From Events to Appliance Identity: Neural-Baseline Statistical Association for Non-intrusive Load Monitoring

A building’s electricity meter records only total consumption. Decisions about waste, operating schedules and maintenance concern individual appliances. Separate appliance meters can supply that information, but they bring the cost of extra sensors and installation. Software metering reconstructs each appliance’s power from shared electrical measurements instead. When an appliance contains several independently switched components — for example a freezer’s compressor and defrost heater — two discoveries become necessary: which changes in the measured supply belong to each component, and which components belong to the same appliance.

This paper presents Holmes, the evolution of NoWatt’s software-metering programme first demonstrated on operating data in 2009. Holmes works in two stages. The first constructs switching events from changes on one to three phases, learns each component’s switch-on and switch-off sizes separately, and reconstructs its runs. It extracts one pattern at a time, removes the pattern’s power from the measured supply, and repeats on the residual. The second introduces a neural-baseline statistical association method for discovering which extracted patterns belong to the same appliance. Small self-supervised networks learn each pattern’s expected switching behaviour and run duration from its own history. Statistical comparisons then identify systematic departures from these learned baselines during or following another pattern’s operation: switching on more or less often while the other runs, or running longer or shorter straight after it. The training targets are the pattern’s own extracted events; no appliance labels or reference-meter data enter the training.

In a seven-day commercial-kitchen evaluation, seven patterns are recovered and joined into four appliances. A freezer’s defrost heater and a dishwasher’s heating element, about 4 % apart in switch-on size, are separated and each assigned to its own appliance. Association reduces time-integrated absolute error from 30.4 % to 8.9 % for a dishwasher, from 30.1 % to 14.6 % for a fridge, and from 13.5 % to 10.2 % for a freezer. A grill is recovered with 0.9 % absolute error. A six-day refrigerator example from the public REDD data set demonstrates extraction on a residential supply.

The contribution is a method that connects measured events to appliance identity through neural-baseline association. It advances software metering from the recognition of stored signatures to the discovery of component behaviour and appliance relationships from the aggregate itself.

**Keywords:** energy disaggregation; non-intrusive load monitoring; event identity; self-supervised learning; neural-baseline association; temporal relationships; software metering