基於大數據與深度學習法之高速公路高乘載車輛車道績效評估

Performance Assessment of Freeway HOV Lanes via Big Data and Deep Learning

高乘載車輛（HOV）車道具有鼓勵小汽車共乘和公共運輸優先通行的誘因，但往往因為使用率低於一般車道而飽受批評。本研究從高速公路局的開放資料庫擷取9個月、77,611個文件檔，進行量化的交通評估，以檢視國道五股楊梅（五楊）高架路段HOV車道和一般車道之間的動態使用關係。分析結果顯示，現階段整體運行狀況平衡，但在某些連續假日出現HOV車道為順暢的A-C級服務水準，一般車道為壅塞的E-F級，形成車道管理熱點，本研究進而開發一種基於深度學習法的多層感知器模型，用於預測壅塞路段和時段，模式預測結果在合理誤差範圍，具有實務應用可行性。文末總結建議（1）調整五楊高架路段的機場系統交流道以南之HOV南向車道管制措施，以優化車道容量使用；（2）健全共乘環境以發揮HOV車道的設置效益；（3）開放資料庫的檔案資料格式整合以增進使用友善性，並以更大量的交通數據升級模型準確性。

High-occupancy vehicle (HOV) lanes provide incentives to encourage carpooling and prioritize public transit, but they usually face criticism for underutilization compared to general-purpose (GP) lanes. This research collected nine months of quantitative traffic data in 77,611 files from the Freeway Bureau’s open database. Such data assessed utilization dynamics between HOV and GP lanes on the Wu-Yang Freeway Viaduct in northern Taiwan. The analysis revealed an overall balanced operational state. However, on certain consecutive holidays, specific HOV segments exhibited surplus capacity with levels of service (LOS) A to C, while GP lanes experienced severe congestion at LOS E to F, identifying lane management hot spots. A deep learning-based multilayer perceptron model was thus developed to predict congested segments and periods, achieving results within a reasonable margin of error and demonstrating practical feasibility. The research recommendations include (1) adjusting control measures for southbound HOV lanes south of the Airport Interchange to optimize lane usage, (2) enhancing carpooling environments to maximize the benefits of HOV infrastructure, and (3) integrating the open database’s file and data formats to improve user-friendliness while incorporating additional traffic data for even better model accuracy.