Actor-critic reinforcement learning to estimate the optimal operating conditions of the hydrocracking process

Abstract Determining the optimal operating conditions for hydrocracking units is imperative due to the changing nature of production requirements. However, it is expensive to optimize the hydrocracking process with mathematical models because hydrocracking units have a limited capacity for quick response and customization. This study proposes an actor-critic reinforcement learning optimization strategy using a DNN surrogate model, which was developed from a validated mathematical model with a marginal error of less than 2%. The surrogate model interacted with the A2C algorithm and the optimal operating conditions were determined with an accuracy of 97.86% and 98.5%. To demonstrate the reliability, case studies were executed; the strategy was found to be consistent, with an average efficiency of 98%. The proposed approach offers the advantages of quick response time, low computational burden and customizability for online implementation, which are essential for practical optimization problems. It can be extended beyond hydrocracking to other chemical industries.