Short-Term Forecasting of Train Demand by Time Series Methods

• Main Authors: Chi-Yuan Yew, 游智元
• Other Authors: Chi-Kang Lee
• Format: Others
• Language: zh-TW
• Published: 2006
• Online Access: http://ndltd.ncl.edu.tw/handle/63209252938252283428

碩士 === 國立成功大學 === 交通管理學系碩博士班 === 94 === The Taiwan Railroad Administration (TRA) is a very important transportation in Taiwan, and transport passengers all the days. How to operate effecticient and reasonable is a very important task. But presently the seat mamgement of TRA still exist some defects, and they will effect revenue of TRA. So my thesis will forcus on: (1) The travel demand of TRA and try to construct suitable short-term forecasting model for daily TRA trains. (2) To compare performance to find some imformation between train datas & models. (3) Finally, try to hybrid different models to get better forecasting performance. Time Series Analysis (TSA) is according to past observation to construct suitable model and predict the tendency of future. The Autoregressive Integrated Moving Average (ARIMA) includes three parts- Autoregressive, AR (p); Differential (d) & Moving average, MA (q), and we can using these parts to construct suitable model. The Exponential Smoothing Model (ESM) can use to construct model with exponential function & smoothing curve by weighted past data. Both models need data observation, so according data features we can construct model with suitable method. After data collection, data analysis, model construction and forecasting, we got some conclusion: (1) Different trains will show different features because of time and date, and these features will effect forecasting performance. (2) ARIMA and ESM can obtain promising performance via trial and error. MAPE of out-of-sample is below 20%. (3) Box-Cox transformation of raw data can improve out-of-sample performance. (4) The proposed hybrid model can further upgrade out-of-sample performance significantly.