摘要
针对以往递归神经网络(RNN)训练算法难,连续搅拌反应釜(CSTR)的强非线性等问题,将回声状态网络(echo state network,ESN)方法应用于模型不确定的CSTR系统辨识中.ESN具有较强的非线性逼近能力和良好的短期记忆能力,且只需要训练网络输出权值,简化了网络训练算法.仿真结果表明,在相同条件下,与带动量的BP(back propagation)神经网络、BP-MLP(back propagation multilayer perceptron)神经网络、最小二乘支持向量机(LS-SVM)、模糊神经网络(FNN)、GAP-RBF神经网络、MGAP-RBF神经网络相比,ESN能给出相当好的性能,表现出较高的辨识精度,ESN比带动量的BP、BP-MLP、LS-SVM神经网络的逼近精度提高了4个数量级,表明了该方法的有效性.
For the difficulties in the algorithms＇ training using the existing recurrent neural network（ RNN） and the high nonlinearity in the CSTR（ continuous stirred tank reactor） system,an echo state network（ ESN） is applied to the identification of the CSTR system with model uncertainty.ESN has strong nonlinear approximation ability and good short-term memory capacity.Futhermore,only the output weights need to be trained,so the network training algorithm is simplified.Simulation results reveal that the proposed method provides considerably better performance and accuracy of identification than methods based on any of the following under same conditions： BP neural networks,BP-MLP neural network,least squares support vector machine（ LS-SVM）,fuzzy neural network（ FNN）,GAP-RBF neural network,and MGAP-RBF neural network.Furthermore,the approximation accuracy of the method based on ESN is four times more accurate than the BP,BP-MLP,and LS-SVM.The simulation results verify the effectiveness of the method.
出处
《信息与控制》
CSCD
北大核心
2014年第2期223-228,共6页
Information and Control
基金
甘肃省财政厅基本业务费项目(620026)
甘肃省教育厅硕导项目(1104-09)
