臺灣博碩士論文加值系統

旅行時間的預測一直都是項重要的交通議題，可靠的旅行時間預測對於駕駛人來說，在旅行或通勤的事先規畫上不但能夠更加精確，還有機會避免疲勞駕駛等相關風險。本篇論文使用一個混合的圖卷積網路架構，以高公局提供的即時國道資訊及時間特徵，來預測國道旅行時間。
我們延伸現有的Graph WaveNet [7]來預測旅行時間，在短路段與長路段的旅行時間預測上，分別使用八層跟六層的Graph WaveNet去處理原始的交通特徵輸入，再連接全連接層、1x1卷積去依據不同路段做旅行時間的計算。我們使用多種不同的相鄰矩陣與擴散步數，去探討兩者間的相互影響，在短路段旅行時間的預測上，平均相鄰節點數落在4到10之間會有較好的結果，擴散步數設為2最為恰當；在長路段旅行時間的預測上，平均相鄰節點數在6以上結果較好，擴散步數設為2或3較為恰當，我們推斷這是因為參考較廣的車輛偵測器(Vehicle Detector, VD)資料對長路段旅行時間的預測較為合適。此圖卷積架構能夠很好的使用圖資料，依據節點間的空間關聯性去模擬車流隨時間的推進與擴散，使突發性的塞車較能被預測，不管在短路段、長路段的旅行時間預測，都有較好的結果，在短路段旅行時間的預測上可達3.39% MAPE (Mean Average Percentage Error)，相較傳統CNN改善了0.64%，在塞車的情況下，相較傳統CNN方法平均改善了2.9%。除此之外，此架構也可以緩解VD資料缺值、路段VD數量稀少所衍生的問題，同時預測多路段多時序的旅行時間，減少分開預測所需的硬體資源及訓練時間。

Travel time prediction plays an important role in intelligent traffic systems. Reliable travel time prediction for drivers not only increase accuracy in prior plaining of travel, but also has the opportunity to avoid risk of fatigue driving. This paper uses a mixed graph convolutional network architecture to predict the travel time of the national freeway with time features and the real-time freeway information provided by Taiwan Freeway Bureau.We extend the existing Graph WaveNet [7] to predict freeway travel time. For short distance route travel time prediction, we use 8-layer GWN, and four long-distance path travel time prediction, 6-layer GWN, to process the traffic speeds and flows collected by VDs (Vehicle Detectors), and then connected to the fully connection layer or 1x1 CNN. We use various adjacency matrices and diffusion steps to explore the dependences among VD data. For the prediction of the route travel time, the average connected nodes of the adjacency matrix between 4 and 10 have better results, and choosing diffusion step as 2 is more appropriate. For the prediction of path travel time, the average connected nodes above 6 is better, and diffusion step of 2 or 3 is more appropriate. We infer that the reference to the wider VD data is more suitable for path travel time prediction. The graph convolution architecture can make good use of road network data, and simulate the progress and diffusion of traffic flow over time according to the spatial correlation between locations, so that traffic jam can be more predictable for both route and path predictions. The MAPE (Mean Average Percentage Error) of route travel time prediction of 30-min ahead can reach 3.39%, which is an improvement of 0.64% compared with the traditional CNN. For the prediction of traffic congestion, the average improvement is 2.9% compared with the traditional CNN method. In addition, this architecture can also alleviate the problems caused by the lack of VD data, and simultaneously predict the travel times of multiple road sections and multiple time series, and thus reduce the hardware resources and training time required for the whole road network predictions.

