METHODS OF PROCESSING BIG DATA AND FORECASTING TIME SERIES USING NEURAL NETWORKS
TEAM International, Oleksandr Klymenko, O.V. Sukhomlynova, V. Kravchenko · Mechanics And Mathematical Method · 2025
The article examines in detail the features of the functioning of neural networks in the context of their application for processing large amounts of data and forecasting time series, paying special attention to the mechanisms of accumulation of intellectual capital. It is established that the key characteristic of such neural networks is their ability to adaptively adjust weight coefficients and interconnections between neurons, which is achieved through the use of self-organization principles. The article analyzes neural network models, in particular Hopfield and Hamming networks, which are focused on preserving reference samples through preliminary initialization, but demonstrate limited flexibility in changing environments. It is proven that networks, the architecture of which is built on the principle of minimizing the Hamming distance, effectively perform the tasks of classifying binary input data and identifying samples with the smallest distance to the given signal. At the same time, it is noted that limitations in computing resources and the lack of feedback with the knowledge base can cause a certain instability in the operation of such networks. The article proposes a method for optimizing the structure of neural networks, which involves replacing individual layers with matrices of weight coefficients, which makes it possible to reduce the computational complexity without losing the accuracy of the obtained results. The effectiveness of neural networks is assessed in the context of analyzing large arrays of time series, in particular in cases where the data are characterized by high dynamics and complex nonlinear dependencies. The article develops recommendations for creating neural networks with improved stability and plasticity indicators, which are able to preserve the formed knowledge even in cases of influence of external signals. An approach is described according to which the information processing process is based on the activation of short-term memory, the functioning of the attention system and the use of long-term connections, which contributes to a more accurate identification of data classes.