SINGLE ANN ARCHITECTURE FOR SHORT-TERM LOAD FORECASTING FOR ALL SEASONS

A single neural network architecture capable of forecasting the load o f week days, weekends days and holidays for all seasons pf the year wi th similar accuracy is described in this paper. Once trained, it can f orecast the load of a power system twenty four hours ahead on an hourl y basis for upto one week. The proposed architecture is a two layer fe ed-forward network with back-propagation learning algorithm. It is tra ined with three weeks' data prior to the forecasting week and provides an ongoing forecast. Using this network forecasts for all seasons hav e been achieved with an average error of slightly over one percent for week days, weekend days and holidays.