GAD Editorial · Published · · English study · Chinese summary · Study dated 2026-10-08
A household electricity bill can identify a large demand without explaining what an architectural retrofit should address. Appliance-level estimates may help a team decide which questions need better measurement before changing a kitchen, service cupboard or electrical installation. NILMTK offers a useful framework for testing those estimates, provided the design brief retains the difference between an inferred load and a measured circuit.
The original NILMTK paper, first submitted on 15 April 2014, describes an open-source toolkit for comparing non-intrusive load monitoring algorithms. This field separates aggregate electricity data into estimated appliance consumption. NILMTK itself is evaluation infrastructure, rather than one AI model. The core repository covers meters, preprocessing and metrics; nilmtk-contrib supplies compatible disaggregation models, including neural approaches. This is a retrospective application review dated 8 October 2026.
For a housing refurbishment, a defensible use would be to investigate whether recurrent evening demand deserves an appliance survey before the design team allocates space for new equipment. A model could produce a shortlist of suspected loads for discussion with the building-services engineer. It should not turn the same electricity trace into an unsupported diagnosis of inadequate insulation. The spatial consequence follows only after the end use and the proposed intervention have been checked.
Consider a proposed pilot covering six dwellings. The project team could obtain authorised aggregate records and temporary reference measurements for a small set of appliances. Four dwellings would support model development; two would remain unseen until the comparison is fixed. This split is an editorial test proposal, not a minimum sample size that establishes general reliability. A winter-only record would also leave seasonal operation unresolved.
Before training, reconcile meter clocks, units, sampling intervals and missing periods. Write down whether a label refers to one appliance or a circuit containing several devices. Keep any rooftop generation, battery activity and unmonitored consumption visible in the data description. Otherwise an apparently precise appliance chart can conceal a mistaken account of what the main meter actually measures.
Compare a selected neural model with a simple reference method using identical time windows. Report both time-series error and error in accumulated appliance energy. Examine periods when several devices operate together, and retain the residual between measured aggregate demand and the estimated components. A tidy total achieved by forcing all demand into known labels would hide the very uncertainty the retrofit team needs to see.
For the worked example, suppose the model assigns a repeated evening peak to cooking equipment. Ask whether the held-out reference meter supports both the timing and the daily energy total. If timing is plausible but the energy allocation is poor, the next useful action is a focused measurement exercise. Replanning the kitchen or resizing a supply from that estimate would run ahead of the evidence. Set acceptable errors with the engineer before viewing results, in units tied to the decision.
The architectural deliverable should be a short load-evidence sheet accompanying the brief: observed quantities, model estimates, unresolved demand and the measurement still required. Store only the household detail necessary for the agreed purpose. This gives AI a practical role in reducing uncertainty at the start of a retrofit while preserving the professional assessment needed to choose and size the work.