Study finds LLMs close the gap on power outage prediction without training data

Overview

Updated 4 days ago

A preprint posted September 2 asked whether an off-the-shelf language model, trained on no utility data at all, can predict which storms cause the worst power outages. Four zero-shot LLMs were benchmarked against two supervised machine learning models on six years of central Texas outage records.

The supervised models won on precision and F1. But the newest LLM generations came close, and they added two things supervised models lack: readable explanations of each risk call, and the ability to transfer to a new region without labeled outage data. The authors conclude that combining LLMs with supervised models is the best practice.

The practical payoff is crew dispatch. Utilities position repair crews before storms, and better severity prediction means putting them where damage will actually hit. The study is one of at least five papers in six months testing LLMs for grid outage risk, a sign that the approach is moving from curiosity to serious evaluation.

Why it matters

If hybrid LLM-ML outage prediction reaches operations, utilities get storm-severity forecasts without years of labeled data — and can dispatch crews before the damage hits.

Questions about this story

Free account needed to ask — your question is kept and asked for you right after sign-up. Answers are public.

No questions yet — be the first to ask.

Key Indicators

Zero-shot LLMs benchmarked

Four off-the-shelf LLMs tested without fine-tuning against two supervised classifiers.

Supervised classifiers used as baseline

Traditional machine learning models trained on the same Texas data held the accuracy edge.

6 years

Historical outage records used

Six years of outage and weather records from a central Texas utility service area.

Forecast horizons tested

Severity classification at 3, 6, and 12 hours ahead.

Voices

Curated perspectives — historical figures and your fellow readers.

Ever wondered what historical figures would say about today's headlines?

People Involved

Christos Petridis

Lead author of the arXiv preprint

Active researcher at Temple University

Zoran Obradovic

Co-author, professor at Temple University

Co-supervised the study

Mladen Kezunovic

Co-author, professor at Texas A&M University

Co-authored the study

Organizations Involved

arXiv

Preprint server

Hosted the outage study

Open-access preprint server where the study was posted September 2, 2026.

PulseAugur

News and research aggregation site

Covered the study's release

Reported on the preprint September 7, highlighting LLM promise for power outage prediction.

Timeline

March 2026 September 2026

4 events Latest: 5 days ago

Tap a bar to jump to that date

September 2026
Press coverage spotlights LLM outage prediction
Latest Media Coverage

PulseAugur reports the study, noting newer LLM generations show competitive performance with added reasoning and scalability strengths.
5 days ago
Zero-shot LLM outage study posted to arXiv
Research Publication

Petridis, Obradovic, and Kezunovic benchmark four LLMs against two supervised classifiers on six years of Texas outage data. Supervised models win on precision and F1; newer LLMs come close.
September 2nd, 2026
OutageDiT foundation model posted to arXiv
Research Publication

Generates seven-day outage trajectories at quarter-hour resolution from US-wide weather and outage records.
September 1st, 2026
March 2026
OutageGPT paper proposes multi-agent LLM for outage queries
Research Publication

Proposes retrieval-augmented multi-agent LLM that beats open models on 2021 severe-weather outage queries.
March 1st, 2026

Scenarios

Utility pilots hybrid LLM-ML outage adviser

Possible Resolves by End of 2027

Discussed by: Study authors Petridis, Obradovic, and Kezunovic, whose preprint recommends combining LLMs with supervised models

A US utility or grid operator pairs an LLM with a supervised classifier for storm-outage severity prediction and publishes results of a field pilot. The LLM handles explanation and transfer to new regions; the classifier carries precision. Peer-reviewed documentation or trade reporting confirms the pilot moved beyond research.

Data-poor region deploys LLM-only outage prediction

Possible Resolves by Q2 2028

Discussed by: The preprint's finding that LLMs scale geographically without labeled training data; the EACL 2026 findings paper

A utility outside North America and Europe, lacking labeled outage records, deploys an LLM-based storm-risk warning system. Zero-shot transfer lets it operate without training data. A peer-reviewed paper or major wire report documents the deployment.

Deployment stalls on validation gaps, LLMs stay in research

Uncertain Resolves by End of 2027

Discussed by: The EACL 2026 findings paper, which found LLM accuracy still trails standard machine learning

Utilities treat LLM outage prediction as promising research but decline to fund operational pilots. They cite accuracy gaps, safety-validation requirements, and unclear liability. Trade coverage confirms no US pilot has moved forward.

Historical Context

2 moments from history that rhyme with this story — and how they unfolded.

2013–2018

IBM Watson for Oncology (2013–2018)

IBM marketed its Watson AI as recommending cancer treatments based on trained datasets, with Memorial Sloan Kettering as a launch partner. By 2017, investigations found Watson's recommendations were often based on hypothetical or unvalidated cases, and some hospitals reported unsafe advice.

Then

IBM wound down Watson Health, selling parts in 2022; medical AI pivoted to narrower, rigorously validated tools.

Now

A defining cautionary tale that high-stakes AI deployment requires clinical-grade validation, not just impressive demos.

Why this matters now

LLMs for grid outage prediction face the same validation threshold: accuracy on historical data plus proven operational safety before a utility bets crews and dollars on them.

July–November 2023

GraphCast and Pangu-Weather (2023)

Two machine-learning weather models matched or beat the physics-based European Centre for Medium-Range Weather Forecasts system for 10-day forecasts at a fraction of the compute cost. Google DeepMind's GraphCast and Huawei's Pangu-Weather showed learned models could challenge numerical weather prediction.

Then

Both were tested in operational forecasting pipelines within months; agencies ran AI models alongside physics models to compare outputs.

Now

ML weather forecasting became a standard benchmark, and operational agencies adopted hybrid physics-ML systems rather than replacing physics models outright.

Why this matters now

The outage study tracks the same adoption pattern: a new ML paradigm proves competitive against incumbent methods, then earns a role in hybrid deployment.