動態貝氏網路於缺水風險評估之探討

Abstract

長期降雨預報之高度不確定性使得水資源管理面臨挑戰，尤其於乾旱期間，決策者需要預見水資源需求和供應之間的潛在差異，以降低缺水風險，但其存在著不確定性且隨眾多因素而變化。面對此不確定性，須有別於傳統風險評估之抗旱決策輔助系統，應改以機率風險評估(PRA, Probabilistic Risk Assessment)概念建之，以序率方式推估未來情境之缺水風險，並在此風險下，提供可能之抗旱決策，進而提前部署。 本研究將嘗試建立水情抗旱決策輔助之動態貝氏網絡(DBN, Dynamic Bayesian network)機率模型，以旬尺度條件機率的方式連結石門水庫集水區降雨量、長期天氣預報、流量預報、河川流量、入庫流量、水庫蓄水量、供水系統、各區域計畫需水量、水庫操作、抗旱措施等，接著使用馬可夫鏈蒙地卡羅(MCMC, Markov Chain Monte Carlo)抽樣法，推估各節點之後驗機率，進而判斷缺水之可能性及不確定性，最後進行貝氏成本效益之風險分析，以提供抗旱決策之可行方案。 本研究之動態貝氏網路推估發現，針對農業供水部份，不採納第三階段限水、停灌，仍可滿足經濟部水利署之抗旱目標，並保留兩個月安全水量進行管控。2015年31旬後採取10旬二階限水、4旬一階限水，可取代7旬三階限水；2017年31旬後採取9旬二階限水、2旬一階限水，可取代14旬三階限水。動態貝氏網路於供水系統，其可依推估之各節點機率密度分佈，進行最佳抗旱決策分析，有機會可以作為抗旱應變決策輔助系統(DSS, Decision Support Systems)之工具。

The high uncertainty of long-term precipitation forecasting introduces a significant challenge in water resources management, especially during the drought period. During the dry season, the decision makers need to foresee the potential discrepancy between the demand and supply of the water resources, which are both uncertain and changing in according to multiple conditions, in order to avoid the potential water shortage risk. This study proposes the water resources management framework against drought based upon Bayesian network model. This model with conditional probability to link rainfall, stream flow, reservoir inflow, reservoir level, water supply system, water planning demand from sectors, operation rules and policies against drought. Then, using Markov chain Monte Carlo (MCMC) sampling to approach posterior probability of each node and analyzing probability and intrinsic uncertainty of water shortage node. Finally, taking Bayesian and cost benefit risk analysis to provide feasible drought decision making. The dynamic Bayesian network (DBN) estimation of this research found that for agricultural water use, the third-stage water restriction and fallow is not executed, which can still meet the target of drought policies of the Water Resources Department of the Ministry of Economic Affairs, and retain two months of safe water volume for control. In 2015, the third-order water restriction can be replaced by the second-order water restriction in ten times and the first-order water restriction in four times can replace In 2017, the third-order water restriction can be replaced by the second-order water restriction in nine times and the first-order water restriction in two times. It is a potential tool to decision support systems (DSS) for drought.