Open AccessNEURAL NETWORK ALGORITHM FOR CHOOSING METHODS OF TIME SERIES FORECASTING
Author(s) -
Yuri Vladimirovich Dubenko,
E. E. Dyshkant
Publication year - 2019
Publication title -
vestnik astrahanskogo gosudarstvennogo tehničeskogo universiteta. seriâ: upravlenie, vyčislitelʹnaâ tehnika i informatika
Language(s) - English
Resource type - Journals
eISSN - 2224-9761
pISSN - 2072-9502
DOI - 10.24143/2072-9502-2019-1-51-60
Subject(s) - artificial neural network , computer science , reliability (semiconductor) , block (permutation group theory) , set (abstract data type) , field (mathematics) , algorithm , time series , data mining , series (stratigraphy) , task (project management) , artificial intelligence , machine learning , engineering , mathematics , systems engineering , quantum mechanics , pure mathematics , programming language , paleontology , biology , power (physics) , physics , geometry
The prediction unit is one of the most important components of intelligent control systems. The results of its operation influence the type of control actions generated by the system. The performance of the unit depends on the prediction methods used. The accuracy of the result of the prediction block depends on the choice of the forecasting method. Thus, uncertainty when choosing a forecasting method is a factor that has a negative impact on the reliability of the result of the prediction block and, as consequence, on the reliability of the control system as a whole. The analysis results of the work in this field suggest that the problem of choosing the best forecasting methods has been worked out mainly at the conceptual level. The disadvantages of the considered works are the lack of specifying the mechanism for implementing the proposed algorithms, as well as the potential result of their work is a wide group of prediction methods that are found optimal. In one of the considered works, the expert system is indicated as a mechanism for solving the problem, and the algorithm for modifying and updating the rules is not specified. We have proposed the algorithm based on the use of a precedent analysis method realized in artificial neural networks, which allows to solve these problems. The statistical indicators of the time series, as well as the forecast horizon are used as characteristics of the object and the forecasting task, which constitute the set of attributes of a precedent. A set of solutions to the problem are the applied prediction methods. The set of results is a general estimate of the solution calculated on the basis of the values of the optimality criteria. At the same time, estimation of the optimality of the forecasting method is performed on the basis of the criteria of accuracy and speed, which are based on the prediction error, as well as the length of time spent on obtaining the forecast. The effectiveness of the proposed algorithm has been proved by the results of the experiment.