Heat pumps are the key technology allowing to reduce co2 emissions within the residential heating sector. For the improvement in efficiency of the functioning of heat pumps advanced Machine Learning algorithms can substantially contribute to the predictive control strategy and reduce potential costs related to the system failures. In this thesis several unwanted states suggested by the experts are accessed and predictive models based on real-world data from multiple heat pump systems are constructed. An optimal time interval within which a reliable prediction can be made as well as an optimal size of the relevant time series data preceding an event are determined. several Deep Learning models, such as CNN and LSTM, with separate and common architectures for the different unwanted states are compared.