Load-Forecasting Challenge 2026: six weeks, eleven teams, one forecast every day

Published

July 20, 2026

Graphic for the Live Load-Forecasting Challenge 2026 with a stylized load curve

Six weeks, eleven teams, one forecast every day: the Live Load-Forecasting Challenge 2026 at TH Köln.

2026-07-20

The Live Load-Forecasting Challenge 2026 at TH Köln has crossed the finish line. Since June, eleven student teams submitted a fresh 24-hour forecast of Germany’s national electricity demand every day before midnight. Scoring was automatic the next day, measured against the actual grid load published by the ENTSO-E Transparency Platform. No toy dataset, no second chances: a live, fully reproducible forecasting competition under real-world conditions.

The challenge was part of the module “Numerische Mathematik” (Numerical Mathematics, summer semester 2026) taught by Prof. Dr. Thomas Bartz-Beielstein at the Faculty of Computer Science and Engineering Science of TH Köln. In the final presentations, the teams presented their approaches, from carefully engineered gradient boosting to modern time-series foundation models. The winning teams of the two award tracks will be honored soon.

Five students presenting in front of two screens showing load-forecast curves

Final presentations at TH Köln: a team presents its daily 24-hour forecasts.

The challenge also doubled as a lesson in open, verifiable research: the teams published model cards, ran OpenSSF security scorecards on their repositories, and cross-certified each other’s results through independent reproduction.

A team presenting the OpenSSF scorecard results of its repository

Open, verifiable research: OpenSSF scorecards were a mandatory part of the challenge.

The challenge also yielded valuable insights into forecasting energy demand in Germany: how much calendar structure — public holidays, bridge days, day types — and weather covariates really contribute, and how carefully engineered gradient boosting holds its own against modern foundation models. More details, including the full methodology and results, will be published soon.

Three students in front of screens showing forecast and load curves

Forecast versus actual grid load: the teams analyzed their submissions day by day.

Impressions from the final presentations

Team in front of two screens showing load-forecast curves
Team of five opening its final presentation
Bar charts on the influence of the amount of training data
Slide on heating and cooling degree hours with time-series plots
Two students preparing their presentation
Two students at the lectern
Duo presenting its 24-hour load forecast for Germany
Title slide of a 24-hour load-forecast presentation
Analysis of an outlier Saturday in the MAE curve
Discussion of the MAE trajectory of the daily submissions
Slide on a backtest over 13 months
Lessons-learned slide
Team of three starting its day-ahead forecast presentation
Title slide of a day-ahead forecast of the German electricity load
Slide on security and compliance
Team of four presenting its machine-learning load forecast
Team at the laptops in front of the title slide
Slide on the data pipeline from ENTSO-E data to the forecast
Team of five in front of the title slide of its load forecast
Group photo of a team in front of the screen
Slide on validating the weather-data cache
Audience following the presentation on the weather-data cache
Team of two opening its final presentation
Talk on the load-forecasting challenge
Slide on how the model evolved over the season

The challenge was based on data from the ENTSO-E Transparency Platform; the challenge infrastructure was provided by the SpotSeven Lab at the Institute for Data Science, Engineering, and Analytics (IDE+A) of TH Köln. Contact: Prof. Dr. Thomas Bartz-Beielstein.