There is considerable interest in supercritical carbon dioxide power cycles for concentrated solar power applications. Therefore, there is a need for accurate and computationally inexpensive models for such plants. The calculation of the radiative heat flux impinging upon a central receiver is required to model the heat transfer to the fluid flowing through it, as well as the temperatures of the tube wall material. The flux profile calculations are computationally expensive since the central tower is surrounded by thousands of individual heliostats, which are all sources of radiative flux.

To reduce the computational expense, a data driven surrogate model was developed using a multilayer perceptron neural network that predicts the flux profile impinging on the receiver for a range of plant configurations and atmospheric conditions at a specific location. It accounts for different power requirements, central tower heights and receiver geometries, different azimuthal and elevation angles of the sun, as well as DNI. The model was trained on data produced using SolarPILOT for Upington in Southern Africa.

The trained model predicts the flux profile of the receiver with an out of sample root mean squared error of less than 0.2%. The surrogate model is more than 300 times faster than directly solving the flux using SolarPILOT. This means that it can produce hourly flux profiles for a full year of operation in oneminute real time, compared to more than six hours when using SolarPILOT.

The flux profile surrogate model was combined with a surrogate model of the central receiver heat transfer network to create an integrated model that can predict the heat transfer and tube wall temperatures. The integrated model is demonstrated via a Monte Carlo analysis to predict the performance over the course of a year for a specific 350 MWth central receiver design.