Ever wondered why a substation that passed its inspection last month suddenly trips offline?
Here’s the thing… those failures are almost never sudden.
The damage was silently brewing inside the insulation for months. Generating small electrical discharges. Corroding into the dielectric. And no one was listening.
Those small sparks are referred to as partial discharge, and they are one of the primary causes of utility substation switchgear failing well before the end of its designed life.
The good news?
Detection has been revolutionized. Instead of an annual crawl through with a handheld probe, you now have a system that listens 24/7 and alerts you to problems weeks before they explode.
What you’ll walk away with:
- What Partial Discharge Actually Is
- Why It Costs Utilities So Much Money
- How AI Diagnostics Quietly Took Over
- What This Means For Ageing Substations
What Partial Discharge Actually Is
Partial discharge refers to localized electrical discharge of solid insulation which does not completely bridge the space between two conductors.
Similar to a crack on your windscreen. It doesn’t mean it has failed but driving over bumps makes it incrementally worse until eventually it all shatters.
Inside switchgear, those “bumps” happen 50 or 60 times a second.
What causes it? Usually one of four things:
- Voids or air pockets left inside solid insulation
- Moisture and condensation getting into the enclosure
- Dust and pollution creating conductive paths across surfaces
- Sloppy installation work, especially on cable terminations
And that’s why enclosure design is so important. With sealed, gas-insulated solutions like Spike Electric GIS, moisture, salt and dust aren’t able to settle on live parts. This eliminates most of the conditions necessary for partial discharge to occur. When your utility substation switchgear is exposed to coastal, dusty or humid environments, a sealed compartment will do more for long-term reliability than any preventative maintenance schedule.
Why It Costs Utilities So Much Money
Partial discharge doesn’t just damage equipment. It takes the lights out.
According to testing specialists, 85% of all disruptive failures experienced by medium and high voltage equipment are due to partial discharge. Partial discharge can be directly attributed to approximately 25% of all MV switchgear failures.
Add to that the fact that the grid isn’t getting any younger. Approximately 70 percent of the country’s large power transformers are now over 25 years old. Much of the switchgear surrounding them is the same age.
Customers see those results in the outage statistics. The average U.S. customer endured 11 hours of interruptions in 2024, more than twice the 10-year average.
Not all of that is partial discharge, obviously. Storms did most of the damage.
But here’s the part that stings:
Insulation failures are the outages that should have been seen coming. The signals were blinking. No one was watching.
The Old Way Of Finding It
Traditional partial discharge testing works like this…
Technicians walk through the switchroom once or twice yearly with a TEV probe or an ultrasonic detector. Readings are recorded. High numbers get escalated.
That approach found plenty of faults. But it has three obvious holes in it:
It’s a snapshot. Partial discharge activity varies with load, temperature and humidity. Take a reading on a cool dry Tuesday morning and a real defect can look perfectly benign.
It’s noisy. Electrical dirties galore. Variable speed drives, fluorescent lighting and radio traffic all create signals very similar to PD on a simple detector.
It depends on the technician. Reading a phase-resolved pattern correctly requires years of training. There are not enough people who know how to read them and the experts are retiring.
So most utilities ended up in the same place. Data everywhere. Very little insight.
How AI Diagnostics Quietly Took Over
This is where things got interesting.
Machine learning excelled at the very task people have trouble with: distinguishing an actual flaw from background noise thousands of times per day without fatiguing.
There wasn’t a big roll out event. It silently snuck into substations one sensor at a time.
It Learns The Difference Between Noise And Trouble
The trained model can examine a discharge pattern to see what it is. Corona. Tracking along the surface. Internal voiding. Floating electrode.
Each defect type has its own fingerprint, and each one has a different urgency.
Surface tracking on a bushing may last you two years. An internal void in a cable termination may last you two months. Knowing the difference will change your entire maintenance approach.
It Never Stops Watching
Permanently installed sensors feed data continuously instead of once a year.
Why that matters, is more than it seems. Partial discharge activity that occurs only when there is heavy summer load, or only when humidity is high (when it rains), is going to be missed by periodic testing every time. Continuous monitoring detects it every time.
It Ranks The Fleet, Not Just The Alarm
Here’s the biggest shift…
Rather than pass/fail reading on a single panel, AI’s score every asset across an entire fleet then rank them by risk.
A large US utility deployed machine learning on 10,000 transformers and 22,000 circuit breakers, and experienced 48% fewer transformer failures across 15 months.
That is what winning looks like. Not less alerts. Knowing precisely which five assets, out of four hundred, should be your focus this quarter.
What This Means For Ageing Substations
Operators rarely have the luxury to rip and replace everything. Budgets just won’t allow it and lead times are horrific.
So the practical play has two parts.
Fully protect new equipment. Sealed and gas-insulated switchgear eliminates the source of moisture and contamination when upgrading switchgear or adding new bays.
Smartly monitor the old legacy stuff. Legacy AI panels are where PD resides. Legacy AI panels are also where continuous monitoring makes its money quickest.
Pair those two and asset replacement stops being a guessing game.
Where It Still Falls Short
AI diagnostics are not magic, and pretending otherwise causes problems.
Training a model on one manufacturer’s switchgear doesn’t mean it can be applied to another manufacturer’s. Sensor location is still critical – place the probe incorrectly and no amount of algorithmic magic will save the measurement.
The tools are good. They still need engineers who understand what the numbers mean.
The Bottom Line
Partial discharge has long been the silent killer of utility substation switchgear. It doesn’t cry out in pain as it weakens structures. You can’t see it on a thermal camera. It kills slowly enough that few pay attention until catastrophic failure occurs.
What has changed is the listening.
Continuous sensors coupled with machine learning changed a once annual snapshot into an ongoing dialogue with the machinery. Faults are categorized, prioritized and slotted for repair rather than identified during an outage.
To recap the whole thing:
- Partial discharge causes the majority of disruptive insulation failures
- Periodic handheld testing misses load and weather dependent activity
- AI models classify defect types and rank risk across an entire fleet
- Sealed enclosure designs stop the problem before it starts
The technology already exists. The question is just how many miscarriages occur before it’s commonplace.