Uso de observadores de estado aplicados no controle inferencial preditivo de processos não lineares.

Abstract

To meet increasingly stringent quality, safety, economy and sustainability standards, modern process control techniques such as predictive model control (MPC) can be used in complex and highly nonlinear systems where the effective implementation of classic control (like the conventional PID algorithm) becomes difficult. In the presence of process variables whose real-time, low-sampling measurements are unavailable or costly to maintain and implement, state observers can be used as estimators in the role of virtual sensors. In this work, a constrained nonlinear predictive inferential control (NMPC) was implemented in a non-isothermal CSTR reactor, propylene glycol producer from propylene oxide hydrolysis and whose inputs are contaminated with noise. Plant states were estimated by extended (EKF) and unscented (UKF) Kalman filter-type state observers, tuned by three distinct approaches. Also, a heuristic approach was used in order to tune the main parameters of the NMPC: sampling time, control horizon and prediction horizon. The performance of the state observers differed little between EKF and UKF, since the correct tuning of the process covariance matrix Q was determinant for the estimation effectiveness. While in classical inferential control (PID-UKF and Cascade-UKF) there were difficulties in controlling the process in unstable regions (also due to plant noise), advanced inferential control (NMPC-UKF) was able to minimize plant oscillations while providing faster control response, with lower overshoot, more stable and meeting the temperature restriction imposed in the process against disturbances at the inputs (regulatory control), at the setpoint (servo control), and at the reactor startup.