| 摘要: | 捷運客流預測是先進交通信息系統的重要組成部分，協助捷運當局執行票務分配、運營規劃、收益管理、捷運站管理、行銷方案規劃等工作，或者在極端情況下協助公司應急管理。許多國內的研究試圖應用參數機器學習模型與深度學習模型執行客流預測，然而參數機器學習模型隨著數據的增加存在一定的局限性，而深度學習訓練模型相當耗時，如今集成式學習模型於國外研究與人工智慧競賽中被廣泛運用，本研究提出以Random Forest、AdaBoost、XGBoost三種集成式學習模型與Neural Network深度學習模型比較。
車站客流受週期、假期、尖離峰時段、特殊節日或大型活動等多種因素影響很大，從數據中提取關鍵特徵對於客流預測模型至關重要。本研究所提出的Random Forest與XGBoost模型可以在真實世界的數據集上達到較佳的預測精度和計算效率。此外，刪除大型活動與特殊節日兩種離異值可獲得更佳的預測結果，而增加降雨量天氣因子，對於桃園捷運客流預測影響不大。
;MRT passenger flow forecasting is an important part of an advanced traffic information system, assisting MRT authorities in performing ticket distribution, operation planning, revenue management, MRT station management, marketing plan planning, etc., or assisting companies in emergency management in extreme cases. Many domestic researches try to apply parametric machine learning models and deep learning models to perform passenger flow forecasting. However, parametric machine learning models have certain limitations with the increase of data, and deep learning training models are quite time-consuming. Today ensemble learning is widely used in foreign research and artificial intelligence competitions. In this study, three integrated learning models, Random Forest, AdaBoost, and XGBoost, are proposed to compare with the Neural Network deep learning model.
Station passenger flow is greatly affected by various factors such as cycles, holidays, off-peak hours, special festivals or large-scale events. Extracting key features from data is crucial for passenger flow prediction models. The Random Forest and XGBoost models proposed in this study can achieve better prediction accuracy and computational efficiency on real-world datasets. In addition, deleting the two outliers of large activities and special festivals can get better prediction results, while adding the weather factor of rainfall has little effect on the forecast of Taoyuan MRT passenger flow. |
