整合中長期氣象系集預報進行水庫供水系統動態脆弱度分析

Abstract

有效的水資源管理關鍵在於全面掌握供需關係，並進行精準的平衡與調配然而，隨著氣候變遷日益加劇，水資源短缺問題愈發嚴峻，特別是在農業領域，由於灌溉用水需求龐大且具高度時效性，因此對水資源調配的挑戰尤為突出。本研究提出一套綜合性框架，用於水資源之動態脆弱度分析，並結合長達6個月的長期氣象預報，以提升水資源管理之前瞻性與調適能力。分析上採用 K-近鄰演算法(K-nearest neighbors, KNN)方法進行空間降尺度，利用長短期記憶法(Long Short-Term Memory, LSTM)模擬降雨逕流，並應用蒙地卡羅模擬(Monte Calro Simulation, MCS)分析水資源系統之脆弱度。透過模型動態更新，並納入歷史水文與氣象資料中的不確定性因素，得以針對未來6個月的水資源供應系統之失效機率進行時序性分析與評估。 本研究以臺灣北部桃園地區之石門水庫供水區為研究區域，並針對歷史氣候條件與作物輪作情境進行分析。歷史條件模擬結果顯示，本模型成功預測 2000 至 2021 年間 7 次由乾旱引發之灌溉停水事件中的6次。透過動態分析，本模型可提供前瞻性之供水系統脆弱度預測，並顯示作物輪作策略能顯著提升灌溉穩定性，即便在乾旱年亦能維持供水系統之韌性。再考量作物適栽性、經濟效益及水資源利用率等3項重要因子，以非支配排序基因演算法(Non-dominated Sorting Genetic Algorithm, NSGA)建立多目標決策機制，提供以工作站為單位之作物種植建議。研究成果可協助管理者與農民更有效地掌握不確定性及脆弱點，並提供調適建議，進而促進水資源之科學化與適應性管理規劃。

Effective water resource management is based on the knowledge of water availability and requirement, as well as an interpretation of the balance between supply and demand. In recent years, the changing climate has caused high water supply pressure of many water resources systems, especially for decision-making of cases with high operational risks, such as agricultural irrigation supply, which has large demand, pressured time, and continuous impacts. Therefore, it is necessary to develop a dynamic analysis framework that can accommodate the medium and long-term meteorological ensemble forecast data (of the next 6 months). In this study, a systematic dynamic vulnerability analysis framework is proposed. In the framework, historical data of rainfall and reservoir inflow, operation, and discharge are incorporated with the meteorological long-term forecast data (TCWB1T1), in order to expand the explaining ability of historical datasets. The rainfall data was first downscaled using the method of K-nearest neighbors (KNN), and the rainfall-runoff process was simulated using the machine learning method-Long Short-Term Memory (LSTM). Then, the local reservoir operating rules were incorporated into the calculation of water supply and demand. Finally, the Monte Carlo simulation was applied to generate the dynamic vulnerability of the water supply system. By continuously updating the factors, the uncertainty was incorporated and the failure probability of the water supply system in the next 18 ten-days was analyzed, to shorten the time delay of generating decision-making assistance information. Taking Taoyuan City in northern Taiwan (the water supply area of Shihmen Reservoir) as an example, the results of model validation show that among the 7 drought-caused cessations of irrigation in this area in recent years, 6 of them can be clearly observed or even predicted. Through the ten-day analysis, dynamic vulnerability forecast can be provided at the key decision-making time (1-2 months before cultivation) in terms of the system failure rate. Considering 3 critical factors—crop suitability, economic profitability, and water-use efficiency—a multi-objective decision-making framework was established using the Non-dominated Sorting Genetic Algorithm (NSGA) to generate crop-planting recommendations at the workstation scale. The outcomes of this study assist managers and farmers in effectively identifying uncertainties and vulnerabilities, while providing adaptive strategies that facilitate more scientific and resilient water resource management planning.