Threshold for true positive, false positive, ...

27 views
Skip to first unread message

Pedro Nascimento

unread,
Jan 26, 2016, 9:04:08 AM1/26/16
to Energy Disaggregation
Hi everyone,

how the threshold for calculating true positive, false positive,... is chosen for NILM algorithms? 
I mean, depending on this threshold the results mean very different things.  Is it always chosen arbitrarily?

Best,
Pedro.

Jack Kelly

unread,
Jan 26, 2016, 9:11:24 AM1/26/16
to Energy Disaggregation
Hi Pedro,

I'm probably being silly but I'm not sure what you mean by "the threshold for calculating true positive".  Do you mean the on-power-threshold for each appliance?  i.e. the power threshold between the appliance being 'off' and the appliance being 'on'?

Thanks,
Jack

Pedro Nascimento

unread,
Jan 26, 2016, 9:34:20 AM1/26/16
to Energy Disaggregation
Hi,

Exactly!! My explanation was very poor. You explained perfectly.

Thanks,
Pedro.

Jack Kelly

unread,
Jan 26, 2016, 10:06:15 AM1/26/16
to Energy Disaggregation
For some appliances it's easy to select an on-power-threshold.  UK kettles draw up to about 3kW so I usually pick a threshold of about 1kW.

Appliances like washing machines can be a little trickier because they use a small amount of power (<100 watts) during some parts of the cycle.  I can't speak for others but I generally find that I need to look through a fair amount of data in order to select valid 'on-power-thresholds' for these sorts of appliance in each house.  I find that the on-power-threshold needs, ideally, to be selected manually for each house because some houses have spurious loads on some channels.  For example, in house 5 in UK-DALE the washing machine channel rarely drops below 15 watts which I presume is because there was an always-on appliance plugged into that channel.  These hand-picked on_power_thresholds are encoded in UK-DALE's metadata for each appliance.

Pedro Nascimento

unread,
Jan 26, 2016, 11:30:41 AM1/26/16
to Energy Disaggregation
Thanks for sharing your experience Jack!! It's always great to learn from you!
Reply all
Reply to author
Forward
0 new messages