary types of incidents, this study focuses on constructing models for estimating the impact of these two majority classes. Considering that the impact duration is non-negative, the Tobit model is applied in this research to identify significant factors and estimate the impact duration. The result indicates that the significant variables are mainly related to real-time traffic flow features. In terms of MAE and RMSE, the models for scattered objects tend to exhibit lower error rates. However, the goodness-of-fit of the Tobit model is not sufficiently satisfactory, indicating that some important factors may not be effectively captured. The impact duration is transformed into discrete variables to address this limitation and treated as a classification problem. The random forest model effectively performs classification by employing appropriate data pre-processing techniques. Although In light of the highly right-skewed data distribution, the overall of the accuracy of each class can still reach 0.8. In general, the proposed models can capture the significant variables influencing the impact duration of traffic incidents, and the machine-learning-based models have satisfying performance in classification. Accordingly, for practical applications, the delays caused by traffic incidents can be predicted based on the associated factors. Accurate estimation can offer valuable information to both traffic management officials and freeway users, enabling them to better understand the current traffic conditions and make informed decisions. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/90810 |
DOI: | 10.6342/NTU202303893 |
全文授權: | 同意授權(限校園內公開) |
顯示於系所單位: | 土木工程學系 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-111-2.pdf 目前未授權公開取用 | 1.92 MB | Adobe PDF | 檢視/開啟 |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。