基于深度神经网络的水文频率分析  

Hydrological Frequency Analysis Based on Deep Neural Networks

作　　者：周楚天 刘攀[1] ZHOU Chutian;LIU Pan(State Key Laboratory of Water Resources and Hydropower Engineering Science,Wuhan University,Wuhan 430072,China)

机构地区：[1]武汉大学水资源与水电工程科学国家重点实验室,湖北武汉430072

出　　处：《水文》2022年第6期1-6,共6页Journal of China Hydrology

基　　金：国家自然科学基金联合基金(U1865201);国家自然科学基金(51861125102);科技部重点领域创新团队(2018RA4014)。

摘　　要：通过给定水文频率分布参数值,随机生成皮尔逊Ⅲ型分布序列,将其作为输入,分别以分布参数或给定设计频率下的设计值作为输出,构建了两种深度神经网络模型:一种可估计水文频率分布参数,另一种可直接估计给定频率下的设计值。分别以序列长为30、50以及100的三组测试集开展实例研究,结果表明:直接估计给定频率下设计值的深度神经网络优于估计分布参数的网络;其设计值估计结果的平均相对误差分别为-0.003、-0.007和0.004,而线性矩法为0.036、0.016和0.010,无偏性更优且有效性全面优于线性矩法,并在分布参数偏大或偏小的情况下表现更加稳定。This paper generated random data of PersonⅢdistribution by given hydrological frequency distribution parameter values,which was inputted into models.Then two deep neural networks models were built,one can output distribution parameters and another one can output specific design values directly.Study is conducted on three sets with sequence length of 30,50 and 100,and the results show that:The deep neural networks that directly estimate the design values at a given frequency are better than networks that estimate distribution parameters;The relative mean relative biases of the design values estimation are-0.003,-0.007,and 0.004 respectively,while they are 0.036,0.016 and 0.010 in L-moments model.It indicates that deep neural networks method performs better than L-moments method in unbiasedness,and the former’s performance is more stable when distribution’s parameters are too large or too small.At the same time,deep neural networks method’s effectiveness significantly better than L-moments method.

关 键 词：水文频率分析 深度神经网络 线性矩法 皮尔逊Ⅲ型分布 

分 类 号：P333.2[天文地球—水文科学]