A Machine Learning (ML) model for performance prediction control strategy for a dual-source direct-expansion solarassisted transcritical carbon-dioxide heat pump
In accordance with the progressive electrification of energy consumption outlined by European Community policies and given the expected increase in the use of heat pumps to replace traditional heating systems based on fossil fuels combustion, it is essential to predict the performance of these machines and develop reliable control strategies to ensure their efficient and sustainable operation. To address this challenge, this work develops a data-driven methodology for real-time source selection in a direct-expansion dual-source heat pump prototype working with carbon-dioxide as working fluid, by combining experimental measurements with Machine Learning-based performance modelling. The system, with a heating capacity of about 15 kW and equipped with both finned-coil heat exchanger and hybrid photovoltaic-thermal panels as evaporators, was tested for domestic hot water application with a temperature lift from 20°C to 60°C in a range of compressor frequencies between 50÷60 Hz, ambient temperature of 16÷25°C and solar irradiances of approximately 225÷725 ????/????2. Based on this data, Artificial Neural Networks were then trained to predict the prototype performances only considering boundary conditions such as ambient temperature and solar irradiance to generate dense performance maps across the operating domain. Thanks to these performance maps, it was possible to reconstruct, for each compressor frequency, an irradiance threshold curve as a function of ambient temperature, above which the Sun-Mode heat pump operation outperforms the Air-Mode operation. This approach provides a basis for reliable, real-time supervisory control strategies for dual-source solar-assisted CO2 heat pumps.