\cite{Dotzauer_2002} develops a simple model with a piece-wise linear temperature component and a stochastic component with 1 parameter for each hour of the week. This model is fitted to actual load data (least square) and used to predict the load. The conclusion is that a simple model like this provides results that are as good as more elaborate commercial tools, but the results are sensitive to the input temperature. Therefor good temperature forecasting is key to good load prediction.
This can be a motivation for using an ensemble of forecasts when planning the load.