Learning-Based Predictive Control Using a Hybrid Model with Adaptive Domain of Validity

The model quality is a key factor for any model predictive control (MPC) scheme to ensure safe and optimal operation of the controlled system. We propose a novel learning-based MPC (LBMPC) approach based on a hybrid model formed of a mechanistic model augmented with a data-based model trained offline which compensates the mismatches of the mechanistic model. If the prediction accuracy of the nominal model deteriorates due to disturbances or new modes of operation, the nominal model is augmented with an additive model learned online using a limited amount of data to capture the residual of the offline trained model. As using data-based models can be unreliable outside their training regions, we estimate their domain of validity via a one-class support vector machine (SVM) trained on low-dimensional projections of the training data. The predicted validity of each data-based model is used to weigh its contribution to the overall model. During online application, the domains of validity of the data-based models are adapted if needed. We show the advantages of the proposed LBMPC approach via a continuous stirred tank reactor (CSTR) example.