Robust nonlinear model predictive control of continuous crystallization using Bayesian last layer surrogate models

In scenarios where high-fidelity physical models are either unavailable or are impractical due to their high complexity, data-based models offer a viable solution to obtain the system model necessary for predictive control. However, the accuracy of the predictions obtained by data-based models is limited. We propose to use neural networks with Bayesian last layer to obtain information about the uncertainty of the predictions. This paper demonstrates the use of Bayesian last layer surrogate models in a robust nonlinear model predictive control setting. The nonlinear model predictive control problem is adapted by considering the predicted uncertainty of the surrogate model, which can be efficiently computed using the Bayesian last layer method, in the cost function. The controller thus takes model uncertainty explicitly into account and by its formulation also avoids areas of extrapolation. The proposed method is applied to a mixed-suspension, mixed-product-removal crystallizer and simulation studies show that it outperforms a standard data-based model. Copyright (C) 2024 The Authors.