The Solar Forecast Arbiter supports operational forecast trials, including features such as:
- Multiple anonymous vendors
- Built-in benchmark forecasts
- Penalties for missing forecasts
- Rejection of late forecasts
- Daily reanalysis of forecast performance
The Solar Forecast Arbiter does not declare trial winners nor declare that one forecast is superior to another. It simply provides an impartial calculation of metrics following clearly defined data processing and forecast specifications.
In addition to the material on this page that is specific to the Solar Forecast Arbiter, we recommend that users review the IEA Wind Task 36 Work Packages for information on weather forecasts, benchmarking, and optimal use of forecasting solutions. The vast majority of the material applies to both solar and wind forecasts and will aid users in defining their forecast requirements.
All participating organizations must sign the Data Use Agreement before a trial can begin.
Initiating a trial
Forecast users may contact firstname.lastname@example.org with a request to initiate a trial. While the trial does not need to be fully-specified when contacting us, it will be most effecient if your request includes information about:
- Sites and variables
- Start and end dates
- Forecast lead time to start
- Forecast interval length
- Forecast issue time of day
- Metrics (MAE, MBE, RMSE, Skill, etc.)
- Time categories (total, day, hour, etc.)
- Reference forecasts
- Data quality flags to exclude
- Rule for missing forecasts (exclude, fill with 0, fill with last valid value)
- Anticipated outcome at trial conclusion (e.g. award contract to vendor)
See Forecast Definitions for an explanation of how the Solar Forecast Arbiter defines forecast parameters. Multiple sites and forecasts (e.g. hour ahead and day ahead) may be analyzed in a single trial.
The Solar Forecast Arbiter administrators will help you refine your requirements if needed.
Next, the Solar Forecast Arbiter administrators will email the proposed requirements to our participating forecast vendor list. Forecast vendors will be given an opportunity to provide feedback on the trial design to the end users. After some iteration, the Solar Forecast Arbiter administrators will finalize the trial requirements and present them to vendors to opt in.
Vendor opt in
The Solar Forecast Arbiter administrators will email a final set of trial specifications to forecast vendors. The specifications will include:
- Start and end datetimes.
- All metadata needed instantiate the relevant Site, Observation, and Forecast objects in the Solar Forecast Arbiter or links to existing metadata. This fully defines the forecast specifications.
- All metadata needed to instantiate reference forecast objects or links to existing reference forecasts.
- All metrics to be computed.
- All time categories over which metrics will be calculated.
- Observation data quality flags that will be excluded from analysis.
- Rule for missing forecasts (exclude, fill with 0, fill with last valid value).
- Rules for visibility of forecast time series with other anonymous vendors:
- during the trial and
- in the final report.
The specification will also include the forecast user’s anticipated outcome at the trial conclusion, but we emphasize that the Solar Forecast Arbiter team will play no role in facilitating that outcome.
Forecast vendors respond with an indication that they would like to participate in the trial or decline to respond.
The Solar Forecast Arbiter will create a new anonymous user account for each vendor for each trial. The mapping between the trial participants and the anonymized names is created automatically so that the administrators do not know the true identities. An automated email will be sent to each trial participant with the anonymous username, a password reset link, and a securely signed text file. This text file can be used to verify that a trial participant was assigned a particular anonymous user account. It should be kept in a safe location.
The Solar Forecast Arbiter will automatically create all required forecast metadata for each forecast vendor. This also defines the API endpoints that forecasters will use to upload data during a trial.
Forecast vendors use the anonymous accounts to log in to the dashboard and to obtain tokens for use with the API. When logging into the dashboard with the anonymous account, users will only see the metadata and data objects that are relevant to this trial. All other user-supplied data and reference data is not accessible by an anonymous trial account, but is of course still accessible with a standard user account.
Once the trial begins, forecast users and vendors upload their data using the Solar Forecast Arbiter’s RESTful HTTP API. Users upload their observation data using the Add Observation Data endpoints and vendors upload their forecast data using the Add Forecast Data endpoints or Add Probabilistic Forecast Data endpoints. The endpoints are unique to each forecast for each vendor. The Solar Forecast Arbiter’s data administration rules prevent other users from uploading data to each vendor’s forecasts. Depending on the trial design, other vendors may or may not be able to view the data.
Forecast vendors must upload their forecasts before the closing time of any forecast run. Forecasts that are not uploaded are handled according to the trial specification agreed to in vendor opt in.
