Visualizing Graph Neural Networks In Order To Learn General Concepts In Power Systems

Neural network-based models play an increasingly important role in many complex decision-making systems. However, their lack of interpretability make it difficult to analyze and trust them when applied to critical infrastructure like the power system. This paper presents visualizations and exploratory analysis of the internal representations of a graph neural network based reinforcement learning agent applied to a power system reliability study. By using low-dimensional t-distributed stochastic neighbor embedding, we show how agents with different generalizing capabilities have different internal representations. We study how the inputs to the agent are processed through the layers of the neural network components. Additionally, we use these visualizations to indicate that the agent can learn general concepts of the power system, and can be expected to scale well to larger power systems. We also show how the agent behaves in a grid expansion scenario with a power system not experienced during training.