基于采样贝叶斯方法的洪水概率预报研究  

Probabilistic flood forecasting by a sampling-based Bayesian model

作　　者：李明亮[1] 杨大文[1] 陈劲松 

机构地区：[1]清华大学水沙科学与水利水电工程国家重点实验室,北京100084 [2]美国劳伦斯伯克利国家实验室地球科学部

出　　处：《水力发电学报》2011年第3期27-33,共7页Journal of Hydroelectric Engineering

基　　金：十一五国家科技支持计划项目课题九(2008BAB29B09);国家自然科学基金项目(50928901)

摘　　要：基于贝叶斯理论建立了新安江模型参数和降雨观测相对误差的联合概率分布函数,利用自适应马尔可夫链蒙特卡洛方法DRAM(Delayed Rejection Adaptive Metropolis)对联合分布进行采样,得到模型参数和降雨误差的后验分布样本,二者相结合模拟得出洪水的概率预报结果。根据对江西滁洲流域32场洪水的模拟,采用交叉验证方法对概率预报结果进行了评估。根据连续概率排位分数CRPS(Continuous Ranked Probability score),相对于确定性预报方法概率洪水预报结果的改善达到26%。P值分位数-分位数图同样表明,概率预报方法对洪水的不确定性有较好的估计,但仍存在低估的现象。A Bayesian model is developed to estimate uncertainties in the parameters of Xin′anjiang model and the errors in rainfall measurement,and the Bayesian joint posterior distribution is sampled by the algorithm of delayed rejection adaptive Metropolis（DRAM）.The posterior distributions of the model parameters and the relative rainfall errors are used in the probabilistic forecast model proposed in this paper.The forecast results of this model are evaluated by applying a two-fold cross validation based on the data of 32 flood events in the Chuzhou catchment of Jiangxi province.In comparison with the deterministic forecast with optimal parameters,the model improves the continuous ranked probability score（CRPS） by 26%.A quantile-quantile plot of predictive cumulative distribution function of observed discharge versus standard uniform distribution shows a better estimate of discharge uncertainty by the proposed method,but still lower than the actual value.

关 键 词：防洪工程 洪水概率预报 贝叶斯方法 参数不确定性 输入不确定性 马尔可夫链蒙特卡洛 新安江模型 

分 类 号：P338.1[天文地球—水文科学]