Development of a Time Series Forecasting Method Based on Entropy Measures of Diagnostic Information

The purpose of this work is to develop an approach to forecasting time series of diagnostic indicators of complex technical systems. The following problems were solved: a conceptual model of entropy time series was developed. Also, within the framework of the project, a methodology for forecasting time series was developed taking into account the constructed entropy time series. Entropy time series are constructed from time series from initial data by calculating a measure of entropy. To evaluate the effectiveness of the proposed models and algorithms, software was developed in C#. Experiments were conducted on more than 10 technical time series obtained from the Computational Intelligence in Forecasting (CIF) dataset using this software. Two sets of experiments were conducted: in the first experiment, 10 row points were predicted based on the training set. The effectiveness of the developed approach was just over 60%. In the second experiment, for each subsequent forecast point, all previous points in the series were used. The forecast accuracy exceeded 80%. As a comparison of the developed approach with experiments on CIF data showed, the result was good, but there are algorithms that showed better results - for example, a hybrid method of an exponential model and a fuzzy model. To improve the efficiency of algorithms in new works, we plan to vary approaches to granulating the original time series data. The article is intended for specialists involved in diagnosing technical systems.