A generalizable neural network model for supercritical CO2 turbine performance and map generation method

Supercritical and transcritical CO2 power cycles are recognized for their high efficiency and compactness. As the pivotal component, a turbine performance map is crucial for efficient cycle design, analysis and optimization. Neural networks serve as powerful tools for generating supercritical CO2 turbine performance maps from the scarce experimental data. This paper constructs a 10-kW demo transcritical CO2 power cycle, and experimental results show that the investigated Turbine-Generator can achieve maximum dynamic power and efficiency values of 11.5 kW and 43.2%, respectively. Based on detailed analyses, four neural networks, including sequential and non-sequential networks, are trained to evaluate the mass flow and power. Through comprehensive evaluation, the Time-Delay Neural Network, a sequential network, demonstrates the best performance. It takes an average of 2.9 s per training and shows the lowest mean squared errors of 7708.8 and 5.28e-5 for power and mass flow, respectively. In addition, it predicts the Turbine-Generator performance accurately for both dynamic (with mean squared errors of 20,499 and 3.58e-4 for power and mass flow, respectively) and steady-state conditions (with mean squared errors of 3035 and 9.59e-6, respectively). Furthermore, a simple data extension method is proposed to plot the overall performance map of Turbine-Generator. This map displays efficiency contour lines with the corrected flow rate and enthalpy drop as the horizontal and vertical axes, respectively, showing a maximum efficiency above 35%. This work points out a sequential network and a data extension method to plot overall turbine performance maps, providing a reliable foundation for cycle design and optimization.