Grid Inspection by the Numbers: What Dominion, FirstEnergy, and ESO Actually Measured

Will Drewes

Founder, Fern Strategy · 3 min read

The inspection backlog is a data problem

Most transmission and distribution assets get inspected on a cycle measured in years. When crews do go out, the bottleneck is rarely the field work. It's what happens after: sorting thousands of images, classifying defects, updating asset records, and turning all of that into a maintenance plan. Three utilities ran structured programs to attack that bottleneck with AI-assisted inspection. The numbers are worth looking at directly.

Dominion Energy: volume and downstream planning

Dominion's transmission team processed 160,000 images in 15 hours during an initial inspection run, then brought that down to 74,000 images in 3.5 hours in a follow-on effort. Detection accuracy for key asset attributes came in at more than 85%, with inspection analysis time cut by roughly 70%.

The number that actually matters for operations: the faster turnaround directly informed the addition of 38 new towers to Dominion's 2025 paint program. That's not a throughput metric. That's a maintenance planning decision that wouldn't have been made on the same timeline with manual review. The inspection output fed a real capital allocation.

FirstEnergy: inventory accuracy and wasted trips

FirstEnergy ran a distribution inspection pilot focused on pole identification and geolocation. The program identified more than 80% of poles on the first pass, located to within 20 feet GPS accuracy. In total, 1,650 poles were geolocated at that precision, and the image database grew fivefold to roughly 5,000 images.

The operational problem this solves is straightforward. Crews dispatched to a pole that isn't where the records say it is waste time and money. An asset inventory that's wrong at the location level also makes defect tracking unreliable. FirstEnergy flagged specific defect types (vegetation overgrowth, corrosion) during the same workflow, which means the inspection pass was doing asset inventory and condition assessment simultaneously.

ESO: classification accuracy on components and insulators

ESO, a Lithuanian energy provider, published accuracy figures from automating what had been manual inspection work. Component detection and classification came in at 96% accuracy. Insulator defect classification reached 92% accuracy.

Insulators are worth calling out specifically. They're a high-consequence failure point, and defect classification on them is visually subtle. A 92% accuracy rate on insulator defects during automated review is a meaningful bar, particularly for a workflow that previously depended on trained inspectors working through images manually.

What utility operators should take from this

These three programs share a structure worth noting. None of them replaced field crews. All of them attacked the analysis layer: the work that happens between collecting inspection data and making a maintenance decision.

That's the right target. Field collection is already reasonably efficient. The bottleneck is turning raw imagery into actionable asset intelligence at scale, and doing it fast enough to influence the planning cycle that's already running.

For regulated utilities, there's a governance layer here that the headline numbers don't capture. An AI system that classifies 74,000 images needs an audit trail: which model version classified which asset, what confidence threshold triggered a defect flag, and who reviewed the output before it influenced a capital decision. Dominion's 38-tower paint program addition is a clean example of AI output driving a budget line. Regulators and internal audit teams will eventually ask how that decision was documented.

Building that audit foundation before the inspection program scales is cheaper than retrofitting it after the fact. The accuracy numbers from these programs are strong enough that the question for most utility operators has shifted from "does this work" to "can we defend how it works."