We are pleased to announce a new scientific publication produced within the scope of the STREAM project, in collaboration with the Jožef Stefan Institute and our partner ComSensus.
The paper, titled “Cost Minimization in Energy Communities With Multi-Agent Deep Reinforcement Learning and Linear Programming”, has been published in IEEE Access in March 2026.
What the paper is about
The authors present MODREC (Mathematical Optimization and Deep Reinforcement learning for Energy Cost minimization), a novel decentralised Community Energy Management System (CEMS). MODREC combines two complementary techniques, Linear Programming (LP) and Multi-Agent Deep Reinforcement Learning (MADRL), to intelligently manage Battery Energy Storage Systems (BESS) across a community of households equipped with photovoltaic generation and dynamic energy pricing.
The key idea is a two-step approach: LP is first used on historical data to compute an optimal “expert” strategy, which is then used to train a set of DRL agents, one per household, capable of making real-time decisions without requiring forecasts of future prices or solar generation. The result is a system that is non-intrusive (no appliance scheduling), resilient to data loss, and scalable to communities of different sizes.
Key results
Up to 29% reduction in community energy costs compared to a conventional baseline. Successfully validated on communities of 5, 10, and 20 households. Effective load shifting to off-peak hours, contributing to grid flexibility. Robust performance even with up to 40% of input data corrupted.
This work directly supports STREAM’s mission of enabling flexible, data-driven energy management on the low-voltage grid side, empowering energy communities to actively participate in flexibility markets.
Read the full paper here.