- 學位論文
應用人工智慧遷移學習預測營造工程物價指數
Predicting Construction Cost Index using Artificial Intelligence and Transfer Learning
摘要
營建成本超支與工程計價爭議常因材料價格波動而發生，因此需準確預測營建物價，且其可協助預測投標價格、合約管理、估算成本與採購管理等，同時減少工程契約計價爭議。深度學習近年發展快速，但過往較少利用深度學習模型預測營建物價且成果有限，多數研究使用營建物價指數或經濟指標預測短期營建物價。然而，收集經濟指標費時費力，且資料量不足以訓練良好的深度學習模型，而預測短期營建物價與實務較不符合。因此本研究提出使用深度學習結合遷移學習預測營造工程物價指數之框架。利用長短期記憶網絡(Long Short-Term Memory, LSTM)配合預訓練與微調以預測短期與長期之台灣營造工程物價指數，以提升機器學習預測營造工程物價指數之表現。此外，將探討利用單一變數衍生之技術分析指標預測營造工程物價指數之可行性與成效。本研究以營造工程物價總指數、水泥及其製品類指數、金屬製品類指數為目標域，並選擇 7 種匯率作為預選源域，使用 14 種源域選擇方法確定最終源域。模型以 2 層 LSTM 層為基礎，通過預訓練和微調的方法提高預測效能。源域選擇結果顯示，營造工程物價總指數、水泥及其製品類指數之最終源域為新台幣兌日圓之匯率、金屬製品類指數之最終源域為新台幣兌韓元之匯率。預測模型方面，預訓練與微調能夠提升預測營造工程物價指數的效能。在預測營造工程物價總指數與水泥及其製品類指數時，目標域資料集使用技術分析指標比經濟指標好。本研究提出應用預訓練與微調技術預測營造工程物價指數之框架。為首個利用此方法協助預測營造工程物價指數以提高預測表現與泛化能力之研究。這種結合允許模型在新的領域中學習，從而更好地適應不同的營建市場。且此方法優於其他機器學習模型，並能準確預測長期之營造工程物價指數，從而使預測結果更具可靠性和實用性。此外，本研究提出了利用技術分析指標的可行性，並對預測效果進行了探討，進一步豐富了預測方法的多樣性。本研究提出之創新性方法為未來相關研究提供了新的思路，同時也為業界提供了更準確地預測和管理營建成本的手段。這些成果將對營建領域的實踐和研究提供重要參考，有助於提升工程項目的管理效率和成功率。
並列摘要
Construction cost overruns and valuation disputes are common due to fluctuating material prices. Accurate predictions are crucial for forecasting bid prices, contract management, cost estimation, and procurement management. Despite advances in deep learning, its use in predicting construction costs has been limited. Most studies rely on construction cost indices or macroeconomic indicators for short-term forecasts, which are time-consuming and labor-intensive. This study explores the feasibility of technical analysis indicators for more accurate cost predictions and management. It proposes a framework using deep learning and transfer learning, specifically utilizing Long Short-Term Memory (LSTM) networks with pre-training and fine-tuning to predict both short-term and long-term construction cost indices in Taiwan. The research focuses on the general construction cost index, the cement and its products index, and the metal products index as target domains. Seven exchange rates are selected as initial source domains, and 14 source domain selection methods determine the final source domains. The model is based on a two-layer LSTM architecture, with pre-training and fine-tuning enhancing prediction performance. Results indicate that the final source domain for the general construction cost index and the cement and its products index is the New Taiwan Dollar to Japanese Yen exchange rate. In contrast, for the metal products index, it is the New Taiwan Dollar to Korean Won exchange rate. Pre-training and fine-tuning significantly improve prediction performance. Technical analysis indicators outperform macroeconomic indicators for the general construction cost index and the cement and its products index. This study presents a novel framework for predicting construction costs through pre-training and fine-tuning, marking the first application of this method in the field. This approach allows the model to better adapt to different markets, outperforming other machine learning models and providing more reliable long-term predictions. The study showcases technical analysis indicators in construction cost prediction, offering a more accurate method for cost management and project success.