Coordinated real-time power dispatch for integrated transmission and distribution networks based on model-based optimization and deep reinforcement learning

Abstract

As massive distributed energy resources are connected to the distribution networks, the distribution networks are gradually transforming into active distribution networks, and the transmission and distribution networks become more coupled. To make the most of these distributed energy resources and improve the safety of the coupled power system, coordinated real-time power dispatch is indispensable. And for privacy protection and actual engineering needs, the real-time power dispatch problem should be optimized in a decentralized manner by different control centers. In this paper, we decompose the global objective function into local ones, which can be locally optimized by control centers of transmission and distribution networks. The coordination is then realized by introducing the alternating direction method of multipliers (ADMM). Based on the decomposition, we design a novel communication architecture, in which only the boundary variables are exchanged, effectively protecting private information. Additionally, in the real power system, the transmission networks have credible models, while the maintenance of accurate models for distribution networks is usually unaffordable. Therefore, this paper also proposes a hybrid real-time power dispatch method: a model-based optimization method for the transmission network and model-free deep reinforcement learning for each subordinate distribution network. This hybrid method not only overcomes the model incompleteness of the distribution networks but also accelerates and stabilizes the learning process. Numerical test results on three different cases justify the effectiveness of the proposed communication architecture and hybrid real-time power dispatch method.