Hybrid Neural Network System for Electric Load Forecasting of Telecommunication Station
Maurizio Caciotta, Sabino Giarnetti, Fabio Leccese · 2009
Abstract − This paper describes a neural network system for power electric load forecasting of telecommunication station. Getting an accuracy useful for contractual purpose a separately daily forecast of both main load and its oscillation is proposed. For the mean daily forecast we used a three layers multilayer perceptron (MLP), while to the oscillation forecasting we realized a system composed by a MLP and a self organizing map (SOM): the typology information obtained by the SOM unsupervised algorithm has been utilized as binary code in MLP input. The proposed system with hourly power load data of a big telecommunication operator has been tested. The total forecast has been obtained combining the two components. The forecasting accuracy for a whole year test data is around 2%. Some problem exists in the forecasted load of summer time.