A Transformer-Based Surrogate Model Predictive Control for Complex Chemical Processes

Quality of the Model Predictive Control (MPC) depends significantly on the predictive model. With the growing availability of industrial data, researchers are increasingly adopting data-driven methods, often using auto-regression and moving average model (ARMA) to accomplish multi-step prediction. However, these models can introduce cumulative prediction errors, affecting control performance. Recently, Transformers, originally designed for Natural Language Processing tasks, have been explored for surrogate MPC modelling. Because of a lack of effective guidance, they might introduce modelling errors caused by irregular temporal dependence and steady-state errors due to model mismatch. To address cumulative and modelling (CM) errors, a new Transformer-based predictive model named CMformer is proposed for MPC. CMformer employs an encoder-decoder architecture to reduce modelling errors and outputs multi-step predictions simultaneously to avoid cumulative errors. Additionally, to eliminate steady-state errors, a selector mechanism which can switch between PI control and MPC control is designed. Experimental validation on a Continuous Stirred Tank Reactor demonstrates superior prediction performance than the encoder-only Transformer, and a constrained closed-loop tracking control task demonstrates the controller's ability to reduce all three types of errors, achieving precise control performance.