Dynamic System Modeling Based On Improved NARX Neural Networks With Applications In Predictive Control

System modeling plays an important role in various fields such as industrial process control,prediction,and optimization.Because that industrial processes usually exhibit complex dynamic nonlinear characteristics,traditional modeling methods are rarely accommodated to these process systems.Alternatively,neural networks can take advantage of process system modeling.In this context,this thesis extensively investigates NARX dynamic neural networks along with improvements on existing problems.The improved NARX neural networks are employed to establish nonlinear dynamic processes.As an application,this network is applied to the predictive control scheme.The achievements and main research contents of the thesis are presented as follows.(1)Improved fruit fly optimization algorithm is proposed.The fruit fly optimization algorithm is extensively studied before some dominant swarm intelligent algorithms are experimentally compared in terms of advantages and disadvantages.Based on the investigation on the influence of algorithmic parameters on the optimization performance of fruit fly algorithms,an adaptive variable step size fruit fly optimization algorithm is proposed along with experimental studies on the performance of the novel algorithm.(2)Combining the improved fruit fly optimization algorithm with the LM algorithm,an IFOALM algorithm is proposed and applied to NARX neural networks,which can overcome the problem of slow training stemmed from random assignments of initial weights and improve the process modeling performance.In regard to nonlinear dynamic systems,the improved NARX neural networks based modeling method is presented.The validity of the method is verified by numerical and process examples.(3)The improved NARX neural networks are applied to predictive control scheme.Therein,a multi-step prediction model is established based on the network;the Golden Section Search algorithm is used for rolling optimization,and the prediction error is used for feedback correction.Numerical examples and pH neutralization experiments were carried out to show the predictive control performance,leading to satisfied results.