Efficient neural network architecture for topology identification in smart grid

Abstract

Identifying arbitrary power grid topologies in real time based on measurements in the grid is studied. A learning based approach is developed: binary classifiers are trained to approximate the maximum a-posteriori probability (MAP) detectors that each identifies the status of a distinct line. An efficient neural network architecture in which features are shared for inferences of all line statuses is developed. This architecture enjoys a significant computational complexity advantage in the training and testing processes. The developed classifiers based on neural networks are evaluated in the IEEE 30-bus system. It is demonstrated that, using the proposed feature sharing neural network architecture, a) the training and testing times are drastically reduced compared with training a separate neural network for each line status inference, and b) a small amount of training data is sufficient for achieving a very good real-time topology identification performance.