Learning MPC for Process Dynamic Working Condition Change Tasks under Model Mismatch

In this study, a learning model predictive control (MPC) algorithm for process dynamic working condition change (DWCC) tasks is proposed. The algorithm continuously compensates for model–plant mismatch (MPM) and improves dynamic performance by predicting multi-step-ahead disturbance from similar DWCC tasks. First, a state-space model augmented by disturbance variables ensures offset-free control for MPM. Second, a dynamic autoencoder is constructed to extract private features from process sequences based on long short-term memory and fully connected networks. DWCC scenarios similar to the current scenario are located from the historical database by calculating the distance between extracted features. Finally, the multi-step-ahead disturbance and its uncertainty representation are predicted through multi-output Gaussian process regression based on the located scenarios. The obtained multi-step-ahead disturbance is incorporated into the state-space MPC framework. A nonlinear case is conducted to demonstrate the effectiveness of the proposed method.