A More Effective Way to Spot Potential Wildfire Origins in the Electric Grid

Sampling method capitalizes on similarities between grid sections to stretch inspection dollars

California’s dry, windy conditions and its more than 35,000 miles of electric grid pose an increasing wildfire risk. Though electrical equipment caused just 8% of the state’s wildfires between 2013 and 2023, those blazes accounted for 60% of structural losses during those years. The 2018 Camp Fire in Paradise, California, alone, sparked by a faulty transmission line, killed 85 people and destroyed over 18,000 structures.

So, regulators now require electric utilities to submit detailed annual plans justifying how their grid inspections reduce wildfire risk. Limited budgets and the grid’s vast scale mean only a tiny fraction of sections can be inspected each year. How to choose?

A working paper by UCLA Anderson’s Abolfazl Taghavi, a Ph.D. student, Johns Hopkins’ Ali Fattahi, USC’s Sriram Dasu and UCLA Anderson’s Reza Ahmadi offers utilities an optimization method that the researchers believe minimizes expected wildfire costs — measured in fatalities, acres burned ($1,100 per acre), structures destroyed ($1 million each) and power outages. Their work may help utilities meet regulatory demands and perhaps reduce wildfires.

The researchers worked with a single utility, whose identity remains confidential. The collaboration informed key modeling choices, such as how grid segments are clustered and how correlations among them are quantified, ensuring that the research reflects real operational data and regulatory context. Moreover, the utility’s current practice uses a method called Highest Mean Approximation, which ranks all segments across the entire grid by expected wildfire costs and inspects the highest-ranking ones. 

Following inspections, under HMA the utility does maintenance on segments above a certain threshold and skips servicing those below it. But when HMA finds problems in a segment, that valuable information is siloed, and risk estimates for similar nearby segments aren’t updated. It’s akin to doctors examining patients with identical symptoms but missing a flu outbreak.

HMA also doesn’t strategically allocate inspection resources. It just chooses the highest-ranked segments without considering if inspections of those segments will provide useful information for decision-making across the grid.

A Different Approach

The researchers recognize that power grid segments in the same area are like neighbors sharing weather and environmental conditions. Inspecting one segment reveals information about nearby segments. It’s like checking a select few homes’ roofs after a hailstorm to assess neighborhood damage. In the case of power grid segments, the researchers posit, “comprehensive inspection is unnecessary — a small, optimal subset suffices.” 

“The marginal value of additional inspection diminishes rapidly, indicating that limited inspection can achieve performance close to that of inspecting all grid segments,” they add.

The researchers’ experiments, using real data sampled from more than 100,000 segments of the utility’s electrical grid, suggest inspecting just 20% of segments achieves results within 0.22% of optimal. The utility’s current approach is 14.31% from optimal under the same constraints. (These percentages measure the gap from the theoretical best result: 0% equals optimal performance and 100% equals taking no action.)

Taghavi, Fattahi, Dasu and Ahmadi used a three-stage process to achieve these improvements: 

1. Select strategic segments for inspection.
2. Update risk estimates from inspection findings.
3. Do maintenance only segments exceeding the critical risk threshold.

Using data from the utility with segments organized into clusters and containing expected wildfire costs and estimated correlation between segments, the researchers calculated two key values: how much inspecting each segment reduces wildfire costs and the need for maintenance post-inspection.

They then performed a two-level optimization resembling planning both a road trip route and the specific stops. First, they allocate inspection quotas, limited by budget constraints, across clusters. Then within each cluster, they select segments maximizing cost reduction while respecting maintenance capacity constraints.

Selecting from tens of thousands of segments is a computationally difficult “NP-hard” problem — unsolvable in a reasonable time even by powerful computers. The researchers developed an approximation method, using a mathematical technique to reduce the complex problem into manageable subproblems.

After selecting inspection targets, the researchers’ process assigns wildfire cost risk estimates to segments, based on inspection findings, that are compared to a predetermined risk threshold. For example, if the utility sets the risk threshold as 50, a segment with an average risk of 80 clearly needs maintenance — no inspection required. One that has an average risk of 20 is fine with a high probability. But a segment whose average risk is 52 is right on the edge and inspecting it provides crucial information to determine whether it crosses the threshold.

Cluster correlations amplify that value. Finding high risk in one inspected segment can inform about others in that cluster — they’ve been exposed to the same environmental conditions, so they might also be riskier than initially expected. Based on correlations, the model updates estimates for uninspected segments, essentially gaining information about multiple segments from a single inspection.

Simulations using 10,000 grid segments from the California utility’s data suggested promising results, reducing extreme risk, or “Tier 3,” segments from 40% to 8.2%, versus 10.9% of segments left in Tier 3 by the utility’s current approach. And when the budget for how many segments can be fixed is increased to 40% from 20%, the researchers’ method eliminated almost all segments in the Tier 3 category. Excluding correlations left 21.1% in Tier 3 — illustrating the importance of incorporating correlation into their method. 

These results lend support to treating the grid as an interconnected network, rather than thousands of isolated parts. Through correlation learning and optimizing resource allocation, utilities may be able to achieve near-optimal performance with fewer inspections — potentially preventing California’s next wildfire.