Learning-based distributed optimal power sharing and frequency control under cyber contingencies

Abstract

This paper proposes a learning-based distributed optimization framework to enhance the resilience of distributed power system algorithms against communication failures. The autoregressive-moving-average (ARMA) model is used to estimate the missing states and control actions of neighboring agents during communication contingencies. The proposed framework allows agents to predict the future control actions of neighbors. Thus, even during complete loss of communication, agents can efficiently perform distributed optimization. We use the Distributed Optimal Frequency Control (DOFC) algorithm, which includes optimal power sharing to achieve frequency stability, as a benchmark application platform to show the effectiveness of the proposed framework. The theoretical findings are evaluated on two practical power systems. The results show that the ARMA-based DOFC algorithm can asymptotically reach the same convergence rate as the power system without communication interruptions.