You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I just completed an end-to-end deployment test with @JanghyunJK:
Exported data from SkySpark
Trained a model
Received the model back
Deployed the model in SkySpark
Called prediction against the deployed model
Synced predictions to SkySpark points
Ok, so first off, this deployment worked great (yay!). But...
The issue is that the prediction seems to have a time zone offset. It isn't a simple shift in the output, though, because then you would expect things to line up just by shifting left. Instead, it seems that perhaps training assumed input timestamps in UTC and prediction assumed UTC-7, or vice versa? I'm surmising this because it appears some predictors are being correctly accounted for (decent fit at some times of day), but the time-of-day predictor is time-shifted (poor fit at other times of day when irradiance has little effect).
I'm not entirely sure how to go about troubleshooting this. Things I do know already:
Original predictor data are delivered in Denver time (MST/MDT) and timestamps are encoded as such
Required start/end times as reported by the Wattile model appear to work correctly, as I can pass the requested window and get a single prediction timestamp out
SkySpark converts SkySpark datetimes to Pandas datetimes complete with timestamp, and converts back the same way; timezones appear to be preserved properly in both directions as shown below.
The text was updated successfully, but these errors were encountered:
In today's meeting we decided that timezone standardization to UTC should happen within prep_for_rnn instead of (or in addition to) loading training data from CSV
I just completed an end-to-end deployment test with @JanghyunJK:
Ok, so first off, this deployment worked great (yay!). But...
The issue is that the prediction seems to have a time zone offset. It isn't a simple shift in the output, though, because then you would expect things to line up just by shifting left. Instead, it seems that perhaps training assumed input timestamps in UTC and prediction assumed UTC-7, or vice versa? I'm surmising this because it appears some predictors are being correctly accounted for (decent fit at some times of day), but the time-of-day predictor is time-shifted (poor fit at other times of day when irradiance has little effect).
I'm not entirely sure how to go about troubleshooting this. Things I do know already:
The text was updated successfully, but these errors were encountered: