Thank you to everyone who joined FlexTrack Challenge 2025 Townhall #1. In this session, we introduced the challenge and explored how digital twin data from commercial buildings can be used to forecast demand response flags and capacities.
Hello. In the competition description, you state: “Participants are to develop a machine learning model that back-cast (from historic time-series data from buildings) to:…”
A back-cast is using present/future data to predict past values. You state around the 40-minute mark that this actually isn’t allowed.
Your competition description explicitly says back-cast but your Q&A states the opposite. Could we please get clarification?
Hello and thank you for this insightful presentation.
Thank you for clarifying that the Competition Phase ends earlier than October 19.
Is the correct date and time October 15 23:59 UTC as seen in the image that I have attached?
Thank you for sharing the fact that Site F is the Private Test Set.
I have one concern with this. We understand that our model should be generalizable to any site. With that in mind, is it fair for the final ranking to depend solely on performance at a single site?
The competition overview says “the private test set will be used for the final ranking.” I fear that it may be possible to design a model that does very well on Site F but is not generalizable to other sites. If there will be additional factors considered when determining the final winners, we would appreciate learning more about them.
Hi @liberifatali in the video at time 21:20, the slide shows that Site F is the private test set and Site E is the public test set. This also implies that our submitted predictions for Site D Capacity are never evaluated. My concern is that the final winning teams will be determined solely by Site F when the goal of the competition is to design a model that works for any site, and I’m asking the organizers if this is a valid concern.
About the back-cast question:
At time 38:30 a question arises that @jack_vandyke and I believe can be interpreted as “are we allowed to use data from timestamps > T when we are predicting for timestamp T?” At time 39:27 Matt answers no to this question, implying that we’re supposed to only use data from timestamps <=T. As the competition overview does not mention this rule, we are awaiting clarification on this topic from the organizers.
Thank you very much Sneha for responding. On the topic of the competition phase ending, I would like to kindly highlight that in the video at time 26:00 Emily says the competition phase “should end a week before the 19th of October,” and I’m wondering if that is true or if we can trust the overview page that states it ends on the 19th.
The competition overview clearly addresses the t+1 problem!
“This challenge focuses on identifying and estimating demand response activity. Participants are to develop a machine learning model that back-cast (from historic time-series data from buildings) to:
determine when demand response events were activated and for how long,
determine how much energy was increased or decreased (over the event duration), compared with normal consumption, as a result of activating demand response mode.
Participants will use ground truth time-series data with known observed demand response events (identified in the form of demand response flags) to learn site consumption behaviour both (i) when demand response mode is not active and (ii) when demand response mode is activated.”
Backcasting in machine learning is the process of generating or predicting unknown historical data using present information. How exactly are we supposed to interpret the quoted text???
Watch the recording: https://youtu.be/oKBcMxAQ3vg
Download the slides: Flextrack Townhall – Google Drive