2026-03-12
【學術亮點】透過一種新型的LSTM-GRU混合深度學習架構增強路面溫度預測，並提高其可解釋性
字體大小
小
中
大
Intelligent Service: Large-scale Agricultural AI Models
【Department of Civil Engineering / Ming-Der Yang / Tenured Distinguished Professor】
智慧服務：可大規模擴展之農業AI 模型【土木工程學系楊明德終身特聘教授】
上架日期2026-03-10
【Department of Civil Engineering / Ming-Der Yang / Tenured Distinguished Professor】
智慧服務：可大規模擴展之農業AI 模型【土木工程學系楊明德終身特聘教授】
| 論文篇名 | 英文：Enhanced pavement temperature prediction via a novel LSTM-GRU hybrid deep learning architecture with interpretability 中文：透過一種新型的LSTM-GRU混合深度學習架構增強路面溫度預測，並提高其可解釋性 |
| 期刊名稱 | Construction and Building Materials |
| 發表年份, 卷數, 起迄頁數 | 2026, 519, no.145838 |
| 作者 | Yared Bitew Kebede, Ming-Der Yang(楊明德)*, Chien-Wei Huang, Henok Desalegn Shikur |
| DOI | 10.1016/j.conbuildmat.2026.145838 |
| 中文摘要 | 精確預測瀝青路面溫度變化對於優化設計和養護策略至關重要，因為溫度變化對材料的剛性、強度和抗疲勞性能有著深遠的影響。然而，傳統的預測方法往往難以應對溫度曲線固有的時間依賴性複雜性。本研究提出了一種新型混合深度學習模型，該模型融合了長短期記憶網路（LSTM）和門控循環單元（GRU），利用即時氣象資料精確預測不同深度下瀝青路面的日最高和最低溫度。 LSTM-GRU架構透過協同結合LSTM對長期依賴性建模的能力和GRU對短期模式捕捉的效率，並克服了LSTM過擬合和GRU在處理複雜序列時複雜度不足等各自的缺陷，顯著提高了預測精度。此混合模型利用2018年5月至2023年3月的大量資料集進行了嚴格的訓練和驗證，結果始終優於一系列基準模型，在預測最高氣溫和最低氣溫方面均取得了卓越的預測精度，平均R²值分別為0.95和0.96，同時在MAE、MSE和RMSE方面也取得了顯著的改進。此外，SHapley Additive exPlanations (SHAP) 分析提供了重要的可解釋性，證實大氣溫度和地表溫度是最具影響力的預測因子，其次是風速、相對濕度和降水量。這些發現無可辯駁地證明了LSTM-GRU混合模型的卓越穩健性、準確性和可解釋性，為路面工程決策和主動管理策略提供了極具前景的先進解決方案。 |
| 英文摘要 | Precise prediction of asphalt pavement temperature variations is critical for optimizing design and maintenance strategies, given their profound influence on material stiffness, strength, and fatigue resistance. Traditional forecasting methods, however, often struggle with the inherent time-dependent complexity of these temperature profiles. This study proposes a novel hybrid deep learning model, integrating Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) networks, to precisely predict daily maximum and minimum asphalt pavement temperatures at various depths using real-time meteorological data. By synergistically combining LSTM's capacity for long-term dependency modeling with GRU's efficiency in capturing short-term patterns and mitigating individual drawbacks like LSTM's overfitting and GRU's reduced complexity for intricate sequences, the LSTM-GRU architecture significantly enhances predictive accuracy. Rigorously trained and validated using an extensive dataset from May 2018 to March 2023, the hybrid model consistently outperformed a comprehensive suite of baseline models, achieving superior predictive accuracy with average R² values of 0.95 for maximum temperatures and 0.96 for minimum temperatures, along with notable improvements across MAE, MSE, and RMSE. Furthermore, SHapley Additive exPlanations (SHAP) analysis provided critical interpretability, confirming atmospheric and surface temperatures as the most influential predictors, followed by wind speed, relative humidity, and precipitation. These findings unequivocally demonstrate the exceptional robustness, accuracy, and interpretability of the LSTM-GRU hybrid, offering a promising and advanced solution vital for informed pavement engineering decisions and proactive management strategies. |
| 發表成果與AI計畫研究主題相關性 | 目前發展的AI 模式以道路鋪面為溫度預測對象，可以應用在農地土壤分層溫度之預測。 |
