Improving energy distribution in collective self-consumption via XGBoost-based allocation coefficients prediction

Energy communities operate under collective self-consumption schemes, where locally generated renewable energy is shared among participating members. In practice, this sharing is commonly governed by static allocation coefficients fixed in advance, which do not capture the time-varying and heterogeneous demand of participants. This mismatch can reduce community self-consumption, increase surplus injections, and raise reliance on the grid. This paper proposes a data-driven framework to dynamically compute allocation coefficients based on predicted individual demand and demonstrates its application in a municipal energy community in Catalonia, Spain. The approach uses an extreme gradient boosting model to forecast hourly consumption profiles and then derive adaptive allocation coefficients that better align shared photovoltaic generation with expected demand. The proposed strategy is evaluated against a static baseline and alternative dynamic schemes using multiple performance indicators, including community self-consumption, surplus energy, and grid dependency. In the case study, the extreme gradient boosting-based allocation increases community self-consumption by 8.4%, reduces surplus energy by 34%, and lowers grid dependency by up to 30% for key members, resulting in a more balanced and efficient distribution of locally generated energy. These results highlight the potential of machine learning-enabled allocation to improve collective self-consumption performance in the existing regulatory framework.