摘要 i
Abstract ii
致謝 iii
目錄 iv
表目錄 vi
圖目錄 vii
一、緒論 1
1.1 研究背景 1
1.2 研究動機 2
1.3 研究目的 4
二、相關研究 7
2.1 Graph Convolution Networks 7
2.2 Temporal Convolution Networks 7
2.3 Graph WaveNet 8
三、研究方法 11
3.1 資料前處理 11
3.1.1 處理VD資料 11
3.1.2 處理M04A資料 12
3.1.3 相鄰矩陣 13
3.1.4 處理預測目標 – 短路段旅行時間 14
3.1.5 處理預測目標 – 長路段旅行時間 14
3.2 類神經網路架構設計 15
3.2.1 短路段旅行時間預測 15
3.2.2 長路段旅行時間預測 17
四、實驗結果 19
4.1 短路段旅行時間預測 19
4.1.1 不同輸入資料與不同架構的比較 19
4.1.2 以不同相鄰矩陣做訓練 20
4.1.3 VD異常與M04A資料缺失的衍生問題 21
4.1.4 比較塞車時段的預測表現 22
4.1.5 比較假日與平日的預測表現 24
4.1.6 自適應相鄰矩陣的訓練結果 25
4.1.7 不同路段的訓練結果 26
4.1.8 特殊路段的塞車時段預測情況 27
4.2 長路段旅行時間預測 29
4.2.1 不同架構的比較 29
4.2.2 以不同相鄰矩陣做訓練 30
4.2.3 比較壅塞時的預測表現 31
4.2.4 比較假日與平日的預測表現 34
4.2.5 自適應相鄰矩陣的訓練結果 35
結論與未來展望 37
參考文獻 39

[1] Thomas N. Kipf, Max Welling “Semi-Supervised Classification with Graph Convolutional Networks” arXiv:1609.02907 Wed, 22 Feb 2017
[2] Michaël Defferrard, Xavier Bresson, Pierre Vandergheynst “Convolutional Neural Networks on Graphs with Fast Localized Spectral Filtering” arXiv:1606.09375 Sun, 5 Feb 2017
[3] Fisher Yu, Vladlen Koltun “Multi-Scale Context Aggregation by Dilated Convolutions” arXiv:1511.07122 Sat, 30 Apr 2016
[4] Keng-Hui Wu “Deep Learning for Short-Term Freeway Travel Time Prediction.” July, 2018
[5] Bing Yu, Haoteng Yin, Zhanxing Zhu “Spatio-Temporal Graph Convolutional Networks: A Deep Learning Framework for Traffic Forecasting” arXiv:1709.04875v4 Thu, 12 Jul 2018
[6] Yaguang Li, Rose Yu, Cyrus Shahabi, Yan Liu “Diffusion Convolutional Recurrent Neural Network: Data-Driven Traffic Forecasting” arXiv:1707.01926 Thu, 22 Feb 2018
[7] Zonghan Wu, Shirui Pan, Guodong Long, Jing Jiang, Chengqi Zhang “Graph WaveNet for Deep Spatial-Temporal Graph Modeling” arXiv:1906.00121 Fri, 31 May 2019
[8] David I Shuman, Sunil K. Narang, Pascal Frossard, Antonio Ortega, Pierre Vandergheynst “The Emerging Field of Signal Processing on Graphs: Extending High-Dimensional Data Analysis to Networks and Other Irregular Domains” arXiv:1211.0053 Sun, 10 Mar 2013
[9] Sam Shleifer, Clara McCreery, Vamsi Chitters “Incrementally Improving Graph WaveNet Performance on Traffic Prediction” arXiv:1912.07390 Wed, 11 Dec 2019
[10] Hoang Thanh Lam, Ernesto Diaz-Aviles, Alessandra Pascale, Yiannis Gkoufas, Bei Chen “(Blue) Taxi Destination and Trip Time Prediction from Partial Trajectories” arXiv:1509.05257 Thu, 17 Sep 2015
[11] Lin- Chung Hsiang “Using GBRT to predict short-term freeway traffic condition” July, 2018
[12] Hau-Chi Lin “Freeway Traffic Congestion Prediction Using Neural Network” July, 2018
[13] Aaron van den Oord, Sander Dieleman, Heiga Zen, Karen Simonyan, Oriol Vinyals, Alex Graves, Nal Kalchbrenner, Andrew Senior, Koray Kavukcuoglu “WaveNet: A Generative Model for Raw Audio” arXiv:1609.03499 Mon, 12 Sep 2016
[14] https://tisvcloud.freeway.gov.tw/
[15] https://localmotors.com/meet-olli/
[16] https://traffic.tycg.gov.tw/businessD/post/upt.aspx?p0=72935