DEATS: Decentralized Energy Agent Trading System for Intelligent Local Energy Management and Dynamic Pricing

Abstract

This thesis proposes a physically-aware multi-agent deep reinforcement learning (MARL) framework for decentralized P2P energy trading in interconnected microgrids. The proposed solution adopts a dual-layer architecture that integrates a decision layer with a physical layer. In the decision layer, a Multi-Agent Deep Deterministic Policy Gradient (MADDPG) algorithm is implemented under a Centralized Training with Decentralized Execution (CTDE) paradigm to learn optimal trading and storage policies. This is coupled with a physical layer that enforces AC power flow constraints on IEEE 33-bus distribution systems.In addition, a genetic algorithm (GA) is employed for optimal placement of distributed energy resources to ensure grid readiness under dynamic operating conditions. Simulation results demonstrate that the proposed framework achieves a 28.3% reduction in energy costs while maintaining full compliance with network constraints. In contrast, unconstrained reinforcement learning achieves higher economic gains (37.5%) but results in critical voltage violations and line overloads. Furthermore, the system ensures fair participation among microgrids, achieving a Jain’s fairness index of 0.92. These results highlight the importance of incorporating physical grid constraints into learning-based energy trading, enabling economically efficient and operationally reliable decentralized energy markets.