Physics-Informed Machine Learning for Solar-Thermal Power Systems: Article No. 119542

Thermal energy system modeling traditionally relies on experimental correlations to estimate heat transfer coefficients using dimensionless numbers and thermophysical properties. Multiple correlations have been proposed for different flow regimes, system geometries, and boundary conditions; however, despite their widespread use, these correlations present significant limitations related to accuracy, valid flow regime ranges, and adaptability to complex system geometries. Moreover, systems powered by intermittent renewable sources, such as solar, face added challenges due to high fluctuations in heat transfer coefficients and system variables. In this work, we change fundamentally this old paradigm. We demonstrate an experimentally validated, physics-informed machine learning approach to accurately estimate time-dependent heat transfer coefficients in solar-thermal energy systems with high variability in energy input, boundary conditions, and load. This method combines data with governing physics, providing a more accurate modeling alternative to purely data-driven frameworks and addressing the limitations of traditional experimental correlations and the need for extensive historical data for offline training. More broadly, the proposed paradigm shift can be employed in other energy plants and complex systems-of-systems.