Data anomaly detection in phasor measurement units using deep learning

What is it about?

Identify false data injection attacks from PMU packet data streams Features are extracted from the PMU buffer to train multiple deep machine learning models This paper is concerned with a class of false data injection attacks (FDIA) which aim to modify PMU measurements resulting in incorrect state estimation solutions. We extract multi-variate time-series signals from PMU data packets aggregated in phasor data concentrators (PDC) corresponding to different events such as line faults and trips, generation and load fluctuations, shunt disconnections and FDIA prior to every cycle of state estimation (SE). A Convolutional Neural Network (CNN) data filter with Nesterov Adam gradient descent and categorical cross entropy loss is proposed to validate the PMU data. This filter extracts inter time-series relationships to classify different power system events by comparing the temporal structure of PMU packet data. The performance of the filter is then compared with (a) deep learning algorithms such as Recurrent Neural Networks (RNN) and Long Short Term Memory (LSTM) and (b) traditional classifiers such as SVM and ensemble methods. It is seen that the proposed CNN-based filter results in higher classification accuracies among all classifiers. This makes the CNN classifier suitable to serve as an independent data filter to identify falsified data streams targeted to alter SE. In order to verify the accuracy of the proposed filter, a hybrid state estimator (HSE) has been used in this study which obtains measurements from both PMU and traditional meters.

Featured Image

Photo by sterlinglanier Lanier on Unsplash

Photo by sterlinglanier Lanier on Unsplash

Why is it important?

Serve as an independent data filter to identify falsified data streams prior to each cycle of State Estimation Early warning system