Zoran Obradovic part of an NSF-funded research team using weather, social media and other data to predict when and where outages will occur

Lightning, wind, extreme cold, fallen trees or heavy snow can leave families, homes and businesses without power, whether it’s for a few hours or a few days.

While utility companies often have the logistical know how, machinery and staff to respond after an outage and turn the lights back on, what if they had the ability to better predict when outages will occur? They could take preemptive steps before power lines go down, deploying people and assets ahead of time to significantly reduce the impact of a power outage.

Zoran Obradovic, Laura H. Carnell Professor of Data Analytics, and his research partners are collecting historical outage data, weather-related data and other types of ‘big data’ and then using advanced machine learning to make predictions on when and where power outages will occur.

Called ALERT: Advanced Learning for Energy Risk Traking, the National Science Foundation-funded initiative looks to develop an early warning protype and outage prediction model. Utility companies will be much better prepared to minimize outages and even prevent catastrophic events like total blackouts. Consumers would have earlier access to important—and potentially lifesaving—information.

“We are building a system that is learning from multiple sources, and this type of multimodal learning is where its power lies,” said Obradovic, who is also director of CST’s Center for Data Analytics and Biomedical Informatics. “In the ongoing NSF project and our related project supported by the US Department of Energy (DOE), we collect large data sets related to weather, weather forecasting, power generation, consumer use of energy, substation failure and information from PMUs.” CST graduate students working on the NSF and DOE projects include Hussain Otudi and Daniel Saranovic.

PMUs or synchophrasors or phasor measurement units are deployed on the electric transmission system to measure voltage and current at a milliseconds rate, more than 100 times faster than older data acquisition systems. And while the data collected by PMUs is fertile ground for machine learning analysis, they too can get knocked offline by weather events.

To fill any potential gaps in data gathering, in complementary projects funded by the US Army Research Laboratory and US Army Engineer Research and Development Center, the Obradovic research team, including CST graduate student Rafaa Aljurbua, turned to social media, a rich source of detail about not only weather events but the effects power outages have on people’s lives.

“We look at publicly available posts to learn more about how outages impact people’s everyday lives, from milk turned bad in the refrigerator to batteries running low on a medical device,” explained Obradovic. “This type of information can be very useful as a complement to other types of data we collect, but we need very sophisticated methodologies for natural language processing to extract useful information from very noisy social media text.”

Research collaborator on the ALERT project include Mladen Kezunovic, from the Department of Electrical and Computer Engineering at Texas A&M University, Alexander Brown from the Department of Economics at Texas A&M, Roger Enriquez from The University of Texas at San Antonio and Paul Pavlou from the University of Houston.

“There are so many layers of information, and so many ways that a down electrical grid overlaps with, for instance, people’s access to water, and other systems,” said Obradovic, who cited a cold snap in Texas a few years ago where earlier and better predictions might have prevented both loss of life and damage to infrastructure. “Our system is called ALERT because it’s about giving utility companies and people the information needed to be proactive, before the power goes out.”