Forecasting Based on Artificial Intelligence Tools and Characteristics of Non-Stationary Objects

Abstract

A comprehensive approach to modeling and predicting stochastic temporal processes in non-stationary systems has been elaborated, utilizing artificial intelligence techniques along with informational properties inherent to such systems. To address the problem of non-stationary object identification, several mechanisms have been introduced: these rely on a nonlinear Volterra polynomial representation, a three-tier neural network architecture, and optimization procedures grounded in statistical, dynamic correlation, and signal-point descriptors of stochastic time series. Furthermore, a framework for enhancing prediction accuracy has been constructed, drawing upon intellectualization methods and the informational attributes of stochastic temporal processes. A scheme for the algorithmic derivation of computational structures intended for variable optimization and adaptation has also been put forward. The degree of informational influence on the identification of stochastic temporal processes has been quantitatively evaluated. A procedure has been formulated for quantifying uncertainty and building an optimization strategy applicable to the description of stochastic temporal processes, encompassing adaptive neural network training and the regulation of information processing speed. Finally, a dedicated software system for non-stationary object identification and forecasting has been implemented, with its constituent functional modules validated against empirical data obtained from a regional power distribution network.