A Directed Acyclic Graph Neural Network for AC Optimal Power Flow

AC optimal power flow (OPF) is of great significance for power system security, reliability, and economy. As an NP-hard problem, its solution can be time consuming by traditional optimization techniques. For more efficient AC OPF algorithms, a Direct Acyclic Graph Neural Network (DAG-NN) is proposed in this paper, which enables an explicit design of a neural network utilizing the intrinsic structural information of the problem to be solved. The approach first reformulates an iterative Newton-Raphson based AC OPF algorithm as a compositional function, accordingly constructs a DAG, and then designs the neural network by realizing its each node by a shallow neural network. The paper also analyzes errors of the DAG-NN. The proposed approach is tested on a modified PJM 5-bus system.