Direct learning of improved control policies from historical plant data

The continuous optimization of the operational performance of chemical plants is of fundamental importance. This research proposes a method that utilizes policy-constrained offline reinforcement learning to learn improved control policies from abundant historical plant data available in industrial settings. As a case study, historical data is generated from a nonlinear chemical system controlled by an economic model predictive controller (EMPC). However, the method's principles are broadly applicable. Theoretically, it is demonstrated that the learning-based controller inherits stability guarantees from the baseline EMPC. Experimentally, we validate that our method enhances the optimality of the baseline controller while preserving stability, improving the baseline policy by 1% to 20%. The results of this study offer a promising direction for the general improvement of advanced control systems, both data-informed and stability-guaranteed.