Machine learning based system-level control for the autonomous S-CO2 power cycle under load varying conditions

This study presents a machine learning-based control methodology for the system-level operation of the Supercritical CO2 (S-CO2) power cycle, advancing from traditional PID-based component-level control. A deep reinforcement learning (DRL) agent was developed and trained using a two-stage simulation based framework: pre-training in a time-series surrogate model environment and fine-tuning through transfer learning in a high-fidelity simulation code environment. The performance of the DRL agent trained under virtual environment exclusively was validated through hardware testing in the S-CO2 power cycle test loop, where it successfully managed dynamic control tasks. The results demonstrated the applicability of the DRL agent trained under virtual environment in real-world systems, highlighting the potential of machine learning-based control strategies to enhance operational efficiency and stability in advanced energy systems. This work provides a foundation for future applications of DRL in large-scale power systems, paving the way for improved control methodologies in the next-generation energy technologies.