quarterly 13 weeks on friday 13:00 to 15:00 (next one dec. 6th)
see slides for agenda
15 min
Progress at Alliander
Daan
forecast evaluation over horizons: forecasts do get a bit worse over time, but within a single day the performance stays the same
improvement idea is to work on a separate model using satelite image-based radiation data to get better forecasts closer in time on the same day
Created Quality gate to label our forecasts based on quality "KTP seal of approval"
Idea that popped up: retrain based on these metrics
We create multiple metrics (probibablistics, deterministic), idea is to use an AND gate on thresholds for these metrics
How much training data do you need plateau the quality metrics? -> Depends on the situation, we currently train on 90 days and are looking into more optimal ways/days. Evaluation is based on last 2 weeks
Migrating away from refrence implementation into Dagster. Not decided yet if it makes sense to update the reference implementation to Dagster.
Pros: allows innovation on the infa part
Cons: Depends a lot on IT infrasturcture, less focus on Datascience part
Suggestion: better communication on difference on reference implementation and openstef
Deep learning model (S4) is starting to being tested in production
Ongoing evaluation, no clear conclusions, but looks promosing
Unclear on how to integrate this into OpenSTEF, many ways to approach this!
10 min
RTE update
Maxime
Checking for data completeness in influx dashboard
Others: no experience with influx ui dashboard
We use grafana, because influx v2 (and so the dashboarding) is quite new
Influx looks really nice!
Question on source_run and created tag
not used, but artifacts of internal data fetchers code
Question on use of weather_tahead
Currently unused, but good to use the co-coding session to make something
10 min
Sigholm update
Jan Maarten
new features that might help on peak detection, based on difference between own model and openstef. TBD to include it
holiday features now configurable, PR sent out for review