Please note some key updates about the challenge:
Phase 1 (Competition Phase)
- Ends on Sunday, 19 October 2025.
- Includes submission of the Solution Documentation.
Phase 2
- Expected to open between 20 and 22 October 2025, 23:59 UTC.
- The Solution Documentation deadline is extended to this window.
Phase 2 Format
- Multiple one-year datasets will be sliced into shorter, equal-length context windows.
- Slices will include a mix of previously seen and new sites.
- In Phase 2, participants will make a single prediction per time series. This is to prevent look-ahead.
- Phase 2 is designed to ensure fairness and reward models that can generalise, are context-aware, and are transferable across sites.
Scoring and Final Ranking
- The final ranking will factor in two components with equal weight:
- the score from the Phase 1 private dataset (SiteF), and
- the score from Phase 2.
- Within Phase 2, scores will be the equal-weighted average across all slices and sites.
Scoring Code and Starter Kit
- We will publish the scoring script and a minimal code sample in the starter kit.
- These materials will be available when Phase 2 opens.
Challenge Description Clarification
- Participants must predict each timestamp t₀ using only inputs where t_input ≤ t₀.
- Training data includes ground-truth time series and demand response flags.
- Models should learn consumption patterns both when demand response is inactive and when it is active.
