Self-healing power systems using reinforcement learning over graphs for controlled grid islanding

Natural disasters and sudden faults create rapidly changing conditions that significantly challenge the power system's resilience. Intelligent algorithms can enable operators to take informed actions for rapid restoration by suggesting efficient switching strategies to enhance the resilience of power grids against extreme events. The intentional islanding restoration strategy isolates vulnerable areas and identifies self-sustaining subsystems to prevent system collapse and minimize risk exposure to extreme events. In this paper, we develop a graph-based reinforcement learning (GRL) model to design AI-assisted switching in transmission networks that mitigate risks by strategically isolating affected areas for self-healing during power outages. To train the AI agent with an awareness of the transmission network's topology for decision-making, the adjacency graph of the transmission network is mapped to the convolutional network of the reinforcement learning model. The intentional controlled islanding problem is modeled as a Markov decision process, where the optimal switching policy is learned using the GRL approach. The dynamic model of the transmission network used in training the agent incorporates generator inertia and load behavior for stability and frequency control, while also reducing power flow mismatches within the formed islands. The effectiveness of this framework is demonstrated using the modified IEEE 118-bus network and validated using dynamic simulations.