基于L2正则化LSTM的非线性动态系统辨识  

Identification of Nonlinear Dynamic System Based on L2 Regularized LSTM

作　　者：徐宝昌[1] 吕爽 董秀娟 王健 XU Bao-chang;LV Shuang;DONG Xiu-juan;WANG Jian(College of Information Science and Engineering,China University of Petroleum(Beijing);PetroChina Beijing Gas Pipeline Co.,Ltd.)

机构地区：[1]中国石油大学(北京)信息科学与工程学院 [2]中石油北京天然气管道有限公司

出　　处：《化工自动化及仪表》2021年第1期1-7,共7页Control and Instruments in Chemical Industry

摘　　要：LSTM本身具有良好的非线性逼近能力,但在应用于化工流程工业建模时,存在模型泛化能力低的问题。对此,提出一种基于L2正则化LSTM的非线性动态系统辨识策略,将L2正则化项引入网络的损失函数中,优化网络结构,从而提高模型泛化能力。同时,利用TE过程进行相关验证实验,仿真结果表明:所提出的方法优于传统的BP神经网络和支持向量回归,能够有效地提高模型的精度和泛化能力,降低对辨识输入数据的要求。LSTM neural network has good nonlinear approximation ability,but low generalization ability exists when it’s applied to the modeling of chemical process.In this paper,a nonlinear dynamic system identification method based on L2 regularized LSTM was proposed and the L2 regularization term was introduced into the loss function of the network to optimize the network structure and improve the generalization ability of the model.In addition,TE process was used to verify the method proposed.The simulation result showed that,the proposed method outperforms both the traditional BPNN and the support vector regression,and it can effectively improve the model accuracy and the generalization ability and reduce the requirements of the input data for identification.

关 键 词：长短期记忆 动态系统建模 正则化方法 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]