Coordination optimization of multi-micro-network and distribution network based on multi-agent reinforcement learning

A distributed coordination method based on hybrid multi-agent reinforcement learning is proposed to solve the problem that the multi-micro-network and distribution network cooperative optimization with discrete devices makes it difficult to solve the problem of efficient computation of non-convex models. By constructing an optimal scheduling model that considers the constraints of regulating resource operation and system security, and introducing multi-agent technology, two agent groups of the distribution micro-network are designed respectively, and the centralized scheduling model is transformed into a distributed optimization problem between cooperative and competitive mixed multi-agents. Then, an improved multi-actor-attention-critic algorithm is proposed to train discrete and continuous action agents at the multi-microgrid-distribution network level to avoid dimension explosion caused by too high state action dimension and realize extensible learning.