Soft faults evaluation for electric heat pumps: Mechanistic models versus machine learning tools

To reduce carbon footprint of heating and cooling, electrical heat pumps (EHP) will have more room of application because of the major use of electricity produced by renewables. To ensure high performances, it is important to develop fault detection, diagnosis and evaluation strategies (FDDE) for soft faults, which do not cause a stop of the EHP and could silently be detrimental (e.g. refrigerant leakages, heat exchangers fouling). In this paper, a surrogate database under faulty conditions generated by a mechanistic model is used to compare the ability in evaluating soft faults and performance degradation of three different approaches: one based on a look-up table implemented remotely and the other two based on machine learning. Among them, one is an artificial neural network (ANN) and the other is a K-Nearest Neighbors (KNN) classification method. All the approaches were developed and tested, considering as inputs 5 measured variables on the machine among pressures and temperatures, characterized by an instrument uncertainty of 0.2°C and 0.2 bar. Results show that all the investigate approaches can similarly evaluate faults, with the ANN able to better evaluate early-stage fault intensities for all the three faults investigated.