Knowledge‐based artificial neural network for power transformer protection. Issue 24 (15th September 2020)

• Record Type: Journal Article
• Authors: Li, Zongbo; Jiao, Zaibin; He, Anyang
• Is Part Of: IET generation, transmission & distribution. Volume 14:Issue 24(2020)
• Page Start: 5782
• Page End: 5791
• Publication Date: 2020-09-15
• Subjects: neural nets -- power transformer protection -- feature extraction -- power engineering computing -- generalisation (artificial intelligence) -- knowledge based systems; S:ANN -- T:ANN -- knowledge‐based ANN -- knowledge‐based artificial neural network -- power transformer protection -- artificial intelligence -- feature transferring -- equivalent magnetisation curve -- data‐driven -- feature extraction; Electric power production -- Periodicals; Electric power transmission -- Periodicals; Electric power distribution -- Periodicals; 621.3105
• DOI: 10.1049/iet-gtd.2020.0542
• Languages: English
• ISSNs: 1751-8687

Abstract

Abstract : Data‐driven and artificial intelligence based transformer protection has attracted increasing attention but not been widely applied in the power system owing to the poor generalisation ability. In this study, a feature transferring method is proposed for a knowledge‐based artificial neural network (ANN) to develop a transformer protection with an improved generalisation ability. Normally, power experts can reliably identify the running states based on the professional knowledge of only focusing on the unsaturated parts of equivalent magnetisation curve (voltage of magnetising branch‐differential current). In order to imitate the power experts, the images of equivalent magnetisation curves whose saturated parts are removed are defined as source domain and the original samples are target domain. An ANN named as S:ANN is firstly trained through the source domain where the extracted features are equivalent to the professional knowledge. Then another ANN with the same structure as S:ANN is trained through the target domain and named as T:ANN. It is specially designed for T:ANN that adaptive layers are employed between S:ANN and T:ANN to reduce the feature differences. Finally, simulations and experiments reveal that the knowledge‐based ANN namely the determined T:ANN shows a better generalisation ability through paying more attention to the unsaturated parts.