Helping stop widespread blackouts before they start

Let’s say it’s winter outside with freezing temperatures and blizzard-like conditions. Then the power goes out — and it stays out for days throughout the majority of the city where you live. Graduate student Birva Sevak experienced this while living in Quincy, Massachusetts in 2022. As she watched the thermostat numbers continue to drop, she became increasingly concerned. 

“I thought it would come back within 15 minutes, but it didn’t. Then I thought, maybe an hour or two. Nothing. It was at least three days,” says Sevak, who earned her undergraduate degree at the University of Massachusetts-Boston. “When everything is electric — even my stove was electric at the apartment — it’s a problem.”

Now Sevak, who is a graduate student earning a Master of Science in data science, is exploring how to stop widespread blackouts before they start. Sevak also references a major blackout that affected many people in the northeastern United States and Canada in 2003. An estimated 50 million people, including Dearborn residents, lost electricity for up to four days and economic damages were in the billions of dollars.

“As we add solar farms, wind turbines and millions of electric vehicles, power grids are becoming more complex and harder to protect. The safety tools operators rely on were largely designed decades ago — running a single check on a large grid can take seconds while operators need hundreds of thousands of these checks every hour,” Sevak says. “I wanted to form a complete AI pipeline for grid resilience.”

Through her graduate research, Sevak has developed two artificial intelligence models to help electric grid operators understand current conditions and predict future failures. She says one model can help operators see the grid clearly in its current state, while another looks ahead and can warn operators before a single tripped electrical line becomes a regional blackout.

Both models use graph neural networks, a specialized type of AI that’s suited to power grids because it analyzes networks of connected components — much like social media algorithms analyze networks of people. Sevak, who works full time as a software developer at Itlize Global, uses GNN to analyze complex, interconnected data at her job. Since she is familiar with that framework, she wanted to explore what public-facing problems GNN may be able to help solve. “When I saw no work like this has been done before, I thought, let's do it and see what’s possible,” she says.

The first model — Topology-Aware Gated Graph Neural Network — figures out the voltage and current at every point in a large grid. That traditionally requires solving thousands of equations from scratch, repeatedly — but Sevak’s model learns the patterns behind those equations and delivers accurate answers in a single pass. The model also learns which connections matter most under different stress conditions. “It was tested on benchmark grids ranging from 30 to over 1,300 connection points,” she says. “It consistently outperformed existing approaches, with speed advantages that grow as the grid gets larger.”

The second model’s design — Physics-Informed Graph Neural Jump Ordinary Differential Equations — incorporates physics to reflect both the gradual stress buildup and the sudden breaking points in electrical grids. In milliseconds, the model predicts which lines will fail, which areas will lose power, how severe the outage will be and the order in which failures will happen. This gives operators a timeline for intervention rather than just a final snapshot.

Using the U-M’s Great Lakes High Performance Computing Cluster, Sevak ran the models on a test grid that could service a mid-sized city. Sometimes each test would take up to seven hours. Sevak says she had remote access to the Great Lakes HPC Cluster while conducting her research thanks to her thesis advisor, Assistant Professor of Electrical and Computer Engineering Van-Hai Bui.

The models correctly identified over 99% of the lines that would fail and over 97% of the areas that would lose power, while also explaining how much electricity would be lost during a blackout at more than 95% of the variation, which means it has high predictive accuracy. Sevak’s work is theoretical at this point, but her goal is to get the models implemented in an actual grid so it can be further tested.

Sevak and her thesis research, “Graph Neural Network Approaches for Real-Time Power Flow Estimation and Cascading Failure Prediction in Transmission Networks,” won the 2026 3 Minute Thesis Competition on campus. The 3MT Competition, sponsored by the Office of Graduate Studies and the Office of Research, cultivates academic, presentation and research communication skills, with students presenting their work in a succinct, easy-to-understand way. Sevak advances to the Midwestern Association of Graduate Schools’ 3MT Regional Competition and will present her research on March 27. Sevak appreciates how the 3MT Competition challenges researchers to explain technical topics without jargon.

“Whether you study information technology, machine learning, economics or medicine, your colleagues often share a similar level of expertise and you naturally use industry vocabulary with them,” she says. “Sometimes it is easy to forget that people outside of your circle aren’t specialized in your subject matter — it is your responsibility to explain it in non-technical terms. Otherwise your stakeholders may not understand the importance of what you are telling them.”

From presentations to research work to publications, Sevak says UM-Dearborn has given her many professional growth opportunities. She chose the university for graduate school because it has a highly rated academic program that combines data science and machine learning. 

“I have gained so much out of my time here. I’m doing research I never expected to do, published papers and obtained grant approvals, all while learning how to apply AI to fields outside of my own,” she says. “UM-Dearborn is a great place to learn and my professors have supported me every step of the way.”

In addition to professors and the research teams at UM-Dearborn, Sevak, who will graduate in April, also wants to thank her family and friends. “My friends in a similar program and I work together on a lot of our research work, helping and learning from each other's research," she says. “My family has supported and encouraged me during this journey too.”

Story by Sarah Tuxbury