具非重現性擁擠特性之高速公路旅行時間預測

Abstract

旅行時間資訊之提供一直為促使智慧型運輸系統(Intelligent Transportation System, ITS)更為成功之重要因素。以往研究更指出非重現性擁擠(non-recurrent congestion)為造成高速公路延滯之主因。因此，如何提高含非重現性擁擠下之旅行時間預測能力，為旅行時間預測領域之重要且必須克服議題。本研究透過類神經 (Artificial Neural Network, ANN)建構台灣中山高速公路36.1公里長路段(包含8個交流道)之旅行時間預測模型，並且為提高模式於含非重現性擁擠之預測能力，本研究除蒐集雨量與時間特徵等資料外，更包含車輛偵測器(Vehicle Detector, VD)所蒐集到之平均線點速度(average spot speed)與大型車流量等資料，並且進一步整合電子收費系統(Electronic Toll Collection, ETC)原始資料所推得之歷史旅行時間(Historical Travel Time, HTT)，並以ETC所推得之真實旅行時間(Actual Travel Time, ATT)做為訓練目標(target)，以建立一穩健型之預測模式。 本研究整合分群(K-means)、分類迴歸樹(Classification and Regression Tree, CART)與類神經網路(Artificial Neural Network, ANN)等三種資料探勘技術，透過創造虛擬變數(dummy variable)與萃取關鍵變數(critical variables extraction)方式，不僅可於不增加設備投資情況下提高預測能力，亦同時獲得關鍵地點之關鍵變數資訊，以有效協助管理單位進行系統維護。本研究所發展之混合專家概念為基之預測方法(Mixture of Experts Based Method, MEBM)與整合分群分類為基之預測方法(Integrated Clustering-Classification Method, ICCM)，不論是否採用ETC系統所產生之資料，相較於直接預測而言，皆能提高模式預測能力，且同時降低MAPE值大於20%之樣本百分比。再者，本研究以決策樹成功萃取出三個關鍵變數，分別為星期(day of week)，時間(Time)，51.6公里車輛偵測器(VD)所蒐集之現點速度(即speed5160)。最後，本研究構建之類神經為基之預測方法(Neural Network Based Method, NNBM)、混合專家概念為基之預測方法(Mixture of Experts Based Method, MEBM)、整合分群分類為基之預測方法(Integrated Clustering-Classification Method, ICCM)等三個預測架構，其MAPE值介於6%~9%之間，皆屬於Highly accurate prediction之模式。

The provision of travel time information has been a major factor facilitating the Intelligent Transportation System (ITS) to become more successful. Previous studies have pointed out that non-recurrent congestion is the major cause of freeway delay. Hence, how to improve the prediction capability of travel time in the case of non-recurrent congestion is an important issue that must be overcome in the field of travel time prediction. This study constructs the travel time prediction model for a segment of 36.1 KMs (including eight interchanges) in the National Freeway No. 1, Taiwan by using the Multilayer Perceptron (MLP). To improve the prediction capability of the model in the case of non-recurrent congestion, this study collects data of average spot speed and heavy vehicle volume gathered by vehicle detectors (VDs), and rainfall and temporal features. Furthermore, the historical travel time inferred from the original data of Electronic Toll Collection (ETC) system is also used as the input variable, and the actual travel time inferred from ETC is used as the training target to establish a robust prediction model. This study integrates three data mining techniques, K-means clustering, classification and regression tree, and neural networks, to predict the travel time of freeway with non-recurrent congestion. By creating dummy variables and identifying important variables, not only increase the prediction performance without additional investment, but also important variables are obtained concerning the important locations of equipment in order to effectively assist public transit agencies with system maintenance. The experimental results show that, whether or not the data generated by the Electronic Toll Collection (ETC) system is used as input variables, the travel time prediction method, Mixture of Experts Based Method (MEBM), Integrated Clustering-Classification Method (ICCM), developed in this study is able to improve the prediction performance. Meanwhile, the proposed approach also reduces the percentage of samples with Mean Absolute Percentage Error (MAPE) > 20%. Furthermore, in this study, important variables are extracted from the decision tree in order to predict the travel time. Finally, the prediction models constructed in Neural Network Based Method (NNBM), Mixture of Experts Based Method (MEBM), Integrated Clustering-Classification Method (ICCM), are highly accurate due to the low MAPE values, which are from 6% to 9%.