Random Process Intelligent Learnable Extrapolator

This research presents an accuracy analysis of the prediction of random signals using adaptive and non-adaptive extrapolators. Extrapolation is performed using Chebyshev polynomials orthogonal over a set of equally spaced points. Several algorithms for an automatic selection of an extrapolating polynomial in real time, based on estimates of the prediction error, are considered. A comparative study of these algorithms is carried out in order to identify the most effective one. The study includes an analysis of prediction errors when using fixed-order extrapolators, as well as adaptive extrapolators using different parameter selection rules. In particular, a selection rule based on the minimum of the error modulus at a given time and a rule based on the minimum of the root mean square error for previous cycles are considered. It is shown that the use of a neural network algorithm based on CatBoost provides the smallest prediction error compared to the alternative approaches considered. This demonstrates the potential of using neural networks to adapt extrapolator parameters to changing characteristics of the predicted random process.