Applying Neural Networks in Measuring Partial Discharges in MV Power Cables

Cables & Accessories

Partial discharge diagnostics is a fundamental process in all modern utilities and progressive research into the physics of PDs has enabled development of monitoring systems with high resolution and reliability. For example, within Singapore Power’s Grid, accretion of spot testing and online monitoring diagnostic efforts have yielded favourable results and continue to expand. Given advancements in this technology, confidence in monitoring systems and competency in expert analysis, there is a good prospect for automation in measurement analysis.

Fig. 1: Overview of PD analysis.

Artificial intelligence based automation, while often oversimplified, can be de-composed to respective stages in order to iteratively construct a dependable assessment structure. As shown in Fig. 1, PD analysis can be anatomized to the following. It is recognized that the requirements for online and offline testing, while comprised of the same roots, contain deviations amongst one another.

Time-Resolved Partial Discharge

Fig. 2: Three pulse PD measurement in offline cable test.

Fig. 2 shows a conventional three-pulse time-resolved partial discharge (TRPD) measurement from a single PD source in an offline cable PD test. The PD direct (PDD), first reflection (FR) and second reflection (SR) pulses are as annotated in the image, with the FR and SR pulses indicative of the discharge location and cable end respectively. Propagation along the cable results in pulse attenuation and dispersions, which is evident from the successive reduction in magnitude and increment in width.

Characteristics of PD measurements is contingent on:

1) cable specification, length and amount of accessories;

2) measurement conditions;

3) noise interference, gain and DC offsets.

Field measurements are particularly affected by noise conditions, which can introduce substantial variations.

Fig. 3: PD measurements with, a) non-stationary noise, b) huge repetitive noise, c) incorrect trigger and d) FR pulse corrupted by noise.

In field cable testing, noise interference is an indisputable factor of concern and unpredictable in nature. For example, it could contain non-stationary properties (as shown in Fig. 3a) or repetitive in frequency (as shown in Fig. 3b). It could also cause incorrect triggering of the measurement (as shown in Fig. 3c) or corrupt the FR pulse (as shown in Fig. 3d).

Attend the 2022 INMR WORLD CONGRESS in Berlin, where Dr. Joel Yeo Wei Wen, an expert in condition monitoring with Singapore Power Group will discuss neural networks coupled with feature engineering applied to PD analysis from offline cable testing, more specifically, time-resolved partial discharge measurements. These measurements contain sequentially two elements of analysis: identification between PD and noise; and localization of discharge location along the cable. He will also explain how successful interpretation relies on accurate identification.