| 摘要: | 民用航空運輸業在全球化經濟體系中扮演重要角色，促進經貿往來及商務旅遊發展。近期面臨國際間戰爭、貿易紛爭擴大及全球供應鏈體系重組等挑戰，導致金融市場、股匯市及能源供需劇烈波動。燃油支出為各航空業者最大營運成本，有效預測航空油價為成本評估與航線效益決策之關鍵。本研究旨在開發航空燃油價格預測模型，探討四種不同時間序列預測方法，包括單變量、多變量、多步驟及遞迴預測模型。使用航空燃油價格數據，結合三大原油市場、標普500指數及美元指數，預測技術比較整合移動平均自迴歸模型和四種深度學習循環神經網路架構包括簡單循環神經網路、長短期記憶網路、閘門循環單元及雙向長短期記憶網路並運用網格搜索方法找到最佳超參數配置強化模型表現。研究結果顯示基於LSTM的深度學習預測模型在各種預測方法中表現皆優於ARIMA模型，在預測次週油價時，GRU模型的平均絕對百分比誤差(MAPE)為3.54%，而在預測次月油價時，雙向LSTM模型的MAPE為7.57%。多變量預測效能略低於單變量預測，原因為多重序列間的特徵權重未優化。在長期預測方面，本研究模型預測未來一年航空油價的誤差較預算油價基準平均降低7.97%，證明LSTM模型在短期預測中表現出色，在長期預測上具有發展潛力。;The civil aviation transport industry plays a significant role in the globalized economic system, facilitating trade and business travel. Recently, it has faced challenges such as international conflicts, trade disputes, and the reorganization of global supply chain, leading to fluctuations in financial markets, stock and exchange rates, and energy supply and demand. Fuel expenses constitute the largest operating cost for airlines, making effective fuel price prediction crucial. This study aims to develop an aviation fuel price prediction model, using four time series forecasting methods: univariate, multivariate, multi-step, and recursive models. The research utilizes historical fuel price data combined with data from major crude oil markets, the S&P 500 index, and the US dollar index. The predictive techniques integrate the Autoregressive Integrated Moving Average model with four deep learning recurrent neural network architectures： Simple Recurrent Neural Network, Long Short-Term Memory, Gated Recurrent Unit, and Bidirectional LSTM, using grid search to find the optimal hyperparameter configuration. The results indicate that the LSTM-based prediction models outperform the ARIMA model across various forecasting methods. For next-week price prediction, the GRU model achieves a Mean Absolute Percentage Error of 3.54%, while for next-month prediction, the BiLSTM model achieves a MAPE of 7.57%. The performance of multivariate prediction is slightly weaker than that of univariate prediction, due to the suboptimal feature weighting among multiple sequences. In long-term forecasting, the proposed model reduces the error in predicting the next year’s fuel prices by an average of 7.97% compared to budgeted price benchmarks, demonstrating the LSTM model′s excellent performance in short-term prediction and its potential for long-term forecasting. |
