188宝金博页面版

第 41 卷第 4 期2013 年 7 月DOI:10. 3876/ j. issn. 1000 1980. 2013. 04. 003河 海 大 学 学 报 ( 自 然 科 学 版 )Journal of Hohai University(Natural Sciences)Vol. 41 No. 4Jul. 2013摇 摇 收稿日期: 2012 05 31基金项目: 国家自然科学重点基金(41130639);江苏省普通高校研究生科研创新计划(CXZZ11-0435)作者简介: 阚光远(1985—),男,天津人,博士研究生,主要从事水文学及水资源研究。 E鄄mail: kanguangyuan@126. com改进的神经网络模型在水文模拟中的应用阚光远1,李致家1,刘志雨2, 李巧玲1, 胡友兵1(1. 河海大学水文水资源学院,江苏 南京摇 210098; 2. 水利部水文局,北京摇 100053)摘要: 将 BP 神经网络与 K鄄最近邻(KNN)算法耦合起来,建立 BK(BP鄄KNN)模型,该模型以前期模拟流量和相应影响要素作为 BP 神经网络的输入,出口断面流量作为网络输出,对产汇流过程进行模拟;采用 K鄄最近邻算法,基于历史样本的模拟误差和相应影响要素对网络输出进行修正,实现了非实时校正模式下的连续模拟。 根据 BK 模型的计算流程将其参数分为 3 个层次,各层次分别使用 NSGA鄄域多目标优化算法进行参数优选,提高了模拟精度、优化效率和网络泛化能力。 分别将新安江模型的产流、产流分水源计算模块与 BK 模型相耦合,建立 XBK (Xinanjiang runoff production鄄BK) 和 XSBK (Xinanjiang runoff production and separation鄄BK)模型,在呈村等 3 个不同类型的流域应用新安江模型、BK 模型、XBK 模型和 XSBK 模型进行模拟精度比较,结果表明改进的模型模拟精度更高,较好地解决了神经网络模型在水文模拟中存在的问题。关键词: 水文模型;BP 神经网络;K鄄最近邻算法;新安江模型;NSGA鄄域算法;呈村流域;东湾流域;大阁流域中图分类号:P338摇 摇 摇 文献标志码:A摇 摇 摇 文章编号:1000 1980(2013)04 0294 06An improved neural network model and its application tohydrological simulationKAN Guangyuan1, LI Zhijia1, LIU Zhiyu2, LI Qiaoling1, HU Youbing1(1. College of Hydrology and Water Resources, Hohai University, Nanjing 210098, China;2. Bureau of Hydrology, Ministry of Water Resources, Beijing 100053, China)Abstract: The BK (BP鄄KNN) model, which is coupled by the BP neural network model and the K鄄nearestneighbor (KNN) model, was established. This model was used to simulate flow concentration and runoff generationwith antecedent simulated outlet flow and relevant influencing factors as inputs of the BP neural network, and theoutlet flow as the output of the network. The output of the network was corrected by the K鄄nearest neighboralgorithm based on historical samples爷 simulation error and relevant influencing factors. The KNN algorithm realizescontinuous simulation in a non鄄real time mode. According to the computational procedure, parameters of the BKmodel were divided into three groups and optimized by the NSGA鄄II multi鄄object algorithm in each group. This wayof calibration improves the accuracy, efficiency, and generalization ability of the BK model.(Xin爷anjiang runoff production鄄BK) and the XSBK (Xin爷anjiang runoff production and separation鄄BK) modelwere coupled by the Xin爷 anjiang runoff production module, the Xin爷 anjiang runoff production and separationmodule, and the BK flow concentration module. BK, XBK, XSBK, and the Xin爷anjiang model were applied to theChengcun, Dongwan, and Dage watersheds to compare their simulation accuracies. The results show that theimproved model has higher accuracy and can be used to solve problems of neural network models duringhydrological simulation.Key words: hydrological model; BP neural network; K鄄nearest neighbor algorithm; Xin爷anjiang model; NSGA鄄域algorithm; Chengcun watershed; Dongwan watershed; Dage watershedThe XBK