Learning from Future: Prediction-based Data Augmentation to Enhance Power Grids Fault Detection

As power grids modernize to include wide area monitoring and advanced metering infrastructures, the applications for data-driven methods based on artificial intelligence (AI) for situational awareness, reliability, and security continue to grow. However, obtaining effective training data that captures normal and anomalous operations in power systems remains a great challenge. This paper proposes a data-augmentation method that can efficiently build effective training data sets, benefiting various data-driven applications used in power grids without increasing the size of training data set. By leveraging load prediction in smart grids, we obtain the knowledge of future operating conditions and potential anomalies, which are integrated with the historical data in the training data set. By utilizing this augmented training data set, we significantly increase the accuracy of data-driven fault detection, e.g., 8.6% on average, compared to the results trained based on historical data only.