NAW-NET: Neural anti-windup control for saturated nonlinear systems

One major issue in industrial control applications is how to handle input constraints due to physical limitations of the actuators. Such constraints introduce nonlinearities in the feedback loop, that are commonly tackled through anti-windup or model predictive control schemes. Since these techniques might result into poor closed-loop performance when an accurate model of the plant is not available, in this work we present an off-line strategy to learn a neural anti-windup control scheme (NAW-NET) from a set of open-loop data collected from an unknown nonlinear process. The proposed scheme, that includes a feedback controller and an anti-windup compensator, is trained to reproduce the desired closed-loop behavior while simultaneously accounting for actuator limits. The effectiveness of the approach is illustrated on a simulation example, involving the control of a Hammerstein-Wiener process with saturated inputs.