目标函数权重对PEST-HSPF水文率定的影响研究

参数率定是水文模型构建与应用的重要基础。在多目标水文校准中,不同目标函数的权重设置直接影响校准的结果,如何确定不同目标函数的权重是水文校准的关键问题。选择典型半分布式水文模型HSPF(Hydrological Simulation Program-Fortran),基于PEST多目标校准模型,以牛栏江上游流域为研究对象,研究目标函数权重对HSPF水文模拟拟合优度的影响,为HSPF模型参数自动率定提供指导与借鉴。结果表明:(1)当PEST-HSPF水文校准的单目标函数权重(日流量、月流量、超流天数)上升时,模型的纳什系数和相对偏差呈现非线性变化特征,不规则波动幅度较大;(2)3个目标函数的权重在1~10量级内模型能够获得较高的预测能力和较低的误差,模型的纳什系数平均可达到0.8以上,相对偏差在10%以内;(3)超流量天数权重设置对模型的预测能力变化影响较大,日流量天数权重对相对偏差的波动影响较大,当超流天数的相对权重在1~1 000范围内变化时,模型纳什系数在0.02~0.86之间波动剧烈,当日流量相对权重值超过30时,相对偏差变化明显。

Impact of Objective Function Weight on Hydrological Calibration of PEST-HSPF

Parameter calibration is essential for establishing and using a hydrological model. In multi-objectives hydrological calibra - tion, different weights of objective functions directly affect the calibration results, and how to determine the weights of different ob - jective functions is a key step of hydrological calibration. The HSPF (Hydrological Simulation Program-Fortran) model that is a typ - ical semi-distributed hydrological model was taken to study the effect of objective weights for different calibration functions on hy - drological simulation. The upstream watershed of the Niulanjiang River was selected as a case study to simulate runoff by the PEST-HSPF model, and the study was supposed to be a guide for parameter optimization of HSPF by automatic calibration approach. Through altering the objective function weights, the goodness of fit of the model was recorded to determine the impact of different weights on the performance of PEST-HSPF simulation. The results indicate that: (1) when the weight of single objective function (daily flow, month flow, hypercritical flow days) of PEST-HSPF rises, Nash coefficient of the model and the relative devia - tion present nonlinear variation response with a vast and unstable variation in value; (2) Higher prediction ability and less error can be obtained when the three weights of the objective functions are set within 1-10 orders of magnitude. The Nash coefficient can reach over 0.8 and the relative error is lower than 10%; (3)The weight of hypercritical flow days has a greater influence on the a - bility to prediction while the weight of daily flow has a greater influence on the fluctuation of relative deviation. When the weight of hypercritical flow days varies between 1 and 1000, the corresponding Nash coefficient is between 0.02 and 0.86; the relative error will change greatly when the weight of daily flow exceeds 30.