AI that stops Electricity Theft

Brazil has a big electricity theft problem. But an AI algorithm tested on several million of the country’s households shows promise as a tool for helping cut this out. It could also offer insights for electricity suppliers elsewhere seeking to do a better job of reading your meter.

Who is responsible for the theft in Brazil isn’t always clear – sending meter readers to check whether meters and overhead cabling have been tampered with is dangerous work, says Adrian Grilli of the Joint Radio Company in London, which does telecommunications for global energy companies.

Electricity theft is hardly limited to Brazil: some countries see as much as 40 per cent of their supply siphoned off largely by users who have tampered with meters. And in the UK, for example, such activity is estimated to cause losses of £440m every year. “Mostly it’s fraud,” says Patrick Glauner at the University of Luxembourg. But big data may soon blow the thieves’ cover.

Glauner and a team of researchers looked at the energy use of 3.6 million Brazilian households over five years. Specifically, they focused on nearly 200 million monthly readings made by professional meter readers.

Then they developed an algorithm that, after sifting through these, could recognise when energy use at a property was suspiciously low. Because the researchers also had data on past inspections against which to check the AI’s conclusions, they could verify that their tool was correctly highlighting potential cases of fraud or faulty metering.

This could help better target physical inspections of properties, which cost time and money. They can also be dangerous. Grilli has seen vulnerable electricity equipment in Brazil first hand, though parts of the favelas were off limits. “I haven’t been into those areas because we were told you’re not allowed in there. They didn’t want to send meter readers in because of the risk to their health and welfare,” he says.

Pinpointing cases
The most accurate versions of the system were able to identify problem cases just over 65 per cent of the time, which the team believes outperforms similar tools. Smart energy firm CHOICE Technologies, which was involved in the research, now plans to implement the technology in commercial software that will be used in Latin America.

Although electricity theft is less common in developed countries, energy firms in the UK are concerned enough that they recently agreed to share data to help identify potential cases.

“It’s interesting that this software can, in theory at least, tell that that building using 10 per cent of what others are using is stealing electricity,” says Paul Ruyssevelt at University College London’s Energy Institute. However, he also points out that it’s normal for adjacent buildings to have very different levels of energy use, which could lead to false positives.

Similar false positives may also be leading to erroneous billing by energy companies. “There are other types of anomaly – meters break, or there could be billing errors,” says Glauner. In the end, adopting an algorithmic approach to understanding meter errors could lead to a better deal for both suppliers and consumers.