About the Competition
To plan the power production ahead of time, the Norwegian transmission system operator, Statnett, requires accurate predictions of expected wind power production in different geographical zones. The WindAI Challenge was a competition to develop more accurate and robust models for forecasting wind-power production in Norway, two days in advance. This is crucial for grid stability and efficient energy management.
- The competition had many purposes in addition to solving an industrial challenge, it has a large component of training students to work with data science in the energy sector. Several teams participated as student projects contributing with innovative ideas, and the challenge can freely be used in teaching, says Signe Riemer-Sørensen (Sintef), co-organiser and head of the evaluation jury.
The challenge was to provide a model framework for predicting wind power production two days ahead, based on public data, that can automatically adapt to changes in installed capacity. But the organizers also asked teams to think beyond pure accuracy and address some big real-world questions: Can your model handle changes in weather and climate? Can it explain how it makes predictions? Can it predict uncertainty — how sure is it? And can it run efficiently, using minimal computing power?
As submissions closed in late September early October, Storm Amy approached Norway’s coasts and brought with it a reminder of the power of wind and the critical importance of accurate forecasting – including protecting infrastructure.
- We launched the WindAI competition to generate novel ideas and models for enhancing our wind power forecasts. Additionally, we aimed to spark broader interest in the field of forecasting by publicly sharing our datasets.
The teams provided valuable insight into their work through a minipaper and a video presentation. These materials demonstrated how various models and features affected the results and showed the feasibility of predicting confidence bands in the forecasts. The top teams utilized gradient boosting models, illustrating that the most advanced models are not always the best in practice. We were surprised that many teams used time features as input, a factor we had not anticipated would affect the results.
Overall, we were pleased with the effort and dedication the teams invested in this competition, and we look forward to testing out the best insights from the top submissions into our own forecasting models, says Kristian Authén, Senior Data Scientist, Statnett.
The participating teams were evaluated based on accuracy, trustworthiness & explainability, implementation & presentation, robustness to changes and computational requirements. The latter being used as a tie-breaker if needed.
There was huge interest in the competition. There were 61 registered teams, of which 9 participated in the full challenge.
About the winners
The winner of the competition is EquiWindNerds who will receive 50 000 NOK in prize money and the opportunity to showcase their solution to Statnett. Members of the team, all Equinor Employees, were Nicolas Petitclerc, Thibaut Forest, Øyvind Steensland, Robert Dibble, Malavika Ramakrishnan and David Augustin.
Runners up were the teams Knowit and WindSight. The teams Brew-Booze and Bits also submitted excellent answers and were praised for novelty and suitability of the methods to the problem.