The forecasting community has long recognized the importance of comparing a forecast to some kind of reference forecast. The choice of reference forecast is critical for a fair comparison. Fair comparisons of test forecasts against reference forecasts can help forecast users and developers understand the relative merits of models. The most appropriate reference will depend on the use case and forecast horizon.
The Arbiter provides a selection of built-in benchmark forecasts for publicly available reference data, but it ultimately leaves the selection of reference forecasts to the user.
Built-in benchmark forecasts
To guide our selection of the built-in benchmark forecasts, we first identified the following desirable attributes of a benchmark forecast:
- Available throughout the US
- Freely available or easily implemented
- Provide quantities of interest to both forecast users and providers
- Stakeholder buy-in
The Arbiter includes the capability to create and evaluate benchmark irradiance, power, and net load forecasts.
The Arbiter provides built-in support for the following benchmark
options. Additional implementation details are available in the
- Persistence of clear sky index
The Arbiter provides reference implementations of persistence algorithms. The averaging window of the Arbiter’s benchmark persistence forecasts are fixed to be equal to a forecast’s run length, but no longer than 1 hour. For example, a benchmark forecast for a 15 minute interval will use the most recent 15 minutes of data to compute the quantity to persist (irradiance, power, or clear sky index). Persistence of clear sky index accounts for the average clear sky index during an interval rather than simply using the interval start or end time.
A week-ahead persistence forecast may also be configured for net load applications (e.g. net load of the next Sunday will be equal to the last Sunday).
- HRRR subhourly irradiance
- RAP cloud cover to irradiance
- NAM cloud cover to irradiance
- GFS cloud cover to irradiance
With the exception of the HRRR, the Arbiter derives irradiance and power forecasts from NWP cloud cover forecasts to accurately account for solar position, inverval labels (beginning or ending), and interval averaging. For these models, the process is:
- Load hourly (or longer) interval data from the NWP grib files.
- For GFS cloud cover, unmix the mixed-intervals average data.
- Resample data to 5 minute intervals.
- For GFS cloud cover, backfill the data.
- For all other NWP data, interpolate the data.
- Convert cloud cover to irradiance following Larson et. al. (2016).
- If desired, use site metadata to compute AC power using pvlib-python functions.
- Compute hourly averages with desired interval labels.
NWP-derived forecasts are not currently available for net load.
- GEFS cloud cover to ranked irradiance or power ensemble
- Persistence ensemble (irradiance, power, net load)
The Arbiter automatically produces benchmark forecasts for every reference forecast site in the Arbiter’s database. Built-in benchmark forecasts are currently restricted to the public dataset, though we hope to add support for user-specifed sites in the future.
The Arbiter creates persistence forecasts with 5, 15, 60, and 1 day lead times. These forecasts are created for each variable that has a corresponding observation at a site. The Arbiter also creates week-ahead persistence forecasts for net load1.
For intraday and day ahead lead times, we chose to define forecast attributes for a GFS Day Ahead, NAM Current Day, HRRR Intraday, and RAP intraday forecast in the local timezones of the sites (Daylight Saving Time not considered). For example, for the Table Mountain SURFRAD site, the following parameters apply (interval label is ending and interval length is 1 hour for all):
|Model||Issue time of day||Run length / Issue frequency||Lead time to start|
|GFS day ahead||7Z||1 day||1 day|
|NAM current day||6Z||1 day||1 hour|
|HRRR intraday||0Z||6 hours||1 hour|
|RAP intraday||0Z||6 hours||1 hour|
Note that, as described in forecast attributes, the combination of issue time of day and run length/issue frequency may describe more than one forecast run per day. Here, the HRRR and RAP forecasts are issued at 0Z, 6Z, 12Z, and 18Z. The Arbiter creates irradiance forecasts at every reference site and also creates power forecasts at the reference power plants.
The Arbiter also creates benchmark probabilistic forecasts. The GEFS drives a day ahead forecast and a persistence ensemble is used for an hour ahead forecast.
User-supplied reference forecasts
Some evalulation applications may require users to provide their own reference forecasts. For example, some Solar Forecasting 2, Topic Area 2 teams will run an earlier version of WRF Solar and then upload its forecasts for existing or new evaluation sites. Or a forecast user may choose to use its existing forecast vendor as a reference. Users may then identify the forecast as a reference when creating a forecast evaluation report so that forecast skill metrics will be calculated with respect to the reference.
Users are encouraged to track metadata about the forecast using the extra parameters field of the forecast creation form. Key modeling details such as grid spacing and schemes, or your internal version of a namelist, may be included in the extra parameters field. As explained in the Data Model, the Arbiter will accept uploaded forecasts for predefined evaluation sites or aggregates. The Arbiter will not accept gridded datasets.
Trial-specific benchmark forecasts
The Arbiter’s operational forecast trials allow for any configuration of the built-in benchmark forecasts or user-supplied reference forecasts. The Arbiter administrators can help prospective trial users determine the most appropriate reference for their application. Contact firstname.lastname@example.org for more information on trials.
Net load is defined here as true system load minus behind the meter PV. Net load is the load that must be served with utility scale resources, regardless of whether or not they are dispatchable, conventional, or renewable. We make no allowance for wind power or utility scale solar power. Therefore, net load is not equal to the load that must be served with conventional generation. ↩