Reinforcement learning for community energy sharing: a cost-fairness analysis

Abstract

This work proposes a reinforcement learning strategy for allocating photovoltaic (PV) energy in energy communities with collective self-consumption. A three-member community with unequal consumption profiles and participation rights is considered as a case study. The proposed approach employs Proximal Policy Optimization (PPO) and is compared with static allocation, a rule-based controller (RBC), and model predictive control (MPC). All strategies determine the same allocation coefficients while keeping the local control layer fixed. The evaluation considers the average cost and the fairness of cumulative compliance with PV participation rights. PPO and MPC share the same 8 h prediction horizon, whereas RBC does not use forecasts. Simulation results show that RBC achieves the lowest average admissible cost, MPC provides the highest dynamic fairness, and PPO offers a compromise between the two. These results demonstrate the potential of reinforcement learning for energy allocation in energy communities.