應用SWAT+模擬高屏溪及曾文水庫流域之水文收支

Abstract

臺灣用水需求節節攀升，可用水資源的調配取決於對流域水文收支之瞭解，然而各流域的水文收支情形迥異，往往需要水文模式的協助才得以獲得全貌。本研究以高屏溪上游的荖濃溪、濁口溪，及曾文溪上游的曾文水庫流域為研究區域，分別以水利署和氣象署測站雨量資料、TCCIP網格雨量資料，搭配SWAT+模式(Soil and Water Assessment Tool+)來模擬其水文收支，應用非支配排序基因演算法(Non-dominated Sorting Genetic Algorithm II, NSGA-II)，透過模式參數的調整，同時優化NSE、logNSE，衡量此模式的模擬結果在高、低流量事件之準確性，找到一組帕累托最優解(Pareto optimal)。研究成果顯示，在NSGA-II檢定後，曾文水庫集水區2005～2022年測站雨量之流量的NSE達0.55~0.58、logNSE達0.74~0.75；網格雨量之流量的NSE達0.2~0.62、logNSE達0.28~0.69。高屏溪流域除了雨流關係較差的兩個年段之外，濁口溪集水區2006～2008年測站雨量之流量的NSE達0.62~0.63、logNSE達0.57~0.64；網格雨量之流量的NSE達0.31~0.33、logNSE達0.56~0.58。而荖濃溪集水區1983～2004年測站雨量之流量的NSE達0.54、logNSE達0.48~0.5；網格雨量之流量的NSE達0.56~0.62、logNSE達0.42~0.47。測站雨量的流量模擬結果在高、低流量事件均優於網格雨量的流量模擬結果。在參數的部分，整體而言檢定參數的選擇對於不同雨量資料之模擬結果較有限，其中最敏感的參數是ESCO、CN2、AWC，三者都是影響土壤層以上水文流動之參數，而最不敏感的參數則是SURLAG，可能反映臺灣的集水區面積較小，地表逕流較少保存在地表逕流儲存庫，故其對於流量模擬的影響力較低。在不同降水資料中，網格雨量之年均降水量比測站雨量更少，使得蒸發散比例明顯較高，為了更好地模擬流量，初始土壤含水量、含水層補注及基流之比例都低於測站雨量，且河道較不易下滲（CHK），顯示出不同雨量來源輸入下對模擬水文收支之影響。曾文溪流域CN值比高屏溪較大，故降水多轉為地表逕流進入河道，且能從更深層之土壤提取蒸發水分（ESCO），導致初始、最終土壤含水量之比例較低，再加上曾文溪流域的土壤水滲透到含水層的時間較長（AQUIFER_DELAY），且地下水向上或下層移動之門檻較低（REVAP_MIN），使其淺層含水層補給略少，導致雖然從參數來看淺層向深層含水層滲漏比例較高（PERCC_LTE），但實際的深層含水層滲漏流出之比例卻略低。在基流方面，雖然基流反應較快（ALPHA），但因為地下水轉變為基流之門檻較高（FLO_MIN），故基流比例也較低。透過本研究除了能解析高屏溪及曾水文庫流域之水文收支之外，也將有助於了解用SWAT+建立臺灣在地化之流量模擬的程序。

The increasing demand for water in Taiwan underscores the importance of understanding watershed water balance for optimal allocation of available resources. Given the diverse hydrological conditions across watersheds, hydrological modeling is often essential for comprehensive assessment. This study focuses on the Laonong and Zhuokou Rivers in the upper Gaoping River Basin, as well as the Zengwen Reservoir watershed in the upper Zengwen River Basin. Using rainfall data from Water Resources Agency and Central Weather Administration gauges, as well as gridded rainfall data from the Taiwan Climate Change Projection Information and Adaptation Platform (TCCIP), the Soil and Water Assessment Tool Plus (SWAT+) model was applied to simulate watershed water balance. The Non-dominated Sorting Genetic Algorithm II (NSGA-II) was employed to calibrate model parameters, simultaneously optimizing the Nash–Sutcliffe Efficiency (NSE) and log-transformed NSE (logNSE) to evaluate simulation accuracy for both high- and low-flow events, yielding a set of Pareto optimal solutions. The results show that, after NSGA-II calibration, simulated streamflow in the Zengwen Reservoir watershed (2005–2022) based on gauge rainfall achieved NSE values of 0.55–0.58 and logNSE values of 0.74–0.75, while simulations based on gridded rainfall achieved NSE values of 0.20–0.62 and logNSE values of 0.28–0.69. In the Gaoping River Basin, excluding two periods with poor rainfall–runoff relationships, streamflow in the Zhuokou River watershed (2006–2008) achieved NSE values of 0.62–0.63 and logNSE values of 0.57–0.64 for gauge rainfall, compared to NSE values of 0.31–0.33 and logNSE values of 0.56–0.58 for gridded rainfall. In the Laonong River watershed (1983–2004), gauge rainfall simulations achieved NSE of 0.54 and logNSE of 0.48–0.50, while gridded rainfall simulations achieved NSE of 0.56–0.62 and logNSE of 0.42–0.47. Overall, gauge rainfall–based simulations outperformed gridded rainfall–based simulations for both high- and low-flow events. In terms of parameter sensitivity, the choice of calibrated parameters had limited impact across different rainfall datasets. The most sensitive parameters were ESCO, CN2, and AWC—variables influencing hydrological processes above the soil layer—while SURLAG was least sensitive, likely reflecting the relatively small catchment sizes in Taiwan, where less surface runoff is retained in the surface runoff storage. Gridded rainfall generally exhibited lower annual precipitation than gauge rainfall, resulting in higher evapotranspiration ratios. To better match observed streamflow, simulations with gridded rainfall required lower initial soil water content, reduced aquifer recharge, smaller baseflow contributions, and reduced channel infiltration (CHK). The Zengwen River Basin exhibited higher CN values than the Gaoping River Basin, indicating greater conversion of rainfall to surface runoff, greater extraction of evaporative water from deeper soil layers (higher ESCO), lower initial and final soil water content, longer percolation time to the aquifer (AQUIFER_DELAY), and lower thresholds for vertical groundwater movement (REVAP_MIN). These factors led to slightly reduced shallow aquifer recharge and, despite higher percolation potential (PERCC_LTE), slightly lower actual deep aquifer outflow. Although baseflow response was quicker (higher ALPHA), a higher threshold for groundwater conversion to baseflow (FLO_MIN) resulted in lower baseflow proportions. This study not only elucidates the hydrological balance of the Gaoping and Zengwen River Basins but also contributes to the development of a localized SWAT+ streamflow simulation framework for Taiwan.