Data-driven energy sharing in collective self-consumption communities: a medium-term predictive ensemble framework

This paper presents a predictive ensemble framework for the allocation of energy in collective self-consumption communities. The proposed approach combining several simple yet complementary models—seasonal persistence, previous month, previous year, rolling average, k-nearest neighbors —within a minimum-variance optimization that dynamically adjusts model weights according to recent predictive performance. The framework can deliver stable and accurate results with limited historical data, reaching full reliability after approximately six months of hourly observations. Using multi-year data from a representative energy community, the ensemble method was benchmarked against individual predictors and a reference baseline representing perfect foresight. Results show that the ensemble consistently minimizes deviations from the baseline across all consumers and time periods, outperforming each single model. Sensitivity analyses with respect to photovoltaic capacity, import prices, and export remuneration confirm the robustness of the approach: the ensemble systematically achieves the lowest cost gap relative to the baseline under all tested conditions, with the greatest advantage observed under high import tariffs and low photovoltaic capacity. Overall, the proposed architecture provides a realistic and scalable solution for managing collective self-consumption under current regulatory frameworks, which often require predefined allocation coefficients set months in advance. Beyond its economic performance, the framework opens the door to future developments incorporating uncertainty management, fairness criteria, and online learning, paving the way toward intelligent and equitable energy communities.