军事知识服务平台
基于不同边缘分布的BMA集合预报比较研究
Comparative Study of BMA Ensemble Forecasts Based on Different Distribution Assumptions
摘要 边缘分布函数的选取是贝叶斯模型平均法能否准确量化水文预报不确定性的重要因素,不恰当的分布假设会导致概率预报区间出现偏差,降低预报的可靠性.为评估不同边缘分布函数对贝叶斯模型平均集合预报效果的影响,选取紫荆关、羊坊闸和屯溪3个具有不同水文气象特征的典型流域,构建了基于Weibull分布、Gamma分布、Box-Cox变换与对数正态变换的贝叶斯模型平均集合预报模型,系统比较各分布在不同流域中的预报表现.结果表明,在洪峰流量较大的流域,Weibull与Gamma分布在确定性预报中具有较强的拟合能力,如紫荆关流域测试集RMSE分别为124.17和140.44,NSE分别为0.74和0.67;而Box-Cox变换与对数正态变换的RMSE分别为142.84和165.91,NSE分别为0.65和0.53.在洪峰流量较小且低流量主导的流域,Box-Cox与对数正态变换则展现出更好的稳健性,而Weibull和Gamma分布因其概率密度函数的左尾特征,参数估计过程易被低值样本主导,导致概率预报结果中5%分位数长期趋近于零且缺乏波动,无法反映真实的基流波动.
Abstract:
The choice of marginal distribution functions is an important factor in whether the Bayesian Model Averaging(BMA)method can accurately quantify hydrological forecast uncertainty.Inappropriate distribution assumptions can lead to biases in probabilistic forecast intervals,reducing the reliability of forecasts.To evaluate the impact of different marginal distribution functions on the performance of BMA ensemble forecasts,three typical basins with different hydrological and meteorological characteristics—Zijingguan,Yangfangzha,and Tunxi—were selected.BMA ensemble forecast models were constructed based on the Weibull distribution,Gamma distribution,Box-Cox transformation,and log-normal transformation,and the forecast performance of each distribution in different basins was systematically compared.The results show that in basins with large flood peaks,the Weibull and Gamma distributions have stronger fitting ability in deterministic forecasts.For example,in the test set of Zijingguan Basin,the root mean square error(RMSE)are 124.17 and 140.44,respectively,and the Nash-Sutcliffe efficiency(NSE)are 0.74 and 0.67.In contrast,the RMSE of Box-Cox and log-normal transformations are 142.84 and 165.91,with NSE of 0.65 and 0.53.In basins with smaller flood peaks dominated by low flows,Box-Cox and log-normal transformations demonstrate better robustness.However,for the Weibull and Gamma distributions,due to the left-tail characteristics of their probability density functions,the parameter estimation process is easily dominated by low-value samples,causing the 5th percentile in probabilistic forecasts to long-term approach zero and lack fluctuation,failing to reflect the true baseflow variation.
刊 名 中国农村水利水电 ISTICPKU
年,卷(期): 2026, (10)
分类号: P338
机标分类号: P338 P457.6 O212.8
在线出版日期: 2026年11月2日
基金项目: 2024YFC3211400:国家重点研发计划B240203007:中央高校基本科研业务费项目2025-18:河北省省级水利科技计划项目,