時間序列模型預測公共自行車短期需求變化比較研究;Comparison of Time Series Models for Public Bicycle Short-Term Demand Prediction

NCUIR > School of Management at National Central University > Executive Master of Information Management > Electronic Thesis & Dissertation > Item 987654321/98277

Please use this identifier to cite or link to this item: https://ir.lib.ncu.edu.tw/handle/987654321/98277

Title:	時間序列模型預測公共自行車短期需求變化比較研究;Comparison of Time Series Models for Public Bicycle Short-Term Demand Prediction
Authors:	邱裕軒;Chiu, Yu-Hsuan
Contributors:	資訊管理學系在職專班
Keywords:	機器學習;時間序列預測;公共自行車;資料探索分析;特徵工程;machine learning;time series forecasting;public bicycle;exploratory data analysis;feature engineering
Date:	2025-07-21
Issue Date:	2025-10-17 12:34:21 (UTC+8)
Publisher:	國立中央大學
Abstract:	隨著永續發展與淨零排放成為全球關注議題，公共自行車系統在推動綠色運輸與智慧城市建設中展現了潛在的價值。透過各區域站點所蒐集之即時資料，結合天氣與時間等外部資訊，可以有效預測需求變化量並有系統地調整營運資源。本研究以臺中市 YouBike 2.0 系統為研究對象，比較不同時間序列模型於短期需求預測之表現。雖然近年來已有部分研究應用機器學習方法預測公共運輸需求，但鮮少有研究系統性比較多種時間序列模型在不同場域與預測區間下之表現，導致營運單位做出即時調度決策時缺乏足夠的參考依據。本研究實驗利用 ARIMA、XGBoost 與 LSTM 三種預測模型，並加入時間與天氣特徵進行預測分析，從站點層級評估模型在短期預測任務下的準確性，並透過消融實驗分析特徵組合對預測準確度之影響。實驗結果顯示，XGBoost 在大多數情境中具備最佳整體預測表現，特別是在高使用量站點與較長預測時間下，R² 可達 0.53 以上。LSTM 模型於部分站點亦表現穩定，惟對資料完整性與特徵變動較為敏感，在較長時間預測下具有較佳準確性。ARIMA 模型整體表現不佳，無法有效處理非線性與高波動性資料。此外，加入過多時間與天氣特徵未能進一步提升預測效能，反而在僅使用單一天氣特徵時表現更佳。空間場域分析結果顯示，學區型與商圈型站點因使用模式較為規律，預測效果較佳。本研究建立一套具擴展性與實務應用價值的預測流程，提供模型選擇與特徵設計之建議，期能作為未來智慧交通管理與綠色運輸調度政策的重要參考。;This study explores short-term demand forecasting for public bicycle systems using the YouBike 2.0 system in Taichung. Although machine learning has been applied to transport demand prediction, few studies have systematically compared statistical, machine learning, and deep learning models across different station types and forecast horizons. This research develops ARIMA, XGBoost, and LSTM models, integrating temporal and weather features, and evaluates their performance through feature combination and ablation experiments. Results show that XGBoost consistently outperforms other models, achieving R2 values above 0.53 at high-demand stations and longer intervals. LSTM demonstrates stable performance at select stations but is more sensitive to data quality and feature dimensions. ARIMA performs poorly overall due to its limited adaptability to nonlinear and volatile demand. Moreover, including too many time and weather features does not improve performance. On the contrary, excessive features reduce accuracy, while using a single weather variable (e.g., rainfall or temperature) yields better results. Spatial analysis also reveals that stations in school and commercial zones, where usage patterns are more regular, yield higher predictive accuracy. This study offers a scalable forecasting framework and practical insights for model selection, feature engineering, and real-time dynamic deployment in public bicycle systems.
Appears in Collections:	[Executive Master of Information Management] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	37	View/Open

社群 sharing

Loading...