Fourier Neural Operator-driven transient analysis and control for supercritical CO2 cycles

The dynamic performance and control strategies of supercritical carbon dioxide (SCO2) systems constitute a fundamental research area that demands thorough scientific exploration. This study systematically addresses several critical gaps in the transient analysis of SCO2 systems, particularly highlighting the insufficient theoretical exploration of the system's dynamic behavior, the predominant reliance on conventional Proportional-Integral-Derivative (PID) control approaches, and the significant computational challenges in solving transient dynamics and control problems. To this end, the following research plan is proposed: first, the intrinsic principles of SCO2 dynamic characteristics are analyzed based on the printed circuit heat exchanger (PCHE) transient model; second, system identification is performed for the SCO2 system; and then, Model Predictive Control (MPC) and PI control-based reinforcement learning (RL-PI) suitable for the SCO2 cycle is designed and compared with traditional PI control; and finally, a Fourier Neural Operator (FNO)-based model for the open-loop response of the SCO2 system and its MPC controller is trained. The results shows that, at the microscopic level, the PCHE outlet temperature consists of an inertial system driven by wall heat transfer and a non-minimum phase system driven by inlet enthalpy. At the macroscopic level, efficiency exhibits non-minimum phase behavior, while net output power does not. A backpropagation (BP) neural network achieved optimal system identification, with training set fits of 97.56 % (efficiency) and 99.19 % (net output power), and validation set fits of 92.79 % and 97.75 %, respectively. MPC outperformed PI, RL-PI, and open-loop control, showing shorter response times, reduced overshoot, stronger noise immunity, and improved signal tracking. FNO successfully predicted the SCO2 cycle's open-loop response, achieving relative errors of 10−4 for net output power and control outputs compared to MPC. The system performance was validated through numerical simulations, and results indicated that FNO has a significant advantage in computational efficiency, with memory and time reduction by 3 orders of magnitude.