Deep-Learning Based Power System Events Detection Technology Using Spatio-Temporal and Frequency Information

Abstract

Power system events such as load disconnection (LD), generator trip (GT) and short circuit fault, affect the stability of power grids. It is significant for grid operators to acquire the classifications and locations of events quickly and accurately, which can help prevent blackouts and economic losses. Utilizing phasor measurement unit (PMU) data, this paper proposes a two-layer algorithm that monitors the power system state and identifies types and locations of events. In the first layer, with Gramian Angular Field (GAF) and Short-Time Fourier Transform (STFT), PMUs data are mapped to 2D images containing frequency information of the events, and then frequency features are extracted to classify the events by Convolutional Neural Network (CNN) and Long Short-term Memory (LSTM). Based on the results of the first layer, the second layer further determines the locations of the events. Graph Attention Network (GAT) is used to extract spatial information in the power system topology, which is insensitive of event types. CNN can enhance the feature gap between different events, and LSTM can extract temporal information from the PMU data. With the extracted temporal-spatial information, the second layer can produce the locations of events with high accuracy. The proposed two-layer algorithm is tested in the IEEE Standard New England 39 Test System. Compared to benchmark models, the proposed method has higher event identification and location estimation accuracy and stronger noise resistance.