Mishchuk O. Neural-like methods and tools to forecast the parameters of atmospheric air pollution

In the dissertation was solved the actual scientific task of development of methods and software using neuro-like structures with non-iterative learning to improve the accuracy of forecasting air pollution parameters, which are focused on mobile and embedded devices. A method was developed for introducing additional attributes (cluster markers into the input vectors) which provides an increasing of accuracy of filling the missing parameters of atmospheric air pollution parameters. A method of short-term forecasting of air pollution parameters was developed, which increased the horizon of time sequence prediction by using a committee of linear and nonlinear neuro-like structures to partially correct separately positive and negative deviations from exact values. The method of functional expansion of Yoh-Han Pao inputs was improved by using rational fractions, which provided an increase in the accuracy of filling the missing values of atmospheric air pollution parameters by reducing deviations at the extrapolation points. The method of constructing a matrix of coefficients of linear polynomials, created by their identification based on the results of learning the linear neuro-like structure of the sequential geometric transformations model, which reduced memory costs when forecasting air pollution parameters was further developed. A software tool with a set of libraries for the implementation of methods for predicting the air pollution parameters on mobile and embedded devices, in particular in the conditions of gaps in atmospheric air monitoring data, has been developed.