HYBRID PID AND REINFORCEMENT LEARNING CONTROL FOR OFFSHORE GAS PROCESSING

This work presents an advanced control framework for an offshore natural gas processing system with high CO₂ content. The objective is to optimize gas export flow while ensuring CO₂ specification compliance and minimizing energy consumption. A digital twin (DT) was developed in Python to model membrane separation, multistage compression, and gas recycling, incorporating real gas behavior and equipment constraints. Control was implemented using conventional PID loops augmented by a reinforcement learning (RL) agent trained in the DT with the Proximal Policy Optimization algorithm. The RL agent optimized setpoints for pressures and flow rates across the system. Results show that the RL-enhanced controller increased the export gas flow by approximately 115%—from 13 kg/s to over 28 kg/s—compared to PID control alone, while maintaining CO₂ levels closer to the 3% specification limit. The hybrid control strategy also maintained suction pressures and discharge temperatures within ±30% of safe operational values, presenting improved performance over static PID tuning in offshore scenarios.