Shell: AI-driven autonomous plant operation for Shell Scotford
3:50 PM - 4:20 PM
Industrial AI

A DRL agent was trained on a cut-down AVEVA™ Dynamic Simulation of the MEG separation column and subjected to a large variety of normal and upset conditions from the point of feed loss hand-over from procedural automation, stabilizing the column in total reflux automation and restarting the column upon re-introduction of feed while also minimizing the unit’s energy usage. The resulting agent was rigorously tested on a version of the Operator Training Simulator and compared to operator benchmark data. The agent’s performance was, in the majority of cases, highly satisfactory and in some cases capable of maintaining the process in tighter control than the operator benchmark. The result of this project shows that DRL agents are an extremely promising addition to process control technology to bring safe, reliable, and repeatable operation during complicated events, thereby reducing the operator workload and the vulnerability to lost production.