發展防災與抗旱導向的人工智慧水庫入流量預報

Abstract

氣候變遷導致極端氣候頻率增加，使得水庫入流量的準確預報成為水資源管理的核心挑戰。本研究針對短期與長期流量預報模式進行深入研究，提出創新的預測框架，以提升水庫操作及用水策略的科學依據。 在短期預報部分，本研究針對傳統簡單平均法(Ensemble Mean, EM)的局限性，提出交替預報模式(Switch Prediction Method, SPM)，整合中央氣象署的系集降雨產品，包括WEPSPro_PM、CWBWRF_W0、ETQPF、QPF與STMAS_WRF，並與支援向量機(Support Vector Machine, SVM)和多步階預報(Multi-step Forecasting, MSF) 結合，實現未來72小時逐時入流量的準確預測。研究以石門水庫為案例，對2004年至2019年間的18場颱風進行測試，結果顯示，SPM在整合降雨資料及水庫入流量預測的表現均優於EM，不僅能減少誤差，亦能提供兼具定率性與機率性的預報結果，為水庫洩洪與蓄洪決策提供強有力的技術支持。 在長期預報部分，本研究整合中央氣象署第一代海氣耦合模式 (TCWB1T1)，並應用K-最近鄰居法 (K-Nearest Neighbor, KNN) 將大尺度雨量及溫度資料降至集水區尺度，進一步結合六種機器學習與深度學習法，包括支援向量機、多層感知器、深度神經網路、遞迴神經網路、長短期記憶及門控遞迴單元，對臺灣中南部六大集水區未來180日入流量進行預測。結果顯示，各研究區域於枯水期與豐水期的預報模式以深度學習法為主，尤其以深度神經網路(DNN)的表現最佳，能有效掌握流量趨勢，並提供穩定準確的預報結果。 本研究所提出的短期與長期流量預報模式，分別針對突發性氣候事件與長期水文變化設計，既能快速反應颱風等極端事件，又能為乾旱期的用水策略提供科學依據。透過提升預報準確度及降低不確定性，本研究成果將對水庫管理及水資源調配提供強有力的支持。

Climate change has led to an increase in the frequency of extreme weather events, which makes accurate reservoir inflow forecasting a critical challenge in water resources management. This study focuses on both short-term and long-term inflow forecasting models, and innovative frameworks to enhance scientific support for reservoir operations and water usage strategies are proposed. For the short-term forecasting, this study addresses the limitations of traditional ensemble mean (EM) methods by introducing the Switch Prediction Method (SPM). The SPM integrates ensemble rainfall products from the Central Weather Bureau, including WEPSPro_PM, CWBWRF_W0, ETQPF, QPF, and STMAS_WRF. Combined with Support Vector Machine (SVM) and Multi-step Forecasting (MSF), the proposed framework accurately predicts reservoir inflows for the next 72 hours on an hourly basis. Using Shimen Reservoir as a case study, 18 typhoon events from 2004 to 2019 are analyzed. The results demonstrate that the SPM outperforms the EM in both rainfall data integration and reservoir inflow prediction; the SPM effectively reduces errors while providing both deterministic and probabilistic forecasts. This approach offers robust technical support for decision-making in reservoir flood and storage management. For the long-term forecasting, this study leverages the first-generation coupled ocean-atmosphere model (TCWB1T1) developed by the Central Weather Bureau and applies the K-Nearest Neighbor (KNN) method to downscale large-scale rainfall and temperature data to the watershed scale. Additionally, six machine learning and deep learning methods are employed, including Support Vector Machine, Multi-layer Perceptron, Deep Neural Network (DNN), Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU), to predict inflows for the next 180 days across six major watersheds in southern Taiwan. Results indicate that deep learning methods, particularly DNN, outperform machine learning methods in capturing inflow trends during both dry and wet seasons. The deep learning methods provide stable and accurate long-term forecasts. The proposed short-term and long-term inflow forecasting models are tailored to address both sudden climatic events and long-term hydrological changes. These models offer rapid responses to extreme events such as typhoons while providing scientific foundations for water usage strategies during droughts. By improving forecast accuracy and reducing uncertainty, the findings of this study provide strong support for reservoir managements and water resources allocation.