Energy efficiency in the built environment has been identified as one of the key enabling technologies to meet global climate change targets. In this paper, we present promising results from a black box method to automatically characterize various aspects of heat pump operation in residential settings. Experimental data is gathered from heat pumps used to provide spatial heating and domestic hot water in recently refurbished net-zero energy houses. This is done by data-driven determination of the heat pump’s performance and the impact of building occupants. These interactions, typically in the form of hot water consumption profiles and preferences for temperature set points, are learnt from sensor data. This allows the formulation of an explicit Markov Decision Process (MDP), which can be solved with the objective to maximize energy efficiency of local heat pump operation. In doing so, we show substantial gains over default policies (grounded in thermodynamics) but which don’t consider occupant behaviour. Three key short-term benefits are envisaged from this research: first, leveraging such synergies allows the energy efficiency of heat pump operation to be improved by, on average, more than 10%. Second, automation unlocks the potential to circumvent the costly, non-generalizable model building step in model predictive control. Finally, it allows direct, unbiased benchmarking of theoretical performance of different types of heat pumps against real world performance. ispartof: Heat Pump Conference location:Rotterdam 2017 status: published