This paper reports on the design and implementation of a short-run for
ecasting model of hourly system loads and an evaluation of the forecas
t performance. The model was applied to historical data for the Puget
Sound Power and Light Company, who did a comparative evaluation of var
ious approaches to forecasting hourly loads, for two years in a row. T
he results of that evaluation are also presented here. The approach is
a multiple regression model, one for each hour of the day (with weeke
nds modelled separately), with a dynamic error structure as well as ad
aptive adjustments to correct for forecast errors of previous hours. T
he results show that it has performed extremely well in tightly contro
lled experiments against a wide range of alternative models. Even when
the participants were allowed to revise their models after the first
year, many of their models were still unable to equal the performance
of the authors' models. (C) 1997 Elsevier Science B.V.