A method for computer-aided learning a computerized recurrent neural network for modeling of a dynamic system

摘要

The invention relates to a method for computer-aided learning a computerized recurrent neural network for modeling of a dynamic system, the respective points in time by means of a observable vector comprising one or more observable is characterized as entries. According to the invention, both a causal network with a temporally forwardly directed flow of information as well as a retroreflecting - causal network with a temporally rearwardly directed flow of information learned. The states of the dynamic system, in the causal network by first state vectors and in the retroreflecting causal network characterized by second state vectors, each of which is observable of the dynamic system as well as hidden states of the dynamic system contain. The observable of the first state vectors are corrected by a first difference vector which, during the learning of the causal network the difference between the observable of the first state vector and the observable of a known observable vector out of training data describes. The method according to the invention is characterized in that the retroreflecting - causal network a separate second difference vector, to which the observable of the second state vectors be corrected, and which in the case of the learning of the retroreflecting - causal network the difference between the observable of the second state vector and a known observable vector of training data describes. The method is dynamically stable and is particularly suitable for the modeling of the temporal development of energy prices and / or raw material costs. Likewise, the method for modeling of observable any technical systems are used, such as, for example, gas turbines and / or wind power plants.