Los Angeles wildfires were 10 times bigger than utility's AI forecast

Straits Times

15 hours ago

FILE PHOTO: A drone view shows workers cleaning debris from destroyed homes for residents of the Eaton Fire in Altadena, California, U.S., March 10, 2025. REUTERS/Mike Blake/File Photo
FILE PHOTO: An aerial view of the fire damage caused by the Eaton Fire is shown in Altadena, California, U.S. January 22, 2025. REUTERS/Mike Blake/File Photo
Southern California Edison's internal wildfire forecasts underestimated the potential size of the Eaton Canyon fire in Los Angeles by a factor of ten in the days leading up to a deadly conflagration in January, according to documents reviewed by Reuters.
The miss suggests potential weaknesses in the utility's fire modeling capabilities that factored into its response to the January wildfire threats, despite being upgraded with improved computing, datasets and artificial intelligence.
At the time, wildfires whipsawed through Los Angeles' western flank near Santa Monica and Eaton Canyon in the east as they consumed more than 34,000 acres (13,750 hectares) - or some 53 square miles - turning entire neighborhoods to ash.
Although no official cause for the Eaton Canyon blaze has been released, numerous lawsuits have claimed SCE's decision to keep power flowing to some lines and towers in the Altadena area led to the circumstances that triggered it.
SCE has said the cause and circumstances around the fire are under investigation and will be for some time, and defended its modeling capabilities.
"We are confident with our fire spread modeling and weather forecasting," Raymond Fugere, SCE's asset intelligence director, told Reuters in an interview.
Fugere said SCE's simulations could have shown higher estimates for acres burned in hard-hit areas. Variations in wind patterns and available fuels in hard-hit areas may not have been fully accounted for in the fire spread modeling, he said.
"But overall, we do feel confident with our modeling because it is giving us very actionable information to be able to make decisions," he said.
SCE's simulations predicted a Jan. 7 ignition in Eaton Canyon that could scorch about 1,000 acres within eight hours without fire suppression, according to an SCE fire potential forecast obtained by Reuters through a public records request. SCE told Reuters those fire spread simulations were factored into the utility's power shutoff decisions as strong seasonal winds and dry conditions escalated the looming wildfire threat.
The Eaton fire ignited as forecast on Jan. 7, but ultimately consumed some 14,000 acres, destroying around 9,400 homes and buildings, and killing 17 civilians – making it the centerpiece of one of the costliest natural disasters in U.S. history.
Joseph Mitchell, a scientist and wildfire expert witness for California utility regulators, said SCE's predictions missed the mark mainly because its models were only running simulations that extend eight hours after an ignition. The bulk of the Eaton fire's damage happened well after the eight-hour mark.
Michael Wara, a wildfire policy expert at Stanford Law School, said the wildfire modeling may also have erred because it is better tuned to simulating fire in dense shrubs and woodlands, instead of blocks of homes and businesses.
"Altadena was a wildland fire for about 20 minutes, and then it became an urban conflagration ... where houses are burning houses down, and where gardens are the fuel type not ... mixed conifer forests," Wara said.
SCE, a unit of Edison International, acknowledged it is evaluating changes to its wildfire risk models, including whether to use 24-hour fire spread simulations in the future.
"The January 2025 wildfires raise important questions regarding the spread of wildfires into built urban environments," the company said in a May 16 filing with regulators.
SCE has previously acknowledged that 24-hour simulations might capture more extreme events where firefighting resources are limited, according to regulatory filings with the California Public Utilities Commission prior to the fires. But SCE's Fugere said the longer simulations produce more uncertainty.
UPGRADED MODELS
SCE's forecast was the biggest test yet of upgraded forecasting capabilities since California Governor Gavin Newsom launched the "Wildfire Innovation Sprint" in 2019 – an initiative to encourage the use of AI to predict disasters and safeguard lives and property.
Since then, SCE has built four supercomputer clusters capable of generating 13 billion simulations across 400 weather scenarios and 29 million ignition points, according to regulatory filings.
The company also began using the services of Technosylva, a La Jolla, California-based company, which received $383,000 in state funding in 2019 to develop forecasting tools for utilities and emergency responders.
Technosylva Chief Executive Bryan Spear told Reuters his company's risk models captured the magnitude of the Los Angeles wildfire five days in advance, allowing firefighters to make better preparations for the looming catastrophe.
SCE's equipment has not been blamed for starting the massive Palisades fire, but the utility's forecast also vastly underestimated its potential size. The separate blaze started on the same day as the Eaton Canyon fire.
SCE's forecast projected up to 1,000 acres burned in the Pacific Palisades area, according to the document.
Actual wildfire destruction there included 23,448 acres burned, 12 civilian deaths and nearly 7,000 structures destroyed, according to Cal Fire.
Together the Eaton and Palisades fires destroyed more than 16,000 structures and caused most of the $250 billion in economic losses estimated by AccuWeather.
SCE plans to spend another $8 million on upgrading fire science and modeling this year, up from $2 million in 2018, company disclosures show. REUTERS
Join ST's Telegram channel and get the latest breaking news delivered to you.