Fast GRNN-Based Method for Distinguishing Inrush Currents in Power Transformers

Abstract

Differential protection, as the key protection element in the power transformers, has always been threatened with sending false trips subjected to external transient disturbances. As a result, differential protection needs an additional block to distinguish between internal faults and external transient disturbances. The protection system should i) be able to perform based on raw data, ii) be able to learn fully temporal features and sudden changes in the transient signals, and iii) impose no assumption on noise. To address these challenges, a fast recurrent neural network, namely fast gated recurrent neural network (FGRNN). By removing reset gate in the gated recurrent unit (GRU), the proposed network is capable of learning abrupt changes in addition to significantly reducing the computational time. Furthermore, a loss function based on an information theory concept is formulated in this paper to enhance the learning ability as well as robustness against non-Gaussian/Gaussian noises. A generalized form of mutual information is also adopted to form a noise model-free loss function, then incorporated with the designed deep network. Simulated and experimental examinations engaging various external factors, in addition to comparison between the proposed fast GRNN, GRU and seven firmly-established methods indicates the faster and more reliable performance of the proposed algorithm.

Deep Learning-Based Algorithm for Internal Fault Detection of Power Transformers during Inrush Current at Distribution Substations

“…It captures 2D feature space as a pattern recognition for each abnormal condition. Methods based on fuzzy and artificial neural networks were proposed in [10,11], and a correlation-based algorithm was developed for inrush current discrimination [12]. For a similar purpose, a method combining a support vector machine as the classifier and a wavelet transform for feature extraction was also proposed in [13].…”

Spatial‐temporal learning structure for short‐term load forecasting

“…Although hybrid methods show satisfactory performance, a range of different techniques lead to disability in presenting a general solution for load time series, in distribution level. Moreover, hybrid methods are very costly in terms of computational complexity and hardware in the practical implementation [22].…”

Countermeasure to Prevent Transformer Differential Protection From False Operations

“…Category 5: Methods based on artificial intelligence algorithms, including artificial neural networks [21], [22], [23], [24], [25], fuzzy control [26], random forest [27], and K nearest neighbor-genetic [28]. Their reliability depends on the effectiveness of the used algorithms, the completeness of the training samples and the empirical knowledge.…”