"過去台灣地區港埠運量預測相關研究皆著重在20至30年的長期預測為 主，長期的預測著重於趨勢分析，若以設施規劃為目的之運量預測則應強 調短中期預測量的準確性，至於觀測資料季節性的變動型態時，短期預測 卻優於長期預測。而港埠運量的預測方法通常可分為多變量與單變量模 式，其中又以多變量模型的預測居多，多變量模式可利用總體外生變數來 解釋與貨櫃吞吐量間之關係，但選擇變數與收集變數資料較費時，而單變 量不易受其他外生變數影響而產生偏誤。 本研究之目的為建立和比較六種單一變數預測方法與類神經網路，並 尋找出適用於港埠貨櫃吞吐量最佳的預測方法，以預測台灣國際港口之每 月貨櫃吞吐量。研究對象為台灣地區三大國際港埠：基隆港、台中港與高 雄港之貨櫃吞吐量。透過古典分解法、三角函數迴歸、季節性虛擬變數、 灰預測、組合灰預測、SARIMA 與類神經網路等七種預測模式，將產生三 港之七種方法的貨櫃吞吐量預測值。再經由實證分析，利用平均絕對誤差 (MAE)、平均絕對誤差百分比(MAPE)及殘差均方根(RMSE)等評估指標比較 後，以驗證何者可提供預測最佳之精確度。 根據本研究之結果找出準確度最高之預測方法，將可提供未來港埠當 局進行港埠規劃、貨櫃運量預測及港埠裝卸能量時之參考。""In the past, the research about the throughput volumes of international ports in Taiwan has focused on the long term forecast with a range between 20 and 30 years. Long term forecast emphasizes on the trend analysis. While the purpose of forecast is used to construct a facility planning, the middle term and short term forecast accuracies should be more emphasized. As the data with seasonal variations, the short term forecast methods provide better results than those of the long term forecast method. In general, the forecast method of the throughput volumes of Taiwan ports can be classified into multivariate and univariate forecast methods. And the multivariate forecast method accounts for a large portion of past research. The advantage of the multivariate forecast method is that it can be used to explain the causal relationship between exogenous variables and the throughput volumes, but the drawback of the method is that it is more time consuming to collect data. On the contrary, the univariate forecast method is less sensitive to the influence of the exogenous variables. The purpose of this study is to make comparisons on six different univariate forecasting methods and the neural network method, to provide a more accurate forecasting model on the container throughput for rendering a reference to authorities. Six different univariate methods, namely the Classical Decomposition Model, the Trigonometric Model, the Regression Model with Seasonal Dummy Variables, the Grey Forecast, the Hybrid Grey model, and the SARIMA as well as the neural network will be used. The monthly data will be collected to construct the forecasting models. The forecasting results of the seven methods will be obtained and compared based on commonly used evaluation criteria, Mean Absolute Error, Mean Absolute Percent Error and Root Mean Squared Error. We will find the most suitable method providing the highest forecast accuracy. The outcome of this work can be helpful to predict the near future demands for the container throughput of the international port for rendering a reference to authorities."