Learning Power System Dynamic Signatures using LSTM-Based Deep Neural Network: A Prototype Study on the New York State Grid

In this paper we present a prototype study of classifying dynamic events using a deep learning (DL) tool for the New York State (NYS) power grid. We use the utility-level full-scale transmission planning model of Eastern Interconnection (EI) in the PSS/E platform for generating simulation data that are used for learning. We specifically use simulation data from those bus locations where Phasor Measurement Units (PMUs) are installed in the actual NYS grid for faster event detection and decision making. We first conduct the coherency study of the NYS grid using a unsupervised learning technique, namely Principal Component Analysis (PCA). Then we consider three different dynamic classification scenarios. We classify the loss of generation scenarios according to the coherent clusters, loss of loads according to the New York Independent System Operator (NYISO) zonal separation and localize the loss of transmission lines along the Central-East (CE) interface. For the transmission line outage scenario, we have generated dynamic event data by changing the grid stress level through the said interface and then tripping different major lines. The deep learning tool used for this study is known as the Long Short Term Memory (LSTM) neural networks. This work is intended toward developing a PMU-based data-driven fast contingency analysis and decision making architecture for the next generation power grid.