IMPLEMENTAÇÃO DE CONTROLE PREDITIVO MULTIVARIÁVEL E DE SENSORES VIRTUAIS EM UMA PLANTA DE CAPTURA DE CO2

Abstract

Investing in studies related to the CO2 capture process, either to improve its efficiency or to ensure that current environmental legislation is met, is a potential alternative to enable some targets present in the SDGs 9, 12 and 13 listed in the 2030 Agenda, on responsible production, industrial technological innovation and action against climate change. Consequently, this work aims to implement the classic (PID) and advanced (MPC) controls in a CO2 capture process, in order to maintain the CO2 capture rate (%CC) and the stripper reboiler temperature (TREB) at the desired values, and develop soft sensors, based on the NARX model, capable of estimating such variables. To control %CC and TREB, the molar flow rates of lean solvent and steam in the stripper (FLS and FST) were manipulated, respectively. The PID controllers were tuned using the Ziegler-Nichols and Tyreus-Luyben methods, and were later submitted to tuning by the Simulink® Tune tool. The MPC controller model is based on the multivariable dynamic matrix (MIMO DMC) under fixed conditions of prediction horizons R equal to 40 s, control L equal to 10 s, stabilization N equal to 80 s and sampling time T equal to 2.5 s. Qualitative graphical analysis guided the choice of the best arrangement of the PIDs and, subsequently, the comparison between conventional and predictive controller in face of the application of disturbances in the manipulated variables. The ANNs were trained with the Levenberg-Marquardt algorithm, hyperbolic tangent as an activation function, and validated with the early stop method combined with cross validation, with the MSE being the decisive criterion for selecting the best sensor. All PID controller arrangements experienced overshoots in the dynamics of the process variables, with the PIDTL arrangement having the best performance, while the MIMO DMC was able to eliminate this phenomenon, accelerate the PVs stabilization by 95% and reject the disturbances used in the variables manipulated, when compared to PID. NTCC5 and NTTR5 sensors, both with 25 neurons in the hidden layer, were the soft sensors chosen to estimate %CC and TREB, under validation errors equal to 9.232x10- 2 and 1.534x10-2 .