Predicting school closures with historical weather prediction data?

Question

I'm interested in predicting when the local schools will be closed due to snow and other winter weather. I have a list of school closure dates and a rough sense of the criteria used to determine the closing of schools including, temperature, wind, precipitation, and their interactions. Using a statistical technique (classical, Bayesian, or machine learning), I'd like to use both the actual weather the day before (t-1) and the predicted weather for the the day (t) to predict the county's decision.

I want to use the predicted weather at t and not the actual because that's the information known to the deciders when they have to make their decisions. For the actual, realized weather for t-1, I was planning on using the NOAA API. But I'm at a loss on where to find weather prediction data. While I could do what @MarkRovetta suggests in the question on Statistical weather prediction and roll my own prediction model, that is harder than using a database and my lousy prediction model won't compare with the National Weather Service predictions available to the country officials. At best that will act like noise that will reduce the quality of my predictions. My primary question is therefore what sources are available for local weather predictions at various lags to the realized weather?

Additionally, are there tables of the predictive quality of forecasts at various time horizons? If, for example, the daily weather was predicted with 99% accuracy by 4 am that day, maybe there isn't an important difference between the predictions and the realized values.

Answer 1

I'm not aware of any datasets that are quite what you want (archived point forecasts for the next day). There is less digestible data out there though.

• NOAA NCDC archived messages

  This website has archives of all messages that NOAA transmits. The vast majority are unrelated to your needs, but if you look through the message types you might find something worthwile.

• NCDC NOMADS

  This site has archived GFS model grids. This model is run 4 times per day and each run will have an analysis grid (current weather) and forecasts every 6 hours through the first few days of forecast time.

Most other archived data I could find are either observations or model reanalysis datasets, which won't work for your needs (except to better define the initial day t). With the above datasets you could potentially put together the current weather at a specific place at 4 different times on day (t) and from each time have forecasts data for the next day, e.g.:

• Observations / model analysis 12Z day t: 24 hour forecast for 12Z day t+1
• Observations / model analysis 18Z day t: 18 hour forecast for 12Z, 24 hour forecast for 18Z day t+1
• Observations / model analysis 00Z day t: 12 hour forecast for 12Z, 18 hour for 18Z and 24 hour for 00Z day t+1

From these you could then do your statistical analysis overall and explore how it may change as the day progresses and the model forecasts are shorter term.

If wading through this data is not something you want to deal with, an alternative is to start archiving forecast data yourself going forward and building up a database over time.

Some caveats...

The NOMADS GFS data is pure model data, no forecaster input. This is not what decision makers will be using directly. This data will be made more consumable and be tweaked by forecasters before dissemination to the public. If the NOAA message archive contains suitable forecasts data, that data will have likely been filtered through a forecaster and close to what would be displayed on the NWS website as a forecast.

School closings are chaotic. The same weather will not yeild the same closures. Sometimes schools will stay open when faced with a worse forecast than a time they recently closed. The criteria for closing may be a moving target as usable "snow days" are used up. School closing thresholds will vary regionally (Florida schools would be closed Dec - March if subjected to a Michigan winter). You may come up with a good forecast for when a school should close, but don't be surprised if it doesn't correlate well with actual closings.

Answer 2

The National Digital Forecast Database (NDFD) is archived by NCDC here: http://nomads.ncdc.noaa.gov/data/ndfd/

The NDFD contains gridded forecasts across the U.S. produced by NWS forecasters. You can also get NDFD output in realtime from a variety of sources, including the NWS itself: http://www.nws.noaa.gov/ndfd/technical.htm

Answer 3

Weatherspark.com, my favorite weather site of late, allows the user to pull forcasts from three sources: NOAA, met.no, and worldweatheronline. Both of the second options provide forecasts for the entire world, and have an API.

You can find information on the met.no API here and information on the WWO API here. In particular, the WWO API page says:

Our local weather API provides current weather conditions and hourly weather forecast for up to 15 days across worldwide locations. Local weather API also provides access to monthly climate average data to deliver aggregate weather across the world.

I have never tried using either service, so I can't promise that they will have exactly what you need, but hopefully this helps.