Cost minimization in energy communities with multi-agent deep reinforcement learning and linear programmingPokorn, Matic (Avtor) Čampa, Andrej (Avtor) Smolnikar, Miha (Avtor) Mohorčič, Mihael (Avtor) Hribar, Jernej (Avtor) With energy costs on the rise and with ever growing concern for environmental impact, energy providers and regulatory bodies have been pushing for dynamic energy prices as a means to encourage load shifting to reduce daily energy demand variance. Coupled with recent advancements in photovoltaics (PV) power generation and Battery Energy Storage System (BESS) technology, this has encouraged the development of energy communities with one of the goals to mitigate the effect of dynamic prices on homeowners’ energy bills without sacrificing comfort, while at the same time utilizing aggregation of Distributed Energy Resources (DER) to contribute to grid flexibility. In this paper, we present Mathematical Optimization and Deep Reinforcement learning for Energy Cost minimization (MODREC), a decentralised Community Energy Management System (CEMS). MODREC leverages Multi-Agent Deep Reinforcement Learning (MADRL) coupled with Linear Programming (LP) to minimize cost in an energy community by intelligently charging and discharging household BESSs while assuming non-elastic consumer loads. MODREC follows an LP-guided training pipeline, where an optimal strategy computed with LP on historical data is employed to train a set of Deep Reinforcement Learning (DRL) agents, each assigned to a household in the community, that minimize a common cost function. Our main contribution lies in the system-level integration of LP-derived expert supervision with decentralized multi-agent control for community energy cost minimization under dynamic pricing. With MODREC, we manage to save up to 30% of energy costs compared to conventional approaches and efficiently shift energy load to off-peak hours.IEEE20262026-03-27 14:46:09Neznano28672UDK: 004.8ISSN pri članku: 2169-3536DOI: 10.1109/ACCESS.2026.3676334COBISS_ID: 273377283ZDAsl© 2026 The Authors.