Practical Nonlinear Model Predictive Control Using an Echo State Network Model

This paper presents a novel suboptimal nonlinear model predictive controller which makes use of Echo State Network (ESN) models. The controller is based on the Practical Nonlinear Predictive Controller algorithm, a general framework which can be used for the implementation of nonlinear model predictive controllers using almost any class of nonlinear model. Echo State Networks explore the Reservoir Computing learning paradigm, which is a computationally efficient strategy for training recurrent artificial neural networks. In the proposed control algorithm, the ESN model is used to obtain on-line a nonlinear prediction of the system free response and a linearized version of the model is used to obtain a local approximation of the systems step response, which is used to build the dynamic matrix of the system at each sampling instant. The proposed controller was tested in a real nonlinear process, comprising the automatic control of the discharge pressure of a refrigerant compressor in a test rig. This process was identified using an ESN and the results are compared to linear and Hammerstein nonlinear models. The controller was implemented in the real test rig and compared with a classical PID, being tested at distinct operating conditions. The modeling approach presented good results, with the Echo State Network model outperforming the linear and Hammerstein nonlinear models. The controller implementation results show that the Practical Nonlinear Model Predictive Controller algorithm with Echo State Networks enables the application of nonlinear model predictive control with recurrent neural networks with low computational burden, both in the process modeling phase and control signal optimization.