Hierarchical Combination of Deep Reinforcement Learning and Quadratic Programming for Distribution System Restoration

This paper proposes a model for hierarchical combination of deep reinforcement learning (DRL) with quadratic programming for distribution system restoration after major outages. In the proposed model, optimal power dispatch of a collection of distributed energy resources, called integrated hybrid resources (IHRs), is determined by a DRL-trained controller, while a grid-level quadratic programming problem checks grid constraints and performs critical restoration operation. DRL is implemented using Soft Actor-Critic (SAC) algorithm, which is shown to outperform the common Deep Deterministic Policy Gradient in continuous action spaces. The numerical studies, performed on the 123-bus test distribution system, demonstrates that the hierarchical combination of DRL and quadratic programming not only speeds up the local operation of multiple IHRs, but also ensures that the network constraints are satisfied during the restoration operation.

Automated summary

This paper proposes a model that combines deep reinforcement learning (DRL) with quadratic programming for distribution system restoration after major outages. The model uses a DRL-trained controller to determine optimal power dispatch of distributed energy resources and a grid-level quadratic programming problem to check grid constraints and perform restoration operation. The implementation of DRL using the SAC algorithm surpasses the Deep Deterministic Policy Gradient in continuous action spaces. Numerical studies on a test distribution system demonstrate that this hierarchical combination accelerates local operation and ensures network constraint satisfaction during restoration.