Experience with FIR BP neural networks for prediction problems

Paper deals with experience of application of FIR filter (Finite Impulse Response) in BP neural networks used for prediction system. Prediction can be defined as extrapolation of unknown function determine by representing data sets. One of the most difficult problems in prediction systems is determination of the size of the time series variable input. It is always very difficult to determine the right size of the history of input variable necessary for prediction - extrapolation of the modeled function. So in general prediction problem can have input e.g. ( x(t), x(t-1), ...,x(t-τ)), where τ - determine the width of the input window size and output (y(t+1), y(t+2), ..., y(t+λ)), where λ - is size of the length of the time series of the predicted variable and x and y are vectors. Usually both τ and λ depend on the problem and in case of output the size of the λ is small then there is bigger chance that prediction will be more reliable. The avoid a problem of τ determination a FIR neural networks offer solution to gather history of the unknown function inside of neural network topology, namely on neural networks synapses in the form of FIR filters. The length of the filters indicates the time-delay of the input. Some modification of FIR concerning Adaptive FIR NN was under research and self-adaptation of length of particular FIR filters associated with each synaptical weight. Experiments were done on selected SANTA-FE data for prediction problems and also on real-world data related to power engineering domain with aim to make electricity load forecast possible. Results indicate that FIR NN is an interesting tool for prediction problem with possibility to avoid determination of the input window size. Comparison study between FIR NN and Adaptive FIR NN shows some advantages of adaptive FIR over the non-adaptive approach.