Noisy PMU Data Recovery in Transient Conditions through Self-Attention Neural Networks

This paper utilizes the self-attention-based Imputation method to effectively manage missing data in Phasor Measurement Units (PMUs) during transient power system disturbances. This self-attention-based method processes multivariate, noisy datasets, improving data accuracy during disturbances under different missing data patterns and ratios. We conducted a comprehensive comparative analysis with other imputation methods using the IEEE 39-bus New England system. As inputs for the imputation, we employed voltage magnitudes and angles. Results demonstrate the superiority of this method in maintaining data integrity and significantly improving the accuracy of imputation under noisy and transient conditions. In comparative testing, this method reduced Mean Absolute Error (MAE) by approximately 5% to 50% across different cases compared to the best result from other methods in most scenarios, although it underperformed slightly in highly sparse data conditions with a missing ratio of 0.9. The method demonstrated robustness through its high imputation accuracy and fast performance, confirming that it is well-suited for real-time applications in smart grid monitoring, thanks to its ability to process data in parallel.