Integrating machine learning and thermodynamic modeling for performance prediction and optimization of supercritical CO2 and gas turbine combined power systems

The supercritical carbon dioxide (SCO2) cycle has drawn attention to extracting energy and generating power from low-grade heat resources due to its higher efficiency, lower cost, and compactness. Combining the SCO2 cycle with a gas turbine (GT) is a promising field of research with applications in industries such as waste heat recovery (WHR) systems. In this research, three different configurations of power generation SCO2 cycles, including recuperation, precompression, and reheating, were thermodynamically designed and modeled. Extracting the main parameters of cycles, and the appropriate objective, which is cycle energy efficiency, up to 80,000 data constructed. Random forest (RF) and support vector regression (SVR), machine learning (ML) techniques, are used for the performance prediction of cycles. The results show the strength of the RF method to predict the cycle efficiency with up to 3% error and the SVR method with more than 90% accuracy. Sensitivity analysis and Particle Swarm Optimization (PSO) algorithm are employed for optimizing the recuperation cycle. The outcomes of the optimization model, based on the prediction models as proxies, have results in nearly close to those of traditional optimization, which is based on a thermodynamic model. In conclusion, ML models can be used in industry for prediction and maintenance since the methods are up to 5 times faster than thermodynamic models.