THE USE OF RECURRENT NEURAL NETWORKS TO ANALYZE TIME SERIES AND IDENTIFY TRENDS IN DATA

Author(s): Chub V.S.

Rubric: Information technology

DOI: 10.21777/2500-2112-2024-2-91-102

Release: 2024-2 (47)

Pages: 91-102

Keywords: recurrent neural networks, temporal series, transformers, long-term dependencies, encoder

Annotation: Traditional models of the recurrent neural network (RNN), such as LSTM (Long Short-Term Memory) and GRU (Gated Recurrent Unit), have long been used for time series analysis, but have limitations in learning on long sequences and high computational requirements. Recent advances in time series forecasting have shown a de- crease in their importance. In this article, a new RNN architecture with linear temporal complexity and lower memory consumption is proposed. During the study, the proposed RNN model was tested and its performance was evaluated for various temporal series analysis tasks. The presented results of the experiment show that the proposed model is superior to modern alternatives. The competitiveness of the model was confirmed by comparison with advanced models such as PatchTST and TimesNet. In addition, the proposed model surpassed the perfor- mance of models based on multilayer perceptrons (MLP) and proved to be more effective than models based on transformers. The proposed RNN architecture may become a promising direction for future research in this field.

Bibliography: Chub V.S. THE USE OF RECURRENT NEURAL NETWORKS TO ANALYZE TIME SERIES AND IDENTIFY TRENDS IN DATA // Education Resources and Technologies. – 2024. – № 2 (47). – С. 91-102. doi: 10.21777/2500-2112-2024-2-91-102