LAUSR

Dynamic network reconfiguration (DNR) and volt-VAR control (VVC) are widely used techniques for the secure and economic operation of active distribution networks (ADNs). Their significance is rising unprecedently due to the increasing integration of renewables in ADNs. This paper hence proposes a bi-graph neural network (BGNN) modeling-based deep reinforcement learning (DRL) framework for effective DNR-VVC real-time coordination featured by high-dimension decision space and complex system dynamics. Specifically, the Gumbel-softmax soft actor critic (GSSAC) algorithm is proposed to effectively decompose the vast discrete decision space resulting from numerous DNR-VVC devices. Its learning efficiency is enhanced by a proposed automated entropy annealing scheme. BGNN is then designed to fully capture both line and bus dynamics of ADNs to further boost coordination performance. Experiments are conducted on several modified ADNs to compare with various benchmarks. Results demonstrate that GSSAC-BGNN can achieve competitive performance for the secure and economic operation of ADNs with a fast decision speed and is superior in managing switching and tapping actions to benefit operators in maintenance cost reduction.