I. Introduction
Accurate power demand forecasting is important for maintaining a stable power grid. With advance warning of demand surges, energy providers would be able to better plan their power generation and/or perform other measures such as peak shaving or load shifting [1], [2]. Various forecasting methods have been proposed based on extrapolation [3], [4], Kalman filtering [5]–[7], fuzzy logic [6], [8]–[10], autoregressive integrated moving average (ARIMA) models [11]–[14], artificial neural networks (ANN) [6], [11], [15], and similar profiles load forecast (SPLF) [16], [17]. All these methods attempt to forecast the aggregate power (all loads combined) directly by relying on the temporal dependence of the aggregate power signal. However, we note that stronger temporal dependence may exist in power signals of individual appliances. This is easily seen in the case of cyclical appliances such as refrigerators, which turn ON and OFF roughly periodically. When the power signals of different appliances are aggregated, such temporal dependence may be disrupted, hence forecasting the aggregate power may be more difficult than forecasting the power usage of individual appliances.