Coordinated energy management of integrated multi-carrier microgrids using attention-enhanced MADDPG

The increasing penetration of renewable energy and the strong interdependence among electricity, gas, heat, and cooling networks make the real-time operation of integrated multi-carrier microgrids (IMCMGs) increasingly complex, particularly under renewable generation and demand uncertainties. To address these challenges, this study proposes an attention-enhanced multi-agent deep deterministic policy gradient (AMADDPG) framework for coordinated energy management. The proposed framework adopts a centralized training and decentralized execution (CTDE) paradigm, while an attention mechanism embedded in the critic network identifies the most relevant inter-agent and inter-carrier interactions, thereby improving coordination and learning stability. Simulation results over a three-IMCMG test system demonstrate that the proposed framework achieves reliable multi-carrier energy coordination and robust operational constraint satisfaction under bounded renewable-generation deviations. Across the evaluated operating scenarios, AMADDPG achieves total operating cost reductions of 54.12%, 45.78%, and 14.31% compared with stochastic programming via particle swarm optimization (SP(PSO)), centralized DDPG, and standard MADDPG, respectively. Furthermore, integrating electrical energy storage (EES) under dynamic electricity pricing reduces the daily operating cost from 246018.051 to 243909.732 (an absolute saving of 2108.319, or 0.86%), reflecting the operational value of storage-enabled load shifting. These results highlight the effectiveness and cost efficiency of the proposed learning-based dispatch framework for modern multi-energy systems.