Non-steady-state Control under Disturbances: Navigating Plant Operation via Simulation-Based Reinforcement Learning

We present an empirical application of neural reinforcement learning to chemical plant control for non-steady operations (e.g. product purity change) under disturbance situations. Given a planned trajectory (i.e. a time-series of predicted process values and control inputs), our system provides corrected control inputs that can cancel disturbance and realise trajectory forecast for the actual plant. In addition, given the desired final steady state provided by a human operator (e.g. setting the product purity to 99% and above), our system also provides the desired trajectories. We combined a novel system identification technique and trajectory pursuit method based on neural reinforcement learning and dynamic simulation. We evaluated our method on an actual distillation plant and demonstrated that our system successfully cancelled the emulated disturbance of a weather change (heavy rain) and achieved the desired trajectory.