基於公共自行車開放數據之使用量預測研究

Abstract

近些年隨著綠色運輸理念的進步，公共自行車系統越來越受到人們的重視，其使用量及站點建置規模也在逐步擴大，在發展過程中也出現了大大小小的問題，對使用者的意願產生了一定的影響，其中無車可借及無位可還的現象對使用者滿意度的影響尤為明顯。因此，能否及時滿足該部分人群的需求，也就成了亟待管理者解決的問題之一。 故一套完善的使用量預測模式將會輔助運營方有效地調補車輛，減少等待對使用者產生的影響。但過去對公共自行車的預測研究主要採用基於歷史資料的平均使用量分析，或者是進一步的線性回歸與決策樹模式，這些方法往往針對橫斷面進行分析而忽視了其時間維度上變動的趨勢。 此外由於公共自行車系統相對於其他大眾運輸運具，更易受到外在環境的影響，天氣狀況的改變也會導致使用量的短期變化。這也導致了單一時間序列模式或針對歷史資料的決策樹模式無法更為精確的描述使用量的變化情況。本研究首先應用自我回歸移動平均整合模式（ARIMA）、三次指數平滑法（Holt-Winters）以及倒傳遞神經網路（BP Neural Network）分別獨立建立預測模式，結果顯示，單一預測模型對使用量預測的精度有限，故考量建立混合預測模型，應用Holt-Winters算法擬合（fit）原始時間序列數據，並應用ARIMA模式進一步擬合Holt-Winters的殘差項，並應用倒傳遞神經網絡結合天氣因素擬合處理過的殘差項，通過將模型動態迭代以得出穩定的時間序列模式及天氣擾動之神經網路模型，建構出精確度更高的混合模型。

In recent years, with the improvement of green transportation idea, increasingly importance has been attached to public bicycle system. The usage and the number of stations increased year by year. There are some big or little problems occurred in the development process of public bicycle system. The two of the serious factors which effect the user satisfaction are no bike for rent and no parking place. So how to meet the part of demand is the urgent problem need operator solve. A completed forecast model could help the operator deal with the problem, reduces the impact for user. However, the most of the past research are concerned on average usage based on the historic data or linear regression and decision tree models. These methods always employing with cross-section cannot include time progress attributes. The tendency of demand along time passage is ignored in them. Compared to other transportation tools, the usage character of bicycle is different. It is more easily influenced by the external environment, especially the weather. The phenomenon causes the single time series or decision tree model is hardly fit the usage data very precisely. The ARIMA model, Holt-Winters model and BP neural network are modeling independent first. In spite of the results show that the single models could also forecast the usage, the accuracy of models are very limited. The Hybrid dynamic model has been built which use the Holt-Winters model break down the original data and employ the ARIMA model fit the residuals. The BP neural network is used to fit the filtered residuals after the process and amend the original data every period. The more stable and precise models could be built after the iteration.