|
The focus real application of author's diploma thesis is on the mutual comparison of
properties of different prediction methods, based on
autoregressive modelling and linear, feed forward and recurrent
neural network. All these models have been used in the adaptive
and in the classical prediction approach. The algorithms of SVD
and QRcp have been used to find the subset of the full
autoregressive model. The same full set has been then used for
three different types of neural networks. Many types of networks
architectures with a wide range of modifications of the learning
algorithms have been tested in the next stage. Suggested
algorithms have been then applied for real data representing gas
consumption in the Czech Republic with the focus on the winter
part of the year. Spectral analysis of these signals with the
sampling period of 2 hours and 1 day has been performed at first
for original data and then for pre-processed signals. The main
part of the work has been devoted to the discussion and comparison
of all models mentioned above.
Aleš Pavelka (May 2002)
|
