appliance-level power

The measurement is the core of the load-monitoring datasets in the household-appliances slice. REFIT Electrical Load Measurements (Cleaned) records 8-second interval power readings per house, split into an aggregate feed plus up to nine appliance-level channels. UK-DALE, distributed through CEDA/UKERC EDC, logs appliance-level power at roughly 6-second intervals with 1-second aggregate data for three homes.

Collection topology varies as much as cadence. Tracebase - Appliance-Level Power Traces takes the direct approach: one CSV per metered wall outlet, grouped into roughly 43 appliance-type folders. Pecan Street Dataport tags its records 'appliances' and 'disaggregation', signaling that its households support both submetered and inferred views.

Paired aggregate-plus-channel structure is what makes these datasets the training substrate for non-intrusive load monitoring (NILM), where a model learns to split a whole-house feed into appliance contributions.

Sampling interval and channel count bound what any downstream model can see. A 300-watt spike lasting two seconds may vanish between 8-second samples; a fridge cycle needs minutes, not seconds, to characterize.

• Disaggregation accuracy is capped at the data's cadence. NILM trained on 8-second REFIT channels cannot learn dynamics faster than that interval records.
• Channel inventories define appliance coverage. Up to nine channels per house (REFIT) or roughly 43 appliance types (Tracebase) is the ceiling of what supervised models can distinguish.
• Aggregate-only delivery hides the ground truth. Without paired per-appliance channels, claimed disaggregations cannot be validated.

Use these checks to test whether a source delivers what the concept promises:

1. Read the sampling interval first. REFIT samples every 8 seconds; UK-DALE logs appliance channels at roughly 6 seconds with 1-second aggregate for three homes - match cadence to the loads you care about.

1. Count the appliance channels. Up to nine per house in REFIT; wall-outlet metering across roughly 43 appliance types in Tracebase.

1. Require aggregate-and-channel pairing. Datasets tagged for disaggregation, like Pecan Street Dataport's, are valuable because the whole-house feed and the submetered channels refer to the same homes.

1. Check household counts against claims. UK-DALE's 1-second aggregate applies to three homes specifically - granular cadence is not uniform across every record.

Where to see it in context: household-appliances data.

The vocabulary around load monitoring:

• nilm
• whole-house-aggregate-power