中文摘要：

基于GEVcdn模型（generalized extreme value conditional density network）构建了珠江流域非一致性洪水频率模型。结果表明：多数站点呈非一致性变化特征，主要分布在西江和东江；在某些传统洪涝灾害易发区域，洪水极值序列呈上升趋势，将面临更加严峻的防洪风险。珠江流域洪水集中发生在两个时段，即1960年代和1990年代中期以后；基于非一致性算法的重现期体现出与一致性算法重现期的明显区别；珠江流域主要堤围基于非一致性算法的失败风险明显高于一致性条件下的失败风险而显示出较大的隐患。

英文摘要：

This paper establishes inconsistent flood frequency model of the Pearl River Basin based on GEV cdn （ generalized extreme value conditional density network） technique. The results indicate that most of station prestent inconsistent variation characteristic, it is mainly distributed in Xijiang River and Dongjiang River, and in some tradi- tional flood easily happening areas the flood extreme presents a rising tendency and the flood preventing risk will be more rigorous ;the flood in Pearl River Basin happened mainly in two periods,i, e. 1960＇ and after middle stage of 1990＇ ;the return period based on nonconsistent algorithm reflects obvious differentiation with that based on consist- ent algorithm, the failure risk of main dyke-dam on Pearl River Basin based on nonconsistent algorithm is obviously higher than that from consistent condition and exists more hidden troubles.