A Novel Deep Recurrent Neural Network for Assessment of Role of Space Charge Effect in Partial Discharge Recognition and Diagnosis

Abstract

Partial Discharge (PD) has evolved into an inevitable tool for diagnosis of insulation of power equipment due to its inherent non-intrusive testing methodology. Since accurate recognition of complex overlapped PD sources is essential for effective diagnosis, recently the focus of research has shifted to challenges related to real-time PD measurement that involve complexities in discriminating multi-source pulse signatures, variations in applied voltage, dynamics of PD patterns related to space charge effect etc. Though research studies have successfully utilized a gamut of machine intelligence techniques such as neural networks, Hidden Markov Models, Support Vector Machines etc, effectiveness of recognition of complex overlapped PD sources based on space charge effect associated with PD has not been comprehensively established. This research focuses on implementation of Deep Recurrent Neural Network (DRNN) as a novel strategy for assessing the role of space charge and its associated memory propagation effect in PD signatures. Since Long Short-Term Memory (LSTM) based DRNN architecture augurs well for data involving sequence similarity recognition, the objective of this research is on establishing an indigenous approach of formulating identification markers to ascertain the role of space charge in PD patterns. The second objective is on establishing a unique approach to decipher state transitions of pulses using transition labels that describes the dynamics of PD signatures due to space charge effect. Detailed case studies based on laboratory benchmark models that replicate complex PD patterns clearly demonstrate the excellent capability of DRNN in deciphering the role of space charge in discriminating complex PD patterns.