AI-powered vegetation management, coming a lot sooner than you may think
Updated: Jan 4
AI-based asset maintenance is difficult, but well understood.
Anyone working in any asset-intensive industry will know about using machine learning. In the power industry, it’s most commonly used in generation assets. That’s not surprising. Modern turbines are shipped with hundreds of sensors as standard. The time series data they create is highly structured, well understood, stored in a historian, relatively easy to access, and existing software can straightforwardly analyse the data. It’s also not too difficult to retrofit older assets with sensors, as and when required.
The hard work starts defining the boundaries of an asset’s permissible behaviour. But these are typically quantifiable parameters for temperature, pressure, vibration, decibels, frequency, etc. And the hard work gets really hard deciding whether a breach of these parameters actually signifies a problem. False positives are a pain—you shut your turbine down because the algo tells you to, only to find there’s nothing wrong. False negatives are even worse: you put your faith in the algo only for you to have an unplanned outage, an uncomfortable conversation with the market operator, and large, unnecessary costs.
But, with careful and considered use of learning datasets, machine learning algos steadily improve and are already delivering significant value. Millions have been saved in the shift from scheduled to predictive and prescriptive maintenance.
AI-based vegetation management is harder by orders of magnitude …
The latest target for AI-based maintenance is vegetation management: the process of clearing vegetation from power lines. Veg management costs utilities billions each year, because it is more or less a manual process, reliant on schedules and the innate knowledge of line engineers and arborists. If ever there were a process in need of automation … even before PG&E’s move into bankruptcy protection brought vegetation management to every utility CEO’s attention.
But—and this is a big ‘but’—it’s going to be a really tough nut to crack. It’s really no surprise that it’s only now, when utilities realise that poorly-executed vegetation management can cause an existential threat, that the sector is waking up to the art of the possible.
… because there is scant available data …
Unlike turbines, there is no cheap, abundant source of structured, easily-analysable data. A turbine spits out terabytes of structured data each year, from hundreds of sensors, connected to an easily-accessed data historian. Not so for vegetation management. A network survey requires a plane to fly over power lines, recording video and images as it goes. Consequently, they are expensive to perform. Most network utilities are lucky to get an annual survey of only part of their lines. Once a survey has been performed, it usually falls to an expert to manually scan images to identify any potential risk of vegetation encroachment, or dying or dead trees. The problem with experts is that, by definition, they are expensive and, because to err is human, still manage to get things wrong.
… which is difficult to interpret.
In a turbine, exceptions are quantifiable and easily defined; exception parameters for vegetation encroachment are subjective and fuzzy. AI-based veg management uses video and image data, rather than structured numeric data. It is a quantum leap from defining what is too hot or too noisy in a turbine to what plant life presents a risk to power system integrity.
For example, vegetation encroachment parameters define ‘how close is too close?’ Initially, these measurements will be constant, irrespective of tree species. A crew will be sent out if vegetation grows within a predefined distance from power lines. But, bear in mind that the northern white cedar grows at about 1cm every hundred years; willows can grow 1m to 3m per year. To remove false positives, a veg management algorithm should accurately ID tree species and only send out crews when necessary.
And it’s not just about how fast a tree grows. For obvious reasons, dead trees are a far greater risk than living ones. But it can be a slow transition from healthy tree, to diseased, to dying, to dead. And that can be difficult to estimate, even for an expert arborist, let alone an algorithm.
Start simple; start now. This will get a lot easier a lot sooner than you think.
Despite these hurdles, utilities can, in fact should, start automating their veg management processes. Even with annual surveys, machine vision algos can augment and validate expert opinion, not to mention ID human error. With carefully managed learning datasets, the algos will only improve. For companies with little input data, current returns may be marginal, but could have tremendous future value. Think about this more as a training exercise in preparation for a very different future. Because soon, the cost of collecting data will decrease substantially, and the science behind machine vision will become more sophisticated. I’ll publish a follow-up blog soon, discussing the reasons why.