Sandia researchers build AI system to detect anomalies across the electrical grid

The electrical grid is a crucial, sometimes fragile, piece of infrastructure. As connectivity to the grid increases, so too does its vulnerability.

Public Service Company of New Mexico, the state’s largest electricity provider serving around 550,000 customers, in the past has said the only way the utility knows about an outage is when someone calls to report it, and only then can it begin to assess the issue.

However, researchers at Sandia National Laboratories have developed a solution for monitoring the grid through artificial intelligence, a technology that has seen rapid growth in recent years.

“This technology is able to detect abnormalities in the grid and help us understand when we’re having a cyberattack or disturbance so we can quickly respond and mitigate it,” said Shamina Hossain-McKenzie, a Sandia cybersecurity researcher and lead on the project.

The three-year project cost about $2 million, Hossain-McKenzie said. Funding came internally from Sandia’s Laboratory Directed Research and Development program, a high-risk, high-reward research initiative.

“This project has about six to nine team members (who) are pretty interdisciplinary. We have computer scientists as well as cybersecurity researchers and electrical engineers to focus on power system analysis,” Hossain-McKenzie said. “We had a great assortment of researchers on this who were crucial in making this a success.”

Hossain-McKenzie said the team collected fiber data, like communication over control traffic, as well as physical data, the underlying physics of the grid, like frequency, voltage and current. Then, she said, they looked at the data sets together using an autoencoder neural network that allows researchers to assess and determine whether there’s an anomaly or not.

Anomalies are classified as cyber events, where grid security is threatened intentionally; physical events, like inclement weather or animal disturbance; or a combined cyber-physical event where, for example, an attacker inserts a device into the network and attempts to take control of a generator, Hossain-McKenzie said.

Unlike other AI models, an autoencoder neural network does not need to be trained on data marking every problem that could happen. Rather, it requires generous amounts of normal operations data.

Hossain-McKenzie said many industries are looking at ways to leverage AI, but deploying a model in the critical infrastructure space is a fairly new use.

“I believe that utilities and others are starting to look into it, but there’s more work in understanding the robustness and performance of your AI algorithms. So, I think that’s where we are right now,” she said.

After constructing the autoencoder neural network, she said the team moved toward testing on smaller hardware, like single-board computers, before moving to lab and fieldwork at PNM’s Prosperity Energy Storage testing site under a Cooperative Research and Development Agreement, commonly referred to as a CRADA.

At the PNM site, Hossain-McKenzie said the team specifically focused on integrating the technology into a live system and collecting cyber-physical data to test if the autoencoder neural network could detect the different anomalies, which they were able to successfully validate.

Hossain-McKenzie described the technology as a “passive monitoring device,” meaning that its only function is to collect data and perform analyses. The technology is not deployed on PNM’s grid.

Don Maez, a PNM project manager in the Advanced Energy Technology department, said the testing facility is a small, 500-kilowatt photovoltaic battery storage site that went live in 2011.

Maez estimates that commercialization of the technology is still years out, but as far as the grid’s concerned, it could help PNM and other utilities see detections faster and therefore have a quicker response time.

“In the future, as we modernize the grid, we’ll continuously look for those types of innovative technologies that can help us detect these types of events as well as other operating characteristics in an effort to operate the grid safely and reliably for our customers,” Maez said. “Any kind of advanced technology, any kind of research that will help us with that, we’re always interested in looking at that.”

Hossain-McKenzie said the next step for this project is finding more utilities to expose the technology to different data characteristics and system configurations, something the team is actively working on.

“We’re still working on improving the robustness, and additionally, we have a nonprovisional patent on the idea,” she said. “So, we are also looking for commercial partners who would actually take that technology and take it into the commercial space.”