Forecast vendors may upload data for intervals beyond what is required at any
given time. The data may be overwritten until the closing time of the forecast
run. Consider a 1 hour lead time, 1 hour interval forecast with
hour-beginning interval labels. In this example, a forecast vendor need only
upload a single value each at each hour of the trial.
For instance, by
2020-01-01T00:00Z a vendor may upload the forecast
# valid upload at any time until 2020-01-01T00:00Z timestamp,value 2020-01-01T01:00Z,10
The vendor could choose to upload additional forecast values such as
# valid upload at any time until 2020-01-01T00:00Z timestamp,value 2020-01-01T01:00Z,10 2020-01-01T02:00Z,20 2020-01-01T03:00Z,30
Now imagine we advance in time to the next forecast interval.
2020-01-01T01:00Z the vendor may change
the value of all forecasts after
2020-01-01T01:00Z. A valid upload could be
# valid upload at any time until 2020-01-01T01:00Z timestamp,value 2020-01-01T02:00Z,25 2020-01-01T03:00Z,35 2020-01-01T04:00Z,45
The API will reject an upload that contains a time that should have already
been issued. For example, during this interval, the upload would be rejected
if it contains any times before
# invalid upload after 2020-01-01T00:00Z timestamp,value 2020-01-01T01:00Z,15 2020-01-01T02:00Z,25 2020-01-01T03:00Z,35 2020-01-01T04:00Z,45
Assuming no additional uploads, the time series used for evaluation would be
# forecast evaluation time series timestamp,value 2020-01-01T01:00Z,10 2020-01-01T02:00Z,25 2020-01-01T03:00Z,35 2020-01-01T04:00Z,45
See the Example script section below for an example of how to program against the API for a trial.
During a trial, the Solar Forecast Arbiter will automatically compute error statistics at a predetermined frequency, typically once per day. The results will be available in a standard Solar Forecast Arbiter report listed under the Dashboard’s Reports section. The reports will compute all metrics for all categories for all forecasts as specified in the trial design. The results are overwritten each time the reports are recomputed.
At the conclusion of a trial, the Solar Forecast Arbiter administrators will email all participants to let them know that the final analysis report has been generated.
We urge users to download html and pdf copies of the report. These downloads contain a PGP signature that may be used to verify the authenticity of the report. The Solar Forecast Arbiter is not a long-term data repository and we make no guarantees about the length of time we will store trial results.
Forecast vendors may or may not choose to directly contact the forecast vendor and reveal their identities (identity may be proven using the signed text file provided when the anonymous account was created). The forecast user may or may not decide to pursue the recorded anticipated outcome. The Solar Forecst Arbiter plays no role in these processes.
Forecast vendor list
All forecast vendors are encouraged to register with the forecast vendor list
by contacting email@example.com.
Please include a link to your vendor website so that we may confirm that you
represent a legitmate forecast vendor. We accept requests for individuals
firstname.lastname@example.org) and lists (e.g.
Signing the Data Use Agreement is not required for addition to the vendor list, but you will not be able to participate in trials until the DUA is signed.
We urge vendors to practice using the API with test forecasts before a trial starts. An example script to upload forecast values for each of the user’s forecast objects in a trial can be found below and in this gist. We do not guarantee the reliability of this script for operational forecast trials - it is only an example to help you get started with the API.
This script uses the solarforecastarbiter-core library to interact with the Solar Forecast Arbiter API. First, a token for API access is requested using the username and password for the anonymous trial user. The script expects a path to a file with the username and password of this user seperated by a new line like
A list of forecasts is then retrieved from the API and filtered for those forecasts relevant to the Trial. For each of these forecasts, a check is performed to determine if the current time is within 10 minutes of the next issue time of the forecast. If it is, a random set of values is uploaded to the API for the expected forecast time range. Otherwise, the script moves on to trying the next forecast in the list.
To run the script, users can make use of the solarforecastarbiter-core Docker image which includes a Python installation and all requirements. Otherwise, the solarforecastarbiter-core Python package can be installed from the Github repository or via pip with the command
pip install git+https://github.com/solararbiter/solarforecastarbiter-core.git
The script should be run periodically to generate new forecasts, either using cron jobs or a cron Python framework like schedule. Further documentation for the solarforecastarbiter-core Python package can be found at https://solarforecastarbiter-core.readthedocs.io/en/latest/.