PQEventCog: Classification of power quality disturbances based on optimized S-transform and CNNs with noisy labeled datasets

Power quality disturbances (PQDs) recognition is a vital topic for smart grid governance. Previous work has achieved remarkable success by promoting signal processing and classifier models. However, for some power grids, it is difficult to obtain well-labeled samples for training. Label noise is ubiquitous, which would fail to assess PQDs due to data distribution discrepancy. This article proposes a hybrid approach called PQEventCog for PQD assessing. Firstly, an improved variational mode decomposition is utilized for signal reconstruction. Then, S-transform combined with singular value decomposition is imported to transfer a time-series signal into a 2D time-frequency image for features enhancement. As a potential tool, Differential Training is applied to reduce label noises of training sets. Finally well-labeled samples are fed into a CNN model. A set of analytical signals, as well as real data in a microgrid platform, are performed to confirm the effectiveness and the excellent label-noisy tolerance of PQEventCog.