Runway

Runway forecasts power outages for your ZIP like a weather app, learns what your home can’t lose, and builds a plan before the lights go out.

Comment

Inspiration

In 2023, one in four US households lost power at least once — 33.9 million homes — and 70% of those outages lasted six hours or more (U.S. Census Bureau, American Housing Survey). That same year, 705,000 households lost refrigerated medicine and 4.9 million lost food.

What struck us is that an outage isn't the same event for every home. A portable oxygen concentrator on one battery may last 2 to 6 hours. Food in a closed fridge stays safe for about 4. A full freezer, about 48. Every home has a different runway, and one alert for everyone can't plan for that.

Outages aren't only a power problem. They're an information problem. That became our project, and its name.

What it does

Runway is the weather app for power outages.

  • Forecasts outage risk for your ZIP code, hourly, 10-day, weekly and monthly, and explains why (wind gusts, heat, tree cover near lines, overhead wires).
  • Learns your household: medical devices, fridge medicine, an EV, a tight budget, family to check on.
  • Builds a personal plan, with each step timed before the riskiest day.
  • Watson, a chat assistant running on IBM watsonx.ai, lets you just say "I keep insulin in the fridge" and adds the right steps.
  • Live map: a risk heatmap plus current outages, radar, weather warnings, fires, power lines, and help nearby.

The idea in one line: for each home with needs $i$, each lasting $t_i$ hours without power,

$$ \text{runway} = \min_i \, t_i $$

and Runway acts when an outage likely to outlast that runway is coming:

$$ P(\text{outage}) \cdot P(\text{duration} > \text{runway}) \;\text{is high} \;\Rightarrow\; \text{prepare now.} $$

Dual use: on an ordinary day it saves money by moving chores and EV charging to cheaper hours. In a crisis it answers the Energy lens question, whose power can wait?, for home-battery fleet dispatchers, utilities, cities, and even military bases.

How we built it

  • Frontend: React + TypeScript + Vite, styled with IBM Carbon and built mobile-first. We also shipped a plain HTML/CSS/JS version that compiles to a single file and runs fully offline, for low-connectivity demos.
  • Maps: Leaflet with a canvas-rendered heatmap, plus layers from NWS alerts, Cal OES statewide outages, NOAA radar (via Iowa Environmental Mesonet), NIFC fire perimeters and OpenStreetMap (hospitals, cooling spots, EV chargers, power lines).
  • Location: a ZIP dataset covering 42,000+ US ZIP codes, built from open zipcode and ZCTA-county datasets, so everything works at ZIP level, not county level.
  • Forecast: National Weather Service forecasts feed a risk model scaled to each ZIP's exposure (tree cover, overhead lines, typical outage length).
  • AI: a FastAPI backend calls IBM watsonx.ai (Granite 4). Watson replies in plain language and returns structured plan items that drop straight into the household plan.
  • IBM Bob was our AI pair programmer for the whole build, from scaffolding the app to iterating on the UI and wiring the chatbot.

Challenges we ran into

  • Making it personal without being creepy. Personalization only works if people trust it, so answers stay on the device and every step explains why.
  • Live data in a sandbox. Our preview environment blocked external map tiles and APIs. We built an offline basemap from ZIP points and sample-weather fallbacks so the demo never breaks.
  • Getting the facts right. We checked every statistic against its source and cut claims that didn't hold up. For example, we dropped a "4 hours for insulin" line once we found that guideline is about food, not insulin.
  • Graceful AI failure. If watsonx is unreachable, Watson says so and falls back to safe, rule-based guidance instead of going silent.
  • Keeping it simple. Early versions had too many sections. We cut them down to one flow: forecast → plan → ask Watson.

What we learned

  • The hard part of resilience isn't the forecast. It's turning a forecast into the right next step for one specific home.
  • Dual-use design forces clarity: if a feature doesn't help on an ordinary Tuesday, nobody will have it open the night the grid fails.
  • Writing for empathy (plain words, one action at a time) matters as much as the model.

What's next

Validate the model on real outage records, add SMS and voice for people without smartphones, launch preset actions (order backup power or a medical kit, alert family), and pilot with an LA hospital, campus or utility. We'll pursue NSF SBIR and DoD ESTCP funding.

Built With

Share this project:

Updates

Submission history