This paper presents the results from a real residential field test in which one of the objectives was to maximize the instantaneous self-consumption of the local photovoltaic production. The field test was part of the REnnovates project and was conducted in different phases on houses in several residential districts located in Soesterberg, Heerhugowaard, Woerden and Soest, the Netherlands. To maximize self-consumption, buffered heat pump installations for domestic hot water and stationary residential battery systems were chosen due to their respective thermal and electrical storage capacities. The algorithm used to tackle the associated sequential decision-making problem was model-based reinforcement learning. The proposed algorithm learns the stochastic occupant behavior, uses predictions of local photovoltaic production and considers the dynamics of the system. The results show that this algorithm increased the average self-consumption percentage of the local PV generation (used instantaneously in situ) on average by 14%, even if only buffered heat pump installations for domestic hot water were used. This increase was achieved without causing any perceived discomfort to the residential end users. The average energy shifted per day from the solar production period to the night by the 2 kW/3.6 kWh batteries was 1.5 kWh. The main contribution of this work was therefore the real field implementation of the proposed algorithm. The results demonstrate that it is possible to improve even further the integration of local production using flexible loads.