Using Machine Learning to Improve Utility Customer Experience

Craig Lawrence Case Studies, Customer Experience, Data Science, ETR, Machine Learning

To most customers, their electric utility has only one job – to keep the lights on. A power outage is one of the few times, other than paying their bill, that a customer engages with their utility – so, the utility interaction with the customer is a high leverage point. Perhaps just as important as getting the lights back on is proactively communicating to the customer when they will get their power back. Not knowing is extremely frustrating to consumers. Utilities are investing in new methods of communicating outages to their customers – real-time outage maps, text alerts and notifications, and new outage reporting systems for customers. But telling a customer that their power is out is of limited value – they probably already know that. What they want to know is when it’s coming back on.

Most utilities have a fairly rudimentary method of calculating the “ETR” – the Estimated Time to Restoration – for an outage.  This might be as straightforward as a static lookup table based on institutional knowledge of historical restoration times, or as simple as a fixed estimate for all outages, regardless of cause.

Treverity believes that utilities can do better – that new modeling and machine learning technologies can be applied to historical data to produce more accurate estimates of time to restoration. TreverityEDGE collects data from multiple utility systems, including the Outage Management System (OMS), Asset Management System (AMS), SCADA, Mobile Workforce Management (WFM), Automated Distribution Management System (ADMS), Customer Information System (CIS), Advanced Metering Infrastructure (AMI), and more. We wanted to challenge ourselves to use all of that data to do better.

We have recently shared the initial results of a program with a large Texas-based electric distribution utility to do just that. The goal was to develop an automated method of calculating ETR using modern machine learning techniques, and to leverage the large data set that TreverityEDGE has collected over many years of operation. The results are shared here.

In summary, we improved the accuracy of ETR predictions by 34% relative to the utility’s existing method, and will soon implement the model so that new estimates are automatically produced every five minutes. This means field and operations crews do not need to spend time and effort making the updates themselves. We expect accuracy to improve even further over time as more data are collected, and as we improve the model itself.

This work is just beginning. We believe we can apply this type of machine learning to a wide range of utility operational challenges, including optimizing utility response to outages, and even predicting outage risks before they occur.  And, with TreverityEDGE, we can operationalize these models, making them immediately useful to the people who need it most.