•Successful field demonstration of distribution system controls such as Volt-VAR control and dynamic distribution network reconfiguration have been reported by many electric utilities. However, there are still many barriers to the wide-spread adoption of the technology. One of the most significant barriers is the lack of robust distribution network topology and parameter information, which are required in model/optimization based distribution system control approaches.

•To overcome the drawbacks of optimization based approaches, we formulate the distribution system control problems as Markov Decision Process (MDP) and Constrained Markov Decision Process (CMDP). We propose safe, sample efficient, near-optimal, and robust deep reinforcement learning algorithms to solve the MDP and CMDP problems. 

•Numerical validation results on IEEE distribution test feeders show that our proposed deep reinforcement learning algorithms achieve similar level of performance as model-based approaches without relying on accurate or complete distribution network topology and parameter information.