Improvement of MLP models for forecasting electrical energy consumption using OBD and OBS methods

In this paper we apply two reduction algorithms of a neural network architecture in order to improve the prediction quality of a multilayer perceptron network (MLP). The first algorithm is Optimal Brain Damage (OBD), whereas the second is Optimal Brain Surgeon (OBS). Our assumptions have been verified experimentally on the models for electricity consumption prediction using real data from the Polish electroenergetic system. Two series of tests have been carried out: the first is hourly forecast of electricity consumption for twenty four hours ahead, and the second is daily forecast of electricity consumption for one day ahead. Taking into account results of performed computations, we have found out that both algorithms OBD and OBS improve the prediction quality of an MLP network. Moreover, simplification of the network speeds up the training process. Presumably, we can assume that these conclusions can be expanded to other time series prediction tasks.