深度遷移學習於水庫入流量預報之研究

Abstract

臺灣位於西太平洋颱風路徑要衝，夏秋兩季常受到颱風侵襲，近10年來平均每年約遭受2.3個颱風侵襲，颱風伴隨來的豐沛雨量，短時間在集水區產生大量逕流，增加管理機關操作失誤的風險，並提高水庫防洪運轉策略研擬困難度。近年來興起以人工智慧 (Artificial Intelligence, AI) 方法建立入庫流量預報，可有效模擬複雜的水文系統，然而資料的數量和質量與深度學習入庫流量預報模式準確度高度相關，當集水區資料相對稀少則無法準確預報，容易產生過度擬合，有效的解決辦法則是增加資料量和簡化模式複雜度，因此本研究於國內首次應用深度遷移學習技術於水文領域，有效利用水文特性相近的流域資料，增加模式訓練資料，提升模式預報的準確度。 爰此，本研究以北臺灣石門水庫和翡翠水庫作為研究區域，蒐集2004至2023年颱風事件之觀測雨量及流量，提出四種深度學習方法，如：深度神經網路(Deep Neural Network, DNN)、遞迴神經網路(Recurrent Neural Network, RNN)、長短期記憶網路(Long short-term memory, LSTM)、門閘遞迴單元(Gate Recurrent Unit, GRU)，結合遷移學習(Transfer Weight, TL)和遷移學習搭配凍結層數技術(Transfer Weight with Frozen layers, TLFL)來建立入庫流量預報。為了比較深度遷移學習模式在不同訓練樣本條件下的訓練效果和模式預報準確性，本研究設定「完整資料情境」和「資料稀少的情境」兩種情境，並將「資料稀少的情境」的模式訓練颱風場次減半，首先以網格搜尋法篩選出輸入因子步長、模式架構及學習率，先使用觀測雨量並利用多步階預報建立未來1至6小時入流量預報，根據評鑑指標優選最佳之深度學習模式後，再介接交替預報(Switch Prediction Method, SPM) 降雨預報，建立未來1至72小時入流量預報，將模式結果根據評鑑指標及流量歷線圖進行比較。 結果顯示在未來1小時預報中，完整資料情境下翡翠水庫以DNN表現最佳，石門水庫以LSTM表現最為優異，相關係數(Correlation Coefficient, CC)皆高達0.96以上，而在資料稀少的情境下，評鑑指標表現皆顯示模式表現略遜於完整資料情境。在完整資料情境下未來1至6小時水庫入流量預報的多步階驗證中，遷移學習能提升深度學習模式表現，有效降低模式誤差，翡翠水庫最優t+6模式的RMSE改善幅度為8%；石門水庫最優t+6模式的RMSE改善幅度則為3%，各模式效率係數(Coefficient of Efficiency, CE)皆高達0.93以上；而在資料稀少的情境下，評鑑指標同樣顯示結合遷移學習的模式表現較佳。在未來1至72小時水庫入流量預報中，遷移學習在流量較大的颱風場次(梅姬、蘇迪勒、瑪莉亞)中相較於原始模式預報流量準確度較高，而大部分流量較小的颱風場次則無明顯進步，受降雨預報影響較大。 在遷移學習效益評估中，本研究分別就「資料情境」、「遷移學習方法」、「水庫集水區特性」三個面向進行探討。結果顯示遷移學習在「資料稀少情境」下改善幅度較大，顯示遷移學習在資料稀少的集水區中，能提升其水庫預報模式的泛化能力；而本研究提出之TL或是TLFL方法皆可有效轉移源域的模式特徵到目標域，降低入庫流量預報誤差；在兩水庫互相遷移的結果顯示，石門水庫遷移至翡翠水庫能有效降低翡翠水庫模式的預報誤差，使模式學習大集水區的水文特性，顯示遷移學習由較大集水區遷移至較小集水區的效果較佳，反之則改善幅度較小。 本研究經短延時及長延時預報結果皆顯示，結合適當的遷移學習方法(TL或TLFL)能提升水庫入流量的預報準確度，同時在資料稀少的情境下也能有效降低模式的誤差，顯示遷移學習可將水文特性從數據豐富的集水區轉移到數據稀少的集水區，提升水庫入流量預報模式的表現，可有效增加深度學習模式的訓練量並降低模式誤差，可供未來水庫相關單位擬訂水庫操作策略之重要依據，作為水利防救災預警工作時之重要參考。

Taiwan is located on one of the main paths of northwestern Pacific typhoons. According to the Central Weather Administration, on average 2.3 typhoons per year strike Taiwan. Since typhoons bring heavy rainfall and severe floods in the watersheds, an accurate forecast of reservoir inflow is needed to reduce the difficulty of operating strategies in flood warning systems. Deep learning techniques have been demonstrated as effective methods to forecast reservoir inflow. However, forecasts in some reservoir watersheds are limited due to the lack of observational data during typhoon events. Hence, novel techniques for forecasting reservoir inflow are proposed to address the challenge posed by the scarcity of data. Transfer Learning (TL) and Transfer Learning with Frozen Layer (TLFL) are used to improve the performance of reservoir inflow forecasting in full-data and sparse-data reservoir watersheds through parameter transferring and fine-tuning. Four deep learning models, including Deep Neural Network (DNN), Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), and Gate Recurrent Unit (GRU), are adopted to fit a sufficiently large source domain dataset and repurpose the learned weights to the model of the target domain. The four models were combined with multi-step forecasting to forecast the reservoir inflow in the next 1 to 72 hours with ensemble rainfall forecasting. The results showed that transfer learning provides significant benefits to target reservoirs in both full data and sparse data scenarios. Compared with models without transfer learning, the optimal model with TL and TLFL for 6-h lead time showed significantly improved performance, where RMSE was reduced by 3% in Feitsui Reservoir and 8% in Shihmen Reservoir. Both models exhibited excellent performance, with coefficient of efficiency (CE) and correlation coefficient (CC) values consistently exceeding 0.93. In reservoir inflow forecasting for the next 1 to 72 hours, models with transfer learning generated more accurate and stable forecasting results during typhoons with larger inflows (e.g., Typhoons Megi, Soudelor, and Maria) while no significant improvement had been observed for smaller typhoons. The comparison of mutual transfer outcomes between Feitsui Reservoir and Shihmen Reservoir indicated that the integration of transfer learning methods, such as TL or TLFL, significantly enhanced the accuracy of reservoir inflow predictions and reduced model errors effectively. The proposed method is expected to provide more accurate forecasts of reservoir inflow for disaster prevention in practical applications.