Repetitive Model Predictive Control Based on a Recurrent Neural Network

Repetitive model predictive control (RMPC) is an advanced optimal control method which includes an internal model based repetitive controller for periodic signal tracking and constraint handling. The solution of RMPC, just like standard model predictive control (MPC) design, involves solving a tedious optimization problem and thus limits its practical use in slow sampled-data control systems. The recent advance in microprocessors has made implementation of this heavy computational control on fast dynamic systems a possible task and developing an efficient algorithm for real-time MPC implementation, namely fast MPC technique, has become an emerging research topic in the community. Following the concept this study presents a neural network based RMPC method by applying a recurrent neural network to solve the quadratic programming problem during the optimization process. The experimental results on repetitive tracking control of a piezo-actuated system demonstrate that the developed neural network based RMPC shows improved performance with less computational burden compared to the previous work using traditional Hildreth´s solver for RMPC implementation.