How computers can catch power line wildfires before they ignite

A falling line

When high-voltage wires snap or touch dry plants, disasters happen fast. These incidents, called single-phase ground faults, can trigger devastating power line wildfires in fractions of a second.

Normally, safety systems wait for manually set temperature limits before shutting down. That delay often comes too late. Researchers built a machine learning pipeline to spot whether an electrical fault will ignite dry vegetation, and exactly when.

Reading the spark

Think of it like an ECG reading for power lines: instead of waiting for a patient to collapse, a doctor reads tiny tremors in a heartbeat to predict a faint. The system listens to the chaotic electrical waveforms, diagnosing disaster before flames erupt.

The system breaks messy electrical signals into tidy sub-signals using Variational Mode Decomposition, or VMD. A Random Forest model checks which features matter most. Next, ablation tests—stripping away data parts to test their worth—proved that simple time-domain features, which track how the signal changes over time, were entirely sufficient.

Using transfer learning, which borrows knowledge gained from previous tasks, the system predicted ignition risk with 0.9848 accuracy. It achieved a precision of 0.9818 and a sensitivity of 0.9878. If ignition was likely, a Long Short-Term Memory network, an algorithm tailored for time sequences, tracked temperature shifts to calculate ignition timing. It pinned down the exact moment of ignition with an average error of only 48 milliseconds—faster than the blink of an eye.

What we still do not know

Unlike human guesswork, this method automates warning systems. It could eventually trip circuit breakers before flames form.

However, the model relies on experimental laboratory data rather than field tests on live power grids. Real outdoor weather varies wildly, and an ECG analogy stops working here because power grids do not have a steady biological pulse. Engineers still need to test the software on real outdoor hardware to see if it handles unpredictable elements without delay.