Optimal design of a supercritical carbon dioxide recompression cycle using deep neural network and data mining techniques

This study aims to explore and optimize the high-dimensional design space of the supercritical CO₂ recompression cycle (sCO₂-RC) using deep learning and data mining techniques. Firstly, the thermodynamic model of the sCO₂-RC for nuclear-powered ships is established. The design space of the system is then comprehensively explored by data mining techniques consisting of Sobol's sensitivity analysis and self-organizing map based on deep neural networks. Subsequently, a baseline optimisation considering the thermal efficiency and total component volume of the system is performed, and the effect of exergoeconomic and specific work on the optimal baseline results is discussed. The results show that the parameters whose multi-order sensitivity is comparable to first-order sensitivity are crucial to the trade-off of the performance indicators. Data-mining techniques can provide useful information for system optimisation by quantifying parameter interactivity and visualizing the nonlinear relationships between parameters and objectives. The optimal thermal efficiency and total component volume under the baseline optimisation are 39.92% and 3.92 m³, respectively, which are correspondingly reduced by 1.44% and increased by 0.68 m³ when exergoeconomic and specific work are considered. This study has guiding significance for the optimal design of the sCO₂-RC with high-dimensional design space in space- and energy-limited scenarios.