Preference-aligned value stacking for household battery storage via a selective policy sharing multi-agent reinforcement learning algorithm

Household battery storage deployed for a single service provision often suffers from insufficient economic returns and overlooked environmental benefits. To overcome these limitations, this paper proposes a preference-aligned value stacking framework that integrates incomes from community energy trading and frequency regulation services, thereby creating both economic and environmental benefits for users. User preferences over these two benefits are elicited through customized binary choice tasks, and a discrete choice model is employed to estimate individual preference parameters. To align service provision with user preferences, a selective policy-sharing multi-agent reinforcement learning algorithm is developed. The algorithm incorporates a preference-embedded reward function to guide policy optimization toward preference-consistent decisions. Moreover, users with similar preferences share policy networks to enhance training scalability in large-scale user settings. Simulation results demonstrate that the proposed method improves the total utility of all users by over 13% compared with a baseline that overlooks user preferences in service stacking.