摘要
传统BP网络需要预先设定网络隐含层的层数和每层的节点数,使得在预测过程中难以确定网络的最优结构。与之相反,梯级-关联算法(CC)要求初始网络仅含有输入层和输出层,通过运算不断向网络增加隐含节点。在介绍梯级-关联算法原理的基础上,分别运用梯级-关联算法和BP算法对拉萨河拉萨站的月流量进行了预测,结果显示:在不损失预测精度的前提下,梯级-关联算法的运算次数仅为5次,而BP算法则需要运算70 000次,运算效率有很大的提高,同时网络的规模也有所减小。
The size and structure of neutral networks must be predefined if the standard backpropagation neural network architecture is used for training. On the contrary, the initial network of cascade-correlation （CC） is composed of the input layer and output layer while the hidden unit is inserted into the network one by one. The principle of CC is presented in this paper, and the monthly streamflow in the Lasa river is forecasted by using the CC and the BP models. The result shows that the CC model only needs to run five times while the BP model needs 70000 times to reach the same precision. The efficiency of the CC model is much higher than that of the BP model. Another conclusion is that the network size of a CC model is smaller than that of the BP model.
出处
《水科学进展》
EI
CAS
CSCD
北大核心
2007年第1期114-117,共4页
Advances in Water Science
基金
国家科技攻关计划资助项目(2005BA901A11)
北京师范大学"京师学者"特聘教授启动经费资助项目
关键词
梯级-关联算法
BP算法
流量预报
拉萨河
cascade-correlation algorithm
backpropagation algorithm
streamflow forecasting
Lasa River
