⚡ Storm Outage Risk — Missouri & Illinois

Which counties lose power when a storm hits — and why does the same storm hit some places much harder than others?

A county-level power-outage model for the 216 counties of Missouri and Illinois, built on public data: DOE EAGLE-I outage records, NOAA HRRR storm ingredients and storm reports, USFS tree canopy and Census housing data. The pipeline runs locally on DuckDB and dbt; models are compared in a benchmark suite tracked with MLflow; results are served in a Streamlit dashboard packaged with Docker.

What it does

1. Predicts for every county and every 6-hour block whether ≥ 1%, 5% or 10% of customers lose power.
2. Scores new outages separately from ongoing ones. Outages persist, so a model can look accurate by predicting "still out"; the hard and useful part is the onset.
3. Explains risk with storm ingredients (wind, gusts, shear, instability, rain, ice) and place: tree canopy, housing age, density, and each county's outage history.
4. Scores resilience: outage hours per customer, events per year, how often a reported storm turns into an outage, repeat outages.

Results

PR-AUC on new outages (onset blocks), XGBoost (800 trees, depth 4) trained on 2020–2024, with settings chosen on 2024 and checked over five seeds, tested once on 2025. For rare events a model with no skill scores the event rate, shown for reference.

Outage size	New-outage rate	Climatology baseline	Weather only	+ place	+ recent outages
Next year (train 2020–2024 → test 2025)
≥ 1%	1.97%	0.032	0.116	0.145	0.153
≥ 5%	0.53%	0.009	0.068	0.091	0.106
≥ 10%	0.26%	0.005	0.047	0.062	0.083
Unseen counties (5-fold by county, 2020–2025)
≥ 1%	1.73%	0.016	0.109	0.110	0.122
≥ 5%	0.46%	0.004	0.070	0.071	0.086
≥ 10%	0.22%	0.002	0.054	0.054	0.068

Source: ml/results/leaderboard_final_2025.md; settings from tune.yaml and seeds_tune.yaml, feature choices from seeds_ablate.yaml.

What the numbers say:

• Weather alone predicts new outages 6–24× better than chance, more so for bigger outages.
• More years of storms helped more than any feature: training on 2020–2024 instead of 2024 alone lifted every score by 17–28% on the same 2025 test (years_2020.yaml, years_2024.yaml).
• Place helps for known counties, not new ones: +25–34% on the next year, about zero on unseen counties.
• Of the 1% of calm county-blocks the model ranks riskiest, about 1 in 7 become a ≥ 5% outage (26× the base rate), catching 26% of all new ≥ 5% outages. With the alert level fixed in advance instead (the top 1% of a 2024 validation run), it flags 1.1% of calm blocks in 2025, 13% become a ≥ 5% outage (25×), and they catch 27%.
• The test year is complete: 1 January to 31 December 2025, 290,002 county-blocks, 271,927 of them with no outage under way, 1,445 new ≥ 5% outages.
• On new outages the model is a weather model: surface wind, minimum temperature, county size and wind × rain lead. Scored on all blocks, the previous block's outage is seven times any weather or place feature — persistence, not prediction.
• Tuning barely matters: the top ten settings on 2024 sit within ±1.2%, about the seed-to-seed spread. Up-weighting rare events hurt.
• Tree canopy: small but real, once tested properly. With LightGBM on one year it looked strong (−12 to −22% without it); with XGBoost it shrank or reversed. Over five seeds and five years, removing it lowers the score with every seed in three of four settings (−2 to −3%).
• Left out: rain in the days before a storm (mixed across seeds) and as-of outage history (no gain on the validation year).

How it compares

PR-AUC on new ≥ 5% outages, next year (ml/results/leaderboard_official.md, ml/results/leaderboard_weathernext.md). Official forecasts are taken as they stood when the block started; the model's HRRR weather covers the block itself.

Approach	Known	PR-AUC
NWS warnings in force at block start (event rate per combination)	in advance	0.024
SPC Day 1 outlook (event rate per risk level)	in advance	0.033
Warnings + outlook	in advance	0.044
weathernext-outage-forecasting's setup, reproduced on this data	during the block	0.053
This model	during the block	0.106
This model with warnings + outlook as features	during the block	0.122

• The model scores 2.4× the official forecasts and gains another 15% (16% on unseen counties) when given them as inputs.
• Warnings are counted only if in force at the block's start. Counting any warning that touched the block nearly doubled their score, because most severe thunderstorm and tornado warnings are issued as the storm, and often the outage, arrives.
• Not a fair race yet: the model sees the block's weather, the forecasts didn't. Matching lead times needs weather forecasts issued before the block (roadmap).
• The weathernext row reproduces that project's method (weather-only features, class-weighted boosted tree, 80/20 county-date hash split) with HRRR in place of WeatherNext. On its own split it scores 0.059, and its precision at a 0.5 cut-off is 2%.

What it would be worth

If a utility had staged repair crews wherever each forecast pointed in 2025, how much would it have saved? pixi run value (ml/value.py) scores every forecast on the same new outages, two ways:

• Relative economic value (the cost-loss model used to value weather forecasts), which needs no dollar figures: the share of a perfect forecast's savings captured. The model reaches 0.52; warnings + outlook 0.34, the SPC outlook 0.31, NWS warnings 0.09.
• Dollars: customer outage costs from Berkeley Lab's ICE models and a typical customer mix, under four scenarios (a staged crew shortens outages by 10% or 25%; staging costs $5,000 or $20,000 per county per 6 hours). Every forecast stages by the same rule, using 2020–2024 data only.

