- In Jamaica, about a quarter of electricity produced is stolen or “lost” through non-paying customers and/or accounting errors. Manual detection has failed to make a difference in reducing this theft.
- ESMAP’s technical assistance team implemented a machine learning model to help Jamaican utility JPS identify and decrease incidents of theft.
- The machine learning model is based on an open source code, and is available for free to any utility.
Globally, billions of dollars are lost every year due to electricity theft , wherein electricity is distributed to customers but is never paid for. In 2014 alone, Jamaica’s total power transmission and distribution system reported 27% of losses (due to technical and non-technical reasons), close to double the regional average. While the utility company absorbs a portion of the cost, it also passes some of that cost onto consumers. Both actors therefore have an incentive to want to change this.
To combat this, JPS would spend more than $10 million (USD) on anti-theft measures every year, only to see theft numbers temporarily dip before climbing back up again. The problem was, these measures relied primarily on human-intensive, manual detection, and customers stealing electricity used more and more sophisticated ways to go around regularly metered use. JPS employees would use their institutional knowledge of serial offenders and would spend hours poring over metering data to uncover irregular patterns in electricity usage to identify shady accounts. But it wasn’t enough to effectively quash incidents of theft.
Now, Jamaica is one of the first countries to use machine learning to tackle its electricity theft problem . The World Bank partnered with Chicago-based data science firm, The Impact Lab, and the Energy Sector Management Assistance Program ( ESMAP) to use machine learning to improve JPS’ theft identification process among large and commercial accounts.
A machine learning model needs much less time to scan spreadsheets than a human; even though its algorithm does the same thing that humans do. It uses experience, history, and past results and gradually trains itself to determine suspicious activity. The Impact Lab’s data scientists used employees’ experience with the types of variables to flag, and integrated the model with company data.
Tom Plagge, Co-founder and Chief Scientist of The Impact Lab shared, “Machine learning models oftentimes feel like magic, so we showed staff how these results build on the company’s staff intuition, fit with what they already know through their experiences. It’s like doing what you’re already doing but in an automated, more precise and faster way.”
The project helped JPS to combine machine learning and human intelligence to produce a digital prototype model. Manual theft detection was replaced by time series visualizations, heat maps of usage, and detailed phase information for each account.
According to a recent World Bank brief, Energy Analytics for Development, the data found thousands of large individual and commercial account holders constituted the largest portion of electricity theft. Armed with this information, the strike rate of successful JPS investigations increased substantially, during the first months of its implementation the strike rate had doubled. Motivated by this success, JPS increased their investment in smart grid advanced metering infrastructure (AMI) and combined that with the machine learning model.
“Automating the analysis of daily transactions on electricity accounts will certainly improve the strike rate for detecting and addressing theft,” confirmed Steve Dixon, Director of Transmission & Distribution Asset Management at JPS.
The developed and tested tool is now housed on GitHub’s development platform. Its code is freely useable and ready to be plugged into the accounts’ data of any utility in a situation when a company has a similar problem. For more information, contact Anna Lerner.
Join the Conversation