Learning-Based Model Predictive Control with Hammerstein Model

The reaction of C3 hydrogenation process is relatively complex and has a strong non-linear characteristic, it is difficult to obtain an accurate first principle model, and there is a large uncertainty in the model with the operating conditions change. Traditional model predictive control (MPC) effect relies on the accuracy of the prediction model, which is difficult to make adaptive adjustments with changes in the controlled process. In this work, a learning-based model predictive control with Hammerstein model is proposed, the linear part of the nominal Hammerstein model is added with an unknown model residual, the residual is learned through Gaussian process regression and the training data dictionary is updated in real-time to improve the accuracy of the model. The optimal intermediate variables are solved using differential dynamic programming according to the linear part considering the residual of the model, and then the real control inputs are obtained according to the inverse model of the nonlinear part. The simulation and industrial application results have verified the feasibility and effectiveness of the proposed method.