為解決私人運具造成的交通壅塞與廢氣排放等問題，大眾捷運系統逐漸扮演著都市運輸的重要角色，因此持續提升捷運服務水準就成為營運管理單位努力的目標。其中站間量可提供重要的績效指標，但電子票證數據無法直接推估旅客之移動軌跡，限制其對站間量的預測與分析。因此本研究先利用手機信令資料追蹤旅客於捷運路網的移動軌跡，並於分叉轉運站與相同匯入站間計算其間多條區段之運量分配比例，將此總區段流量比例應用於分岔轉運站之進出旅客數預測值，即可估算未來各區段的站間量。 本研究採用三種模型預測捷運站進出旅客數：函數資料分析(FDA)、長短期記憶(LSTM)與閥控遞迴單元(GRU)。結合總區段流量比例與捷運站進出旅客數預測值，進一步提出捷運站間量的推估方法，實證結果貼近旅客實際的移動軌跡，精度高於現行之「建議路線」軌跡，這主要原因為旅客選擇路徑會受到不同因素影響，包含：路線長度、路線車站數、乘車時間、轉乘時間、擁擠度與彎繞度等。 就臺北車站之預測誤差而言，機器學習模型LSTM(4.58%)與GRU(4.60%)較統計模型FDA(7.04%)為佳，但需較長的訓練時間。針對已觀測區間長度 與未觀測區間長度 的探討，各模型無明顯趨勢，綜合來說，最佳預測組合 仍需依不同模型與數據，根據經驗或測試後選定。最後，根據信令站間量之預測結果顯示，文湖線存在超載狀況，應彈性調派列車或增加列車容量，其餘路線的服務供過於求，則應調整列車服務計畫或採彈性編組的服務模式，以降低成本。 ;Link flow of Mass rapid systems is an important performance indicators to manager. In this research, we use sighting data for tracking the trajectory of passengers in MRT network and calculating the segment proportionality of the transfer station. Then, we can combine the result of flow prediction and the segment proportionality to predict short-term link flow. In this research, we use functional data analysis (FDA), long short-term memory (LSTM) and gate recurrent unit (GRU) model to predict passenger flow. Combined Total segment proportionality and flow prediction, we propose an estimation method for link flow. The empirical results are close to real situation. The accuracy is higher than practical method which use “suggested route” trajectory. Because the route choice of passengers is affected by many factors such as travel time, transfer time, degree of crowding and degree of circuitousness of path. The results of passenger predict in Taipei Main Station show that the performance with machine learning method (LSTM:4.58%; GRU:4.60%) are better than FDA(7.04%). The result of link flow predict shows that Wenhu Line overloading situation occurs during peak hours on weekdays and weekends. The flexible schedule of the train should be assigned by manager. Other routes supply exceed demand. Manager should adjust the train service plan or use elastic way of combination train to lower the cost.