Net savings in 2025	Range over the four scenarios
Perfect forecast	$11.6M to $99.5M
This model	$3.5M to $38.8M
Warnings + outlook	−$14.9M to −$0.6M
SPC outlook	−$18.9M to −$1.4M
NWS warnings	−$57.7M to $2.0M

The estimate: about $4–39 million saved in 2025, where acting on the official forecasts would mostly have lost money: they point at far more counties than lose power. At each forecast's best cutoff in hindsight, the model still saves about 3× as much. Almost all the avoided cost is businesses' (for households alone, staging never pays), the figures are customers' losses rather than the utility's, and like every score here they are a best case because the weather inputs are near-observed. Results: ml/results/value_2025.csv, ml/results/value_curves_2025.csv.

Limitations

• Weather is short-range, not a 24–48 h forecast. The HRRR archive stitches the first hours of each model run, which is close to observed weather. Scores are an upper bound for planning a day ahead.
• One test year. Settings were chosen on 2024 and 2025 was scored once, but a single test year gives no confidence interval; the seed spread (about ±1%) measures model noise only.
• One weather point per county. Storm winds are local; a county-wide grid would be better.
• Canopy is county-wide and correlated with rurality and the serving utility (its link to outages comes almost entirely from Missouri's Ozark counties), so its effect is an association, not a cause.
• EAGLE-I coverage is patchier before 2020, which limits how far back training can go.
• St. Louis City and County are one region: EAGLE-I records the County's outages under the City's code from mid-2020 to 2024.

How it's built

ingest/ (Python)          dbt/ (SQL on DuckDB)                    ml/ (Python)                 app/
EAGLE-I, HRRR,     ──►  staging → intermediate → marts   ──►  benchmark suite (MLflow)   ──►  Streamlit
storm reports,          47 tests; 3.5M-row modeling table      YAML configs, 3 splits,        (Docker)
canopy, Census          + resilience scorecard                  onset + all-block metrics

Layer	Tool
Environment	pixi — one lock file; default, app and ci environments
Storage / SQL	DuckDB
Transformations	dbt Core + dbt-duckdb, dbt_utils tests
Geo	geopandas, rasterio
ML	XGBoost, LightGBM, scikit-learn, MLflow
Dashboard	Streamlit, Plotly
Packaging	Docker (linux/amd64), Hugging Face Spaces
Quality	pytest, ruff, GitHub Actions

Design choices worth knowing:

• Region × 6-hour block is the unit; the label is the block's peak % of customers out.
• Missing rows are zeros only when EAGLE-I was running. A national "heartbeat" (any county in the U.S. reporting) separates "no outages" from "no data".
• Leakage guards: history features come from 2020–2023 only; recent-outage features use previous blocks; storm reports are context, never inputs.
• Validation: next year (time) and held-out counties (spatial), always against baselines. A random region-day split is in the code to show how it overstates skill.
• The dashboard reads precomputed Parquet only — it never runs a model.

The data traps found along the way (a renamed column in 2023, duplicated snapshots in 2025, a raster whose no-data value was a real value, county codes swapped for years, and more) are in docs/data-notes.md.

Quick start

pixi install
# EAGLE-I outages: manual download — see data/README.md
pixi run fetch-counties && pixi run fetch-tree-canopy && pixi run fetch-acs && pixi run fetch-storm-events
pixi run fetch-weather 2024 2025      # rate-limited (free tier): ~a day for 216 counties
pixi run build                        # dbt: every model + test
pixi run train                        # quick benchmark → ml/results/
pixi run export-app && pixi run app   # dashboard at http://localhost:8501

docker compose up --build runs the dashboard in a container.

Repository layout

Folder	Contents
ingest/	download scripts, one per source
dbt/	SQL models (staging → intermediate → marts), seeds, tests
ml/	data loading, splits, model registry, metrics, benchmark runner, configs, results
app/	Streamlit dashboard
notebooks/	explore_warehouse (inspect tables), benchmark_lab (try configs interactively)
tests/	pytest: benchmark machinery and dashboard pages
scripts/	download queue, Hugging Face publishing
legacy/	the first St. Louis hourly prototype
docs/	data dictionary, data notes, decisions, experiments, audit, development guide, roadmap

Data sources

• Power outages: Tansakul, V. et al. (2023). EAGLE-I Power Outage Data. Oak Ridge National Laboratory. https://doi.org/10.13139/ORNLNCCS/1975202
• Storm ingredients: NOAA High-Resolution Rapid Refresh (HRRR), via the Open-Meteo Historical Forecast API.
• Storm reports: NOAA NCEI Storm Events Database; NWS zone–county correlation file.
• Official forecasts: NWS warnings and SPC Day 1 convective outlooks, archived by the Iowa Environmental Mesonet.
• Tree canopy: USDA Forest Service / NLCD Tree Canopy Cover 2021, via MRLC.
• Housing and population: U.S. Census Bureau, ACS 2020–2024 5-year estimates.
• County boundaries: U.S. Census Bureau cartographic boundary files (2023).
• Airport weather (prototype): NOAA NCEI Global Hourly (ISD), AWS Open Data.

Full source list with URLs: docs/data-dictionary.md.