Chepenko T. Nonstationary time series prediction on the base of artificial neural networks with time delay elements

The thesis is devoted to learning of the artificial neural networks with time delay elements for forecasting of time series that describe the behavior of the multivariable systems. The adaptive predictive models of stochastic processes and their learning methods have received further development. The architecture of artificial neural networks with the feed forward propagation of information and that have robust properties in conditions of disturbances with unknown distribution and the learning method of artificial neurons on the base on robust Welsh's criterion had been improved. New methods of learning predictive recurrent neural networks based on dynamic neurons-filters with finite-impulse and infinite-impulse response are presented.The efficiency of the proposed methods was experimentally confirmed by the instrumentality of simulation modeling. The proposed learning methods and networks structures were used in the practical task of the intrusion alarm system modeling.