Data-driven multi-objective optimization of the helium-xenon closed Brayton cycle system

The helium-xenon (He-Xe) closed Brayton cycle (CBC) system is a space nuclear power thermoelectric conversion system with a compact layout, which has been often optimized using multi-objective optimization (MOO). The traditional model-driven multi-objective (MDMO) optimization approach is time-consuming and fails to reuse optimization data, requiring repeated model calls for new problems. This study introduces a data-driven multi-objective (DDMO) optimization approach using machine learning to reduce time costs and utilize optimization data more effectively. A He-Xe CBC model was developed and calculated with Modelica language, achieving a maximum relative error of 2.8 % compared to the reference data, and a dataset containing 5000 parameter sets was generated using a Python program by randomly varying parameters within the feasible domain. A neural network was trained on this dataset and validated against the original system model. This neural network was then combined with the NSGA-II algorithm for DDMO optimization of the He-Xe CBC system. Compared to traditional MDMO optimization results, DDMO optimization has a maximum relative error of around 1 %, with computation time accelerated by nearly 10,000 times. The integration of system simulation and machine learning for DDMO optimization is crucial for rapid nuclear energy system design optimization and can offer insights into data utilization in nuclear power operations.