文档介绍：IEEE TRANSACTIONS ON WORKS, VOL. 9, NO. 5, SEPTEMBER 1998 1471
—Multiscale work Architecture
for Time Series Prediction
Amir B. Geva, Member, IEEE
Abstract— The effectiveness of a multiscale work give good insight into the system, and knowledge about the
(NN) architecture for the time series prediction of nonlinear laws underlying the data. With the knowledge gained good
dynamic systems has been investigated. The prediction task is predictions of the system’s future behavior can be made. The
simplified by posing different scales of past windows into
different scales of wavelets (local frequencies), and predicting techniques for time series analysis and predicting [1] can
the coefficients of each scale of wavelets by means of a sepa- be classified into roughly two general categories: 1) If there
rate multilayer perceptron NN. The short-term history (short are known underlying deterministic equations describing the
past windows) is posed into the lower scales of wavelet series, in principle they can be solved to make a forecast. 2)
coefficients (high frequencies) which are utilized for “detailed”
analysis and prediction, while the long-term history (long past If the equations are not known, one must find rules governing
window) is posed into higher scales of wavelet coefficients the data and information about the underlying model of the
(low frequencies) that are used for the analysis and prediction of time series (such as whether it is linear, quadratic, periodic,
slow trends in the time series. These coordinated scales of time chaotic, etc.).
and frequency provides an interpretation of the series structures, Linear models (such as MA, AR, and ARMA) have been
and more information about the history of the series, using
fewer coefficients than other methods. The prediction’s results most frequently used for time series analysis, though often
concerning all the different scales of time and frequencies are there is no inherent reason to restrict consider