Forecasting electricity consumption at dierent locations
in electric distribution grids on short time scales
is a crucial ingredient of systems that will enable higher
renewable penetration without sacricing the security of
electricity supply. The goal of this paper is to evaluate
the performance of state-of-the-art forecasting methods
based on actual data. Overall, we observed that most
of the algorithms benet from larger training sets and
splitting the data into training sets of particular day
types. In addition we observed that predictions based
on disaggregated data from individual appliances lead
to better results. Generally, our analysis has revealed
that if the forecasting methods are applied without individual
tuning, they are able to beat the accuracy of
persistence forecasting only in rare cases. Furthermore,
the achievable accuracy in terms of average MAPE is
surprisingly low, ranging between 5 and 50% for one of
the considered data sets, and between 30 and 150% for
the other, more variable, demand prole. Our work thus
motivates more research investigating how accuracy can
be increased.