Identification For Nonlinear Dynamic System Based On LSTM

Chemical process dynamic model usually has many characteristic,like strong coupling,nonlinear,large time delay,large inertia and complex mechanism.Nonlinear structure and model accuracy based on some classical identification methods are difficult to meet the needs for further control or other applications.At the same time,it strongly depends on the data obtained from the strict identification experiment,which leads to the great limitation of its application scenarios.If a large number of historical data can be fully used which are from factory's DCS to establish a good and be easy to apply dynamic model,and through the model structure and parameters to express the nonlinear dynamic process,it will be better able to grasp the intrinsic characteristics of the process.It is easy to develop more efficient and stable control method,to improve economic efficiency and production safety.In this paper,Long Short-Term Memory,a kind of deep learning model,is applied to study the identification of nonlinear dynamic systems,in order to discover the information of a large number of dynamic data in the DCS of the factory,complete the identification of the production process,and improve the identification accuracy.First,the data from TE process are used to verify the feasibility of LSTM which identify nonlinear dynamic system.The grid search method is used to select hyper-parameter,and the influence of learning rate and other hyper-parameters on identification accuracy in LSTM is also discussed.Secondly,in order to solve the problem of the low model accuracy and model generalization ability when LSTM is used for nonlinear system identification,an identification method based on L2 regularization LSTM is proposed.By comparison with BPNN and Support Vector Regression methods,it is verified that it can effectively improve the accuracy and generalization ability of the model and reduce the requirements on input data.Finally,to avoid falling into the local optimal,make full use of the past information,improve the model accuracy,the Gray Wolf Optimization algorithm and the fractional order derivative are used to optimize the LSTM respectively.By combining the simulation results and the advantages of the two above,the GWO-FOD-LSTM method is proposed.The proposed method is used to identify TE process and quadruple tank model.Identification results show that the proposed method has low requirements for identifying input data,high identification accuracy,and can fully express the nonlinear dynamic characteristics,and can fully simulate and reproduce the actual production process.