Robust nonlinear control with neural networks

Abstract

A new method for robust nonlinear control of single-input single-output systems is presented. The control law utilizes the universal approximation characteristic of neural networks augmented with the ability for adaptation. The presence of neural networks obviates the need for a mechanistic model for control law computations and the difficulties associated with model-based approaches become irrelevant. The new control law called N-RNCL incorporates the ability for adaptation through an adjustment of bias neurons and ensures offset-free performance in the presence of load and unmeasured disturbances. The performance of N-RNCL is demonstrated using the examples of a strong-acid strong-base pH control system and a nonlinear heat exchanger system. The state-of-the-art controller shows excellent servo and regulatory performance.