本研究以桃園市境內之月租借數量最高的租賃站―中壢火車站前站為研究對象，先將借車需求量每十五分鐘作為一個時間單元，以今日變動、長期變動及今日變動結合長期變動共三方向做實驗設計。本研究引用類神經網路，以遞歸神經網路建構一個公共自行車之借車需求預測模式，採用三層、完全連結節線及回饋式的網路架構，配合即時遞歸演算法建構不同輸入資料之預測，利用歷史借車需求量資料作為遞歸神經網路訓練與測試基礎。經由各種情境條件設計下不斷測試，比較分析結果得知，本研究所構建之借車需求量預測模式，以當日結合前三周之輸入方法預測效果最好，欲使平均誤差降到最低，輸入數量以前10筆最佳，準確度可達92.169%。因此在公共自行車即時借車需求量預測方面，本研究可提供未來相關單位，作為後續營運上策略研擬之參考。 ;Borrowing public bicycle problem is one of the main concerns of public for a long time, and the thorny problems of the government. Therefore, to solve borrowing problem is the most urgent. Owing to the development of Intelligent Transportation Systems and user’s information have attracted much interest, real-time information of borrowing bicycle is getting more and more important. As to numbers of borrowed bicycles are changeable with time, it’s necessary to make efficiently real-time controlling policies by forecast rental demands accurately.
This study have selected the rental station of Zhongli rail-way station (Front) which rented bicycle the most per month in Taoyuan City. In order to promote the forecasting ability of model, it reviews kinds of rental demand forecasting modeling and analyzes the rental demand of public bicycle, quote from Recurrent Neural Network with three layers, fully connected and feedback network, and Real-Time Recurrent Learning to build forecasting model. Using the historical rental data of public bicycle would be the base of training and testing mode of Recurrent Neural Network. After repeatedly correcting and testing, this model would forecast effectively with small error and high accuracy. As a result, this thesis can be provided a way to forecast rental station in real-time rental demand estimation.
