Learning of model-plant mismatch map via neural network modeling and its application to offset-free model predictive control

Self-Evolving Offset-Free Model Predictive Control with Model–Plant Mismatch for Dynamic Working-Point Change Tasks in Industrial Processes

“…Although the proposed method exhibits excellent dynamic behavior in DWPC tasks, the following challenges need to be addressed in future research. This study specifically focuses on the disturbances caused by MPM, which dominates the performance of MPC. − However, in practical processes, external disturbances driven by the environment, such as noise, temperature, and changes in feed composition, are another important factor affecting MPC performance. External and internal disturbances have different characteristics, and it is necessary to identify both types of disturbances from measurement and design control methods for each type. The established MOGPR model in Section can accurately predict future multistep-ahead disturbances and their uncertainties.…”

“…However, most studies estimate the current disturbance based on the output error and use it for the MPC calculation at the next time step, which is insufficient when performing dynamic tracking tasks because it introduces delays in disturbance estimation and MPM compensation. Moreover, most research and industrial MPC implementations use a constant output disturbance model, assuming that the disturbance estimated remains unchanged within the MPC prediction horizon . This assumption does not hold in practice because MPM varies with the system state in dynamic tasks, and a disturbance estimate at one state is unavailable for another.…”

“…Offset-free MPC usually augments the prediction model with disturbances and designs an observer to estimate the disturbances. 5 Related studies can be divided into two categories: linear and nonlinear. Focusing on linear MPC, Muske and Badgwell analyzed the design approach for the disturbance model in the state−space MPC framework.…”

“…Moreover, most research and industrial MPC implementations use a constant output disturbance model, assuming that the disturbance estimated remains unchanged within the MPC prediction horizon. 5 This assumption does not hold in practice because MPM varies with the system state in dynamic tasks, and a disturbance estimate at one state is unavailable for another. Due to these limitations, the above offset-free methods cannot rapidly handle the MPM in dynamic tasks, which results in a significant degradation in the dynamic performance of MPC.…”

“…To compensate for MPM in MPC and eliminate the resulting offset control behavior, offset-free MPC is widely studied. Offset-free MPC usually augments the prediction model with disturbances and designs an observer to estimate the disturbances . Related studies can be divided into two categories: linear and nonlinear.…”