Fig. 2: The learning framework developed which includes the environment and the policy network architecture with graph neural network (GNN) based feature abstraction.
From: Real-time outage management in active distribution networks using reinforcement learning over graphs
The environment is composed of a distribution network modeled using OpenDSS with a Python-based API depicted as the blue-colored blocks. The graph replica of the DSS circuit and the corresponding evaluations in the graph domain are represented using orange-colored blocks. Voltage sources are introduced at virtual slacks specifically for intentional islanding scenarios. The policy network uses the state information from the environment to compute graph node embeddings and context node embeddings using GNN and feedforward networks, respectively. Both the policy and value networks encompass the RL network module, color-coded in light red. The GNN-based graph abstraction is further elaborated in the green section. A final feature vector that encompasses the two embeddings is computed by a multi-layer perceptron (MLP).
