Pismo

Pismo tests if forecasts can tell home batteries to charge before NWS warnings. In Feb 2024, our Humboldt trigger fired 27 hours before the High Wind Warning.

Comment

Inspiration

Home batteries with storm features, like Tesla's Storm Watch and Enphase's Storm Guard, charge when the National Weather Service (NWS) issues selected alerts. A mentor told us people complain these can come too late. The risk is highest in rural areas, where outages last longer, crews have to travel farther, and long single-path lines leave few backup options. We asked whether a weather forecast could tell a battery to fill up before the warning arrives.

What it does

Pismo replays the February 2024 storm in northern California. For Humboldt, Sonoma, Santa Cruz and San Mateo counties, it shows when a forecast-based trigger would have fired compared with when the NWS issued watches, advisories and warnings. In Humboldt the trigger fired 27 hours before the High Wind Warning.

  • A simulator lets you set the battery size, charge rate, starting charge and outage start, and shows how full the battery would be under each strategy.
  • A rural-county comparison shows why we focused on Humboldt.
  • An evidence section states what the data does and doesn't support.

How we built it

  • Weather: NOAA GFS forecasts through Google Earth Engine, averaged by county, with a 24-hour window checked once a day.
  • Alerts: the NWS alert archive from the Iowa Environmental Mesonet for the San Francisco, Eureka and Sacramento offices. We mapped forecast zones to counties by area overlap.
  • Outages: EAGLE-I county outage records for 2018 to 2021, used to compare counties.
  • Trigger rule: forecast rain of 15 mm or more, or peak wind of 18 m/s or more. We set it from December 2023 and January 2024 data and did not adjust it for the February storm.
  • Prototype: a single interactive web page with the real data embedded.

Challenges we ran into

  • Earth Engine was slow, about 8 months of forecasts in 1.5 hours, so we cut the scope to the storm replay and two calibration months.
  • The outage data has gaps. Only about 1 day in 7 is fully recorded, so quiet days are hard to confirm and outage counts are more reliable than non-outage counts.
  • Forecast zones don't line up with county lines, so counts changed with our matching rule.
  • Large data downloads failed or ran out of disk space, so we filtered the files down to a few counties.
  • County-average weather predicted outages only weakly on Bay Area training months.

Accomplishments that we're proud of

  • A trigger set before the test storm that beat the NWS warning by 27 hours in Humboldt.
  • A clear finding that Santa Cruz and San Mateo received no warning at all, so a warning-only battery would never have charged there.
  • A working replay and simulator built in a day from public data.
  • A page that states its limits openly.

What we learned

  • Warnings are rare. Humboldt had one warning-level wind or flood alert in 62 winter days.
  • Forecast rain tracks NWS flood alerts well, but wind tracks wind alerts and outages much less.
  • A battery fills in a few hours, so earlier notice only helps when an alert would otherwise land too close to the outage. The cost of forecasting is charging on more days.
  • Outage labels are the hardest part, since a missing record could mean no outage or no data.

What's next for Pismo

  • Add the actual February 2024 outage start times for Humboldt and replace the scenario slider with real onsets.
  • Test the trigger on more storms and winters, and add Shasta as a second rural county.
  • Train a real outage model with confirmed outage labels, using finer-scale weather than county averages.
  • Verify how soon forecasts are available, and compare against battery activation data from real devices.
  • Add community vulnerability layers, such as the HHS emPOWER medical-equipment counts, so outreach goes where outages hurt most.

Built With

Share this project:

Updates

Submission history