Grid Alarms Marquee@2x

Grid Alarms

Kyle DickmanScience Writer

Share

An AI-driven method of detecting electrical faults could stop the most destructive fires before they begin.

June 1, 2024

Around 6:30 a.m. on November 8, 2018, another hammering gust of wind funneled up Northern California’s Feather River Canyon, a steep-walled gorge about eight miles northeast of the 26,000-person town of Paradise. Powerlines in the canyon swung wildly. A branch fell across a line, prompting a high-impedance fault (HIF), the same breed of hard-to-detect electrical fault that had ignited the fires that burned more than 5000 buildings in Napa and Sonoma counties the year before. Soon thereafter, sparks landed in dry grass, and the Camp Fire began—a catastrophic burn that would leave 19,000 buildings ash, most of them homes and businesses in Paradise, and 85 people dead. The consensus among experts is that if that HIF had been detected sooner, and if that defective section of grid had depowered automatically within the milliseconds or seconds it takes for faults to produce sparks, Paradise would still be standing. It just wasn’t possible. The capability to detect HIFs wasn’t yet robust or scalable. But now, because of pioneering work from Wenting Li and Deepjyoti Deka, a pair of Los Alamos coders and artificial intelligence experts, it could be soon. “We could prevent future Camp Fires,” says Li.

For the past three years, Li and Deka have been developing a new method that uses artificial intelligence (AI) to detect HIFs. Called PICAE, or the Physics-Informed Convolutional Auto-Encoder, the neural network–based detection system does what current systems cannot, and monitors grid health holistically. The biggest impediment to existing HIF detection lies in the limited number of sensors that make up the backbone of grid monitoring. Called phasor measurement units (PMUs), these high-speed sensors, which measure voltage and current throughout the power system with microsecond accuracy, are trained to recognize HIFs as a variation on a particular time or time-series pattern, a practice called labeling. Once identified, they can shut off power to the affected part of the grid before the sparks fly. 

PMUs are incredibly good at their jobs. There just aren’t enough of them. Each unit costs between 40 thousand and 180 thousand dollars, and many large grids, like Pacific Gas & Electric’s, whose 18,000 miles of powerlines brought electricity to Paradise, lack the sensors needed to cover the entire system. When HIFs occur far from a PMU-equipped substation, the sensors miss the faults about 10 percent of the time because the change in electrical flow can be so small that it’s absorbed by the dominant current. “If we had unlimited data and unlimited PMUs, detecting faults wouldn’t be a challenge,” says Deka. 

With PICAE, Li and Deka developed a new method that optimizes the data gathered by wide but sparsely distributed networks of PMUs. Rather than labeling faults, they train an AI model on the data the PMUs gather under normal operating conditions. Like ecosystem modeling, this physics-informed approach models not just the metrics measured by the sensors but how those forces interact. By doing this, the AI can extract more information out of the data-lean environment and create a holistic representation of grid operations. “We’re modeling with greater sensitivity how the grid functions under normal conditions,” says Deka. “Because of this, we’re able to detect smaller deviations from the norm.” What the AI model revealed is that when HIFs occur in one part of the grid, there are corresponding changes to voltage or electrical flow patterns throughout. By focusing on these subtle signals, not only is the model more robust against the noise native to all electrical grids, but it’s able to detect orders of magnitude more HIFs than existing systems.

“It means we don’t have to rely on the hardware alone to detect the faults,” says Li. She and Deka are now working to run PICAE on an actual grid, instead of the training data used to create the model. Given that less than 2 percent of fires account for 90 percent of firefighting costs, and that HIFs that arc on windy days cause the lion’s share of those extreme blazes, Li and Deka’s new detection method could soon save billions of dollars and untold numbers of homes and lives.

 

More 1663 Stories

1663 Home