Research Of Nonlinear Model Predictive Control Based On Growing And Merging RBFNN

Modern petrochemical production plants are increasingly developed in the direction of large,complex,and highly coupled parameters,with strong nonlinear and uncertain characteristics,which make it difficult to establish accurate mechanism models,and traditional linearization methods and simplified mechanism models have limited accuracy,thus their process modeling and automatic control still face challenges.Based on this,this paper deeply investigates the process modeling method based on radial basis function neural network(RBFNN),and on this basis,the data-driven nonlinear model predictive control(NMPC)method is studied.The main research works are as follows:(1)An improved growth-merge RBF network(IGM-RBF)modeling method is proposed.The network can achieve simultaneous optimization of network structure and parameters,and solve the problem that the number of nodes in the hidden layer is difficult to determine.First,a network initialization method for sample output clustering is designed.Then,its structure is optimized based on the relative error of the model and the cosine similarity between nodes,while a discriminant method based on the sample distance is used to make the network more adaptable.Its parameters are then trained using the improved LM algorithm,and the convergence of the algorithm is demonstrated in two steps based on Lyapunov theory.Experimental simulation results show that the IGM-RBF proposed in this paper can achieve a compact structure in a short training time,and also has good prediction accuracy and generalization ability.(2)The NMPC controller is designed based on the IGM-RBF network.The IGM-RBF network is applied to the NMPC as a prediction model for nonlinear processes,and the multi-step prediction results are obtained by a recursive strategy,while the control law is solved using the gradient method to ensure the real-time performance of the MPC,and a detailed procedure for solving the control law is derived and given.To ensure the accuracy of the model,a correction method combining feedback correction and network weight update is designed.And the stability of the proposed control system is proved by constructing the Lyapunov function.The effectiveness of the control scheme is verified by numerical cases and continuous stirred kettle reactor system,which has good tracking control ability and anti-disturbance and anti-noise capability.