Forecasting electricity demand for a single household is much harder than forecasting it for a city. A city averages out into something smooth. One home is somebody's routine, and routines differ wildly: a night shift worker, a family with school runs, a house that empties every weekend.

The usual approach treats that behavioural pattern as a label you sort houses by first, then forecast within each group. This paper asks whether the behaviour can live inside the forecasting model instead of outside it. Their framework infers a discrete hidden variable representing behavioural structure from whatever context is available, and feeds it directly into the part of the model that produces the forecast.

The distinction sounds academic but it changes what the model can do. A grouping decided in advance is fixed, while something inferred from context can shift as the evidence does. Each household's load profile is treated as its own small forecasting task rather than as one more row in a shared table.