Skip to Main Content
Neural networks with switching units were originally designed for classification tasks. However it was also shown that simple neural networks with switching units are capable to predict seasonal time series with results comparable to common stochastic methods. This paper presents enhanced model of neural network with switching units with aim to improve the forecasting performance of non-stationary time series. The presented model of neural network network is build of neurons with feedback and continuous activation function and it has a two level topology. The paper further describes the application of genetic algorithms to the optimization of the first level of topology. Finally, the performance of the proposed model was tested on the time series of currency in circulation and two artificial seasonal stochastic processes. Experimental results confirm that the new model outperforms the basic one as well as common stochastic methods.