| 摘要: | 本研究聚焦於高速公路壅塞預測，旨在建立一套以電子道路收費系統(Electronic Toll Collection, ETC)資料為核心，結合監督式機器學習技術之區間中位數車速預測模型。研究對象選定國道三號北上土城至中和路段，透過蒐集2022年1月之ETC車速與流量資料，並以5分鐘為時間單位，進行系統性資料處理與建模實作。研究特別著重於時間序列特性與交通流動變異性之解析，自變數設計涵蓋歷史車速、歷史交通流量及時間特徵(小時、星期與是否假日)三大類別，共計18項變數；預測目標則設定為未來5、10、15、30、45與60分鐘後之區間中位數車速，反映短中長期不同交通管理需求。
在模型建構部分，本研究實際採用六種常見之監督式機器學習模型進行實驗比較，包括線性回歸、k -近鄰、隨機森林、梯度提升機、類神經網與適應性增強法。原規劃之支持向量回歸與長短期記憶網路經初步測試後表現不佳，故未納入正式實驗比較。建模流程皆透過Orange資料探勘平台實現，確保實驗重現性與結果一致性。預測效能則採用五項指標進行量化評估：均方誤差、均方根誤差、平均絕對誤差、均方根百分比誤差與決定係數。
實驗結果顯示，完整時間特徵組合(包含小時、星期與假日變數)於中長期預測表現最穩定，能有效捕捉交通流變化之周期性與異質性；模型方面則以隨機森林與梯度提升機於大多數時間窗中呈現最佳預測準確度，優於其他模型。整體而言，本研究證實ETC資料具備高度覆蓋率與即時性，適合應用於高速公路壅塞預測之實務場景，亦顯示機器學習方法於交通流預測領域之發展潛力。本研究不僅具備理論創新性，亦具備高度實務導向，可提供交通主管機關於壅塞預警、資訊推播與路網調度決策上作為重要參考，進一步推動智慧交通系統之發展與應用。
;This study focuses on the issue of highway congestion prediction and aims to develop a segment-based median vehicle speed prediction model using data from the Electronic Toll Collection (ETC) system, integrated with supervised machine learning techniques. The selected study area is the northbound section of National Freeway No. 3, from Tucheng to Zhonghe. ETC speed and traffic volume data from January 2022 were collected at 5-minute intervals, followed by systematic data processing and modeling implementation. Emphasis was placed on analyzing the temporal characteristics and variability of traffic flow. The independent variables include 18 features across three main categories: historical speeds, historical traffic volumes, and temporal features (hour of day, day of the week, and holiday indicator). The prediction targets are the median speeds of the segment 5, 10, 15, 30, 45, and 60 minutes into the future, covering short-, medium-, and long-term traffic management needs.
In terms of model construction, six commonly used supervised machine learning models were tested and compared, including Linear Regression, k-Nearest Neighbors, Random Forest, Gradient Boosting Machine (GBM), Neural Networks (NN), and Adaptive Boosting. Support Vector Regression and Long Short-Term Memory (LSTM) were initially considered but excluded from final experiments due to unsatisfactory preliminary results. The entire modeling workflow was implemented using the Orange data mining platform to ensure reproducibility and consistency of results. Model performance was evaluated using five metrics: Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), Root Mean Squared Percentage Error (RMSPE), and Coefficient of Determination (R²).
Experimental results show that the complete set of temporal features (hour, weekday, and holiday) provided the most stable performance for medium- to long-term predictions, effectively capturing periodic and heterogeneous traffic patterns. Among the models, Random Forest and GBM consistently achieved the best prediction accuracy across most time windows, outperforming the other methods. Overall, this study confirms that ETC data, due to its high coverage and timeliness, is highly suitable for practical applications in highway congestion prediction. It also demonstrates the promising potential of machine learning approaches in traffic flow forecasting. The study contributes both theoretical innovation and practical value, offering transportation authorities critical insights for congestion warning systems, information dissemination, and road network management—thus promoting the development and application of intelligent transportation systems (ITS).
Keywords: Highway Congestion Prediction, Electronic Toll Collection, Machine Learning, Time Series Forecasting, Intelligent Transportation Systems. |
