Data-Driven Volt/VAR Optimization for Modern Distribution Networks: A Review

Abstract

The Volt/Var optimization (VVO) enables advanced control strategy development for voltage regulation. With the recent advancement of data-driven approaches and communication infrastructure, realtime decision-making through VVO can effectively address distributed energy resources (DERs) uncertainties without relying on models and topologies of distribution networks. In this paper, a comprehensive review on data-driven VVO in distribution networks is presented, focusing on statistics and machine learning (supervised/unsupervised, ensemble, and reinforcement learning (RL)). State-of-the-art monitoring devices essential in data-driven VVO frameworks are firstly discussed. How data-driven structures serve as primary or supplementary tools in VVO frameworks is then detailed. Since RL is increasingly used, RL-based algorithms (value-based, policy-based, actor-critic-based, and graph-based algorithms) are reviewed. Decision-making processes for RL-based VVO frameworks, such as the Markov decision process (MDP), Markov game, constrained Markov decision process, constrained Marko game, and adversarial Markov decision process, are also surveyed. Future research directions in this area are recommended in the paper.

Keywords

• Data-driven decision-making
• distribution networks
• supervised
• unsupervised
• ensemble learning
• reinforcement learning