WindAI – Forecasting wind power production

Note

Dear forecasting teams!

We have become aware of a naming inconsistency in the MET nowcast file, two of the wind parks have slightly different names in the daily updated files than in met_nowcast.parquet used during training. These are:

• "Sørfj" have been renamed to "Sørfj. Vindkraft" in the daily updated files
• "Tonst" have been renamed to "Tonst. Vindpark" in the daily updated files

Our apologies for the mistake!

Leaderboard

Last updated : 2025-Oct-08 10:00

Motivation/ about

To plan the day ahead power production the Norwegian transmission system operator, Statnett, requires accurate predictions of expected wind power production for each Norwegian bidding zone.

Problem Statement/ RQ (Innovation, value and domain focus)

The challenge is to provide a model framework to predict wind power production 2 days ahead based on public data, that can automatically adapt to changes in the installed capacity.

The produced power will depend on characteristics of each wind park combined with wind speed, and parameters affecting the air density like pressure and temperature. These are predicted forward in time in the weather forecasts but may not always be accurate. 

The teams should address at least one of the suggested tasks but are free to supplement with their own ideas. All solved tasks should be explained and justified in the submission.

Tasks

We recommend an exploratory approach when solving the suggested tasks and encourage innovation when developing solutions.

Provide a wind power forecast for each bidding zone in Norway on an hourly resolution for the entire day 2 days ahead (24 hours in total) . The model may use the recent power production up until the first timestamp of the weather forecast. The model may also use unavailability messages given to the market (REMIT UMM) from NordPool.

Additional tasks: 

1. How can you demonstrate robustness of the model including robustness to changes in weather patterns/climate changes and installed wind power capacity?
2. How do you handle trustworthiness and explainability?
3. Can you provide reliable uncertainty predictions?
4. How lightweight can the model be in terms of computational resources? 

Evaluation setup and criteria

The submission will mimic a production setting. For the two weeks following the deadline [Sep 22-26 and Sep 29-Oct 3], all teams will deliver a daily forecast on workdays (10 in total, one can be missed).

The forecast should be uploaded to the allocated disk before midnight CET spanning the length of the MET MEPS weather forecast issued at 09:00UTC for that day (24-48 hours). Input data (weather, recent power production and UMMs) will be made available on a shared container. Submissions should be uploaded as csv-files following the convention in example provided with the data.

The predictions will be shown on a public dashboard (root mean square error and plots)

• The forecasts will be scored using the root mean square error metric.
• The top 10 performing teams on accuracy will be evaluated according to the following criteria.

The model will be evaluated on usefulness for Statnett:

• Accuracy 65%
  • Firstly, used to rank all contributions, after which the other factors will also contribute.
• Trustworthiness and explainability 20%
  • Has the team understood the task and provide meaningful models?
  • Does the model provide uncertainty on the predictions?
  • Does the team provide a benchmark against other models?
  • Does the team provide explanations for the model predictions?
  • Is it simple to understand how the data influence the model?
  • Is there an obvious simpler model with similar performance?
• Implementation and presentation 10%
  • Is the model easy to use? Can it easily be retrained?
  • Does the team explain the overall concepts in a good way?
  • Is it feasible for Statnett to implement the model in their daily operations?
• Robustness to changes (installed wind power capacity, climate changes etc.) 5%
  • How will the model handle changes to wind power infrastructure?
  • Is the model sensitive to climate changes?
• Other criteria that can be used to break a tie
  • Computational requirements such as prediction time of the model and required computational to train and run the model?

Submission

The upload link will be shared soon—please stay tuned for further instructions.

In addition to the 10 daily submissions, all teams should submit the following by Sep 20:

• A 5 min video: A video explaining the judges what you have done and why. Recording a slide presentation with Teams or Zoom is sufficient. The judges will be a mix of data science experts and power production experts. You can assume that they are familiar with the information given on this page.
• A short paper: A 2-3 page mini-paper (double column, plus 1 additional page for references) describing the method and results. The submitted papers are supposed to aid the judges understand the suggested method. The submitted papers will be offered peer review and if quality criteria are met, the papers can be included in a joint publication description of the competition results. Outstanding submissions will be invited to submit a full-length paper in the Nordic Machine Intelligence Journal. For more information about the NMI journal and to download templates, please visit the journal homepage https://journals.uio.no/NMI
• A zip file with self-contained code and a notebook explaining how to run the code.
• A pdf-file of the notebook.

To be considered for the prizes, participants must complete every task. Submitting just one sub-task is permitted, but entries that do so will not be eligible for any awards.

About the data

The competition will evolve around three open data sets:

• Weather forecasts (MET) - MEPS weather forecast containing 15 forecast simulations for each wind park.
• Weather observed (MET) - Now-cast weather for each wind park.
• Wind power observed (Statnett) - Observed wind power for each of the Norwegian bidding areas.
• Wind power metadata (Statnett) - Metadata for mapping wind park weather to bidding area. 

Detailed information about the datasets is provided in the accompanying PDF.

The complete datasets, together with a descriptive notebook, can be accessed through this link.

Computational resources

Norwegian residents may obtain free access to computational resources from the Norwegian AI Cloud.

Frequently Asked Questions

Read our FAQ page