Identifying network harmonics from condition data
A common health monitoring technique for high voltage equipment is to measure small electrical currents called partial discharge (PD) in the insulation. The pattern of PD can be used to diagnose the type of defect present, which allows the utility to plan the correct repair. However, it has recently been discovered that harmonics on the high voltage network can also alter the PD pattern. Harmonic monitoring is expensive, but an accurate diagnosis can’t be made unless the harmonics are known.
The Data Science
I recognised that particular harmonics distort the PD pattern in predictable ways. Also, certain harmonics are more likely to be present on the network that others, with the 5th and 7th harmonics particularly common. Pattern recognition techniques are typically used to diagnose the defect, so I could use the same approach to diagnose the presence of key harmonics.
Since classifying harmonics is different from classifying defects, the most informative features in the data may be quite different. I generated a large set of features, and used a process of feature selection to identify those that contributed best to harmonic classification. I built C4.5 rule induction classifiers to identify the presence of the 5th and 7th harmonics, with moderate accuracy on unseen data (around 60%).
Results
While the accuracy on a single PD pattern was modest, a single pattern represents just 80ms of data. Typically, PD will be recorded over a few minutes to a few days, so a conclusion about harmonics can be drawn from a number of patterns in a row. I tried to increase the accuracy by using majority voting on sequential patterns. This worked, with accuracy tending towards an asymptote as batch size increases.
Resources
Papers detailing this work: