Advanced control of compressor inlet temperature in supercritical CO2 Brayton cycle

The supercritical CO2 (sCO2) Brayton cycle has gained much interest because of its flexibility, compactness and high efficiency. The sCO2 at the inlet of compressor should work near the critical point to obtain high system efficiency, while the physical properties of sCO2 near the critical point change dramatically, which brings great challenges to the control of the compressor inlet temperature. At present, cooling far from the critical point and adjusting the cooling water flow rate with Proportion- Integration-Differentiation (PID) controller is the commonly used method, but it loses system efficiency and may be out of control sometimes. Therefore, in this study Linear Model Predictive Control (LMPC) and Deep reinforcement learning (DRL) are used to control the compressor inlet temperature and compared with PID. A dynamic model of a recompression sCO2 Brayton cycle is established, and the cooler model is carefully validated against experiment data. The results indicate that both the LMPC and DRL can control the sCO2 temperature near the critical temperature much better than the PID. LMPC works the best because the cold end parameters fluctuate slightly and the cooler model can be regarded as approximately linear, thus LMPC can find almost the global optimal solution. Nevertheless, DRL control exhibits the fastest real-time computation ability and proves good extrapolation ability.