Hybrid modeling and energy optimization of NGL demethanizers using deep reinforcement learning

Abstract

Abstract Cryogenic NGL demethanizers suffer from energy inefficiencies due to non-convex thermodynamic constraints. This study aims to implement a Deep Reinforcement Learning (DRL) framework to optimize energy consumption while respecting complex operational boundaries. A Soft Actor-Critic (SAC) agent was employed, trained on 2,000 rigorous Aspen HYSYS digital twin scenarios via a Deep Neural Network (DNN) surrogate. Unlike simplified proxy models, this supervisory strategy directly interacts with the high-fidelity simulation to navigate nonlinear process dynamics. The results demonstrate a 12 % reduction in reboiler duty compared to baseline operations, yielding potential annual savings of $4.0 million and a reduction of approximately 15,000 tonnes of CO₂. Sensitivity analysis confirms economic robustness under conservative market conditions. The proposed framework demonstrates the potential of DRL as a computationally efficient alternative for steady-state supervisory optimization in complex cryogenic processes. By utilizing a high-fidelity steady-state process surrogate, it provides a promising framework for industrial energy optimization, setting the stage for future real-time industrial applications.