Decentralized Carbon-Aware Dispatch for Virtual Power Plants in an Unbalanced Distribution Network via Heterogeneous Reinforcement Learning

Environmentally sustainable operation is critical for distribution networks with high DER penetration. This paper proposes a decentralized carbon-aware virtual power plant (CAVPP) framework that jointly minimizes operational cost and carbon emissions while respecting network constraints. We introduce a carbon-emission flow (CEF) model that allocates demand-side emissions and, unlike prior single-phase approximations suffering from under/over-allocation, develop a realistic unbalanced DN formulation with nodal-phase emission tracing. The CAVPP coordination problem is recast as a cooperative Markov game and solved via a heterogeneous-agent soft actor–critic (HASAC) algorithm that accommodates diverse agent roles and dynamic action–observation spaces. Tested on the IEEE 123-bus unbalanced system, the proposed approach achieves an 8.76% reduction in operating cost and a 21.74% reduction in total carbon emissions relative to a state-of-the-art benchmark, demonstrating coordinated scalability and deployability for large, heterogeneous DER populations.