Several spatio-temporal load forecasting approaches are proposed for residential units.

The proposed approaches exploit sparse relational patterns among the time series of houses and behavioral similarities between end-users.

A “decompose-forecast-aggregate” framework is proposed to further improve the forecasts.

Using Pecan Street datasets, we testify our methods on real data recorded from 173 houses in Austin, TX.

The proposed methods significantly improve forecasts compared to the considered benchmark methods.