摘要: | 公共自行車系統(Bicycle-Sharing System, BSS)作為解決公共運輸第一和最後一哩路問題的關鍵,各國政府積極推動其發展並不斷增設新站點。然而,隨著站點增加,BSS的騎乘路線和調度變得複雜,而需要更多資源來優化調度策略和規劃騎乘環境建置。本研究提出一種基於旅運需求預測的旅次分佈長期預測模型,旨在有效預測新設站點建立後的旅次分佈並針對其安排高效的調度策略以及騎乘環境建置成本投入區域。 而在進行旅運需求預測中的旅次發生借車量與還車量的預測,本研究探討議題將牽涉到具備歷史票證資料的既有站點和欠缺歷史票證資料的新設站點,兩種站點形式,有鑑於時間序列資料分析法相對於利用變數推估在預測上有更高的精度,對於既有站點以時間序列資料分析為主提出了統計模型的差分整合移動平均自我迴歸模型(ARIMA)和季節性差分整合移動平均自我迴歸模型(SARIMA)和機器學習的長短期記憶(LSTM)和閥控遞迴單元(GRU)進行預測與比較,比較結果發現LSTM有最佳的預測精度;新設站點則利用氣候、人口、土地使用和周遭站點使用四個主要影響BSS借車量與還車量的因素,並應用統計模型的多元迴歸模型(MLR)和機器學習的極限梯度提升(XGBoost)進行預測和比較,XGBoost 有最佳的預測精度。因此本研究旅次發生階段則採納LSTM和XGBoost分別對既有站點和新設站點進行預測。旅次分佈則是應用傳統的重力模型與深度學習結合而成的深度重力(DG)模型,並以旅運需求預測程序為主要的概念,旅次發生的輸出結果加上距離作為旅次分佈的輸入,對BSS進行旅次分佈的預測,預測結果為0.7884,成功地捕捉到BSS旅次分佈樣態。 最後透過旅次分佈的結果配合新設站點的站點特性,發現若站點是以休閒觀光特性為主,則會以其他包含休閒光觀特性的站點為主要往來;住宅區則以前往休閒光觀景點、交通節點和就學為主要用途。 ;The Bicycle-Sharing System (BSS) is a key solution to the first and last mile problem in public transportation. Governments worldwide actively promote its development by continuously adding new stations. However, as the number of stations increases, BSS routes and dispatching become more complex, requiring more resources to optimize dispatch strategies and plan the riding environment. This study proposes a trip distribution prediction model based on travel demand prediction to effectively predict trip distribution after the establishment of new stations and to arrange efficient dispatch strategies and investment areas for the construction of the riding environment. In predicting the number of borrowings and returns of trips in travel demand forecasting, this study explores the issues related to both existing stations with historical ticket data and new stations lacking this data. Given that time series analysis methods show higher accuracy than variable-based estimates, for existing stations, this study proposes the statistical models: the Autoregressive Integrated Moving Average (ARIMA) model, the Seasonal Autoregressive Integrated Moving Average (SARIMA) model, and machine learning models: the Long Short-Term Memory (LSTM) model and the Gated Recurrent Unit (GRU) model; and performs predictions and comparisons. Results indicate that LSTM has the best prediction accuracy. For new stations, the study utilizes four main factors that influence BSS borrowings and returns: climate, population, land use, and the usage of surrounding stations, and applies the statistical Multiple Linear Regression (MLR) model and the machine learning Extreme Gradient Boosting (XGBoost) model for predictions and comparisons, with XGBoost showing the best prediction accuracy. Therefore, this study adopts LSTM and XGBoost for predicting the trip generation of existing and new stations, respectively.For trip distribution, the study employs a Deep Gravity (DG) model, combining traditional gravity models with deep learning, using the travel demand prediction process as the main concept. The output results of trip generation, combined with distance, serve as inputs for trip distribution prediction. The prediction result has an R² value of 0.7884, accurately capturing the BSS trip distribution patterns. Finally, through the trip distribution results and the characteristics of new stations, the study finds that if a station primarily serves recreational and tourism purposes, it will predominantly interact with other stations with similar characteristics. Residential areas mainly travel to recreational spots, transportation nodes, and educational institutions. |