摘要
松花江干流水位受上游来水和水库调蓄等多因素的共同作用而表现出非线性响应,采用典型的3层BP神经网络模型来模拟松花江干流肇源、哈尔滨、通河、佳木斯和富锦5个代表站点水位。鉴于BP神经网络学习收敛速度慢、参数选择困难、易陷入局部极小值等缺点,分别采用Levenberg-Marquart算法和基于遗传算法的BP算法来建立水位预报模型,并对预报结果进行了分析和比较。结果表明:两种算法收敛速度快,预报精度均能达到预报要求。特别是将遗传算法的全局搜索和BP网络局部精确搜索的特性有机结合,做到了优势互补,在河流水位预报方面有着广阔的应用前景。
Water‐level of mainstream of Songhuajiang River has a nonlinear response to the upper river discharges and the regulation and storage capacity of reservoir ,the three‐layer back propagation neural network has the ability to simulate water level for 5 representative hydrological observation stations in the mainstream of Songhuajiang Rvier ,including Zhaoyuan Station ,Harbin Station ,Tonghe Station ,Jiamusi Station ,and Fujin Station .Based on slow convergence of learning ,difficulty selecting of parameters and liability of dropping into local minimum for BP neural networks ,the outcomes of these models is analyzed and compared ,by establishing water‐level forecasting models based on the Levenberg‐Marquart algorithm and the genetic algorithm respectively . It show s both algorithms all have the fast convergence and excellent performance in forecasting water‐level .Especially for that the paper join the genetic algorithm w hich is good in global searching and BP algorithm w hich is effective on accurate local searching together and supplementing mutually , the method has a good application prospects . in water‐level forecasting field .
出处
《黑龙江大学工程学报》
2015年第4期6-11,共6页
Journal of Engineering of Heilongjiang University
基金
国家自然科学基金资助项目(41371047)
中国科学院战略性先导科技专项(XDA05110102)
水文水资源与水利工程科学国家重点实验室专项经费(1069-514031112)
