Artificial Intelligence for multi-agent peer-to-peer energy trading in dairy-farm microgrids

The global energy transition is driving a rapid rise in distributed energy resources (DERs) and prosumers, making peer-to-peer (P2P) energy trading a key paradigm for local coordination. This thesis investigates how multi-agent systems (MAS) and multi-agent reinforcement learning (MARL) can support P2P trading in dairy-farm microgrids, with a particular focus on increasing revenues from electricity sales and reducing electricity bills and peak-hour demand. First, a survey of MAS-based P2P trading identifies dominant architectures, market designs, and open challenges, highlighting the lack of customization for domains such as agriculture. Next, a MAS framework and P2P market tailored to dairy farms is developed, demonstrating reduced energy costs, higher renewable self-consumption, and lower electricity demand from the grid during peak hours, while also revealing the limitations of static, hand-crafted agent strategies. To address these limitations, MARL is introduced to jointly optimize trading and local resource scheduling using value-based and policy-based reinforcement learning algorithms. MARL agents outperform heuristic MAS strategies in cost reduction and flexibility, yet remain vulnerable to forecast errors and environmental uncertainty. The thesis then proposes an Uncertainty-Aware Knowledge Transformer for MARL, which embeds probabilistic forecasts and domain knowledge into agents’ state representations. This transformer-based module enables agents to better handle uncertainty in renewable generation, demand, and prices, significantly improving the robustness, risk sensitivity, and fairness of P2P trading outcomes. Overall, the work demonstrates that MAS provide the structural backbone for decentralized coordination, while MARL delivers adaptability, provided that uncertainty is modeled explicitly. The results have practical implications for dairy-farm operators, distribution system operators (DSOs), and policymakers, and point toward future research in safe, constrained, and human-centric MARL for real-world energy systems.

Publisher

University of Galway

Rights

CC BY-NC-ND

cbn

Collections

University of Galway Theses (PhD Theses)