A method for forming a multi-tiered neural network forecasting system with the possibility of reconfiguration

Purpose of research. Improving the accuracy of forecasting by identifying logical connections in unstructured datasets and forming a multi-tiered structure of a specialized neural network computing system. Methods. A parallel algorithm for determining the fragmented structure of the training sample is proposed, which is used to isolate fragments containing training data based on the logical dependencies of the sample. Based on the generated fragmented sample, a method for assembling neural networks has been developed, which is used to form an effective structure of a cascade forecasting system. Results. Forecasting the results of the unofficial team competition of the International Student Sports Festival 2023 was chosen as the main experiment. A fragmented training sample has been formed on the basis of which a cascade of neural network modules has been built. Four cascade variants were tested in experiments, which showed a significant increase in prediction accuracy compared to single-module analogues. To significantly improve the performance of a neural network system with ultra-short-term forecasts, the hardware implementation of cascades based on the decisive field of FPGA is considered. The structure of the complex with the possibility of its reconfiguration is proposed. Conclusion. The use of artificial neural networks in forecasting is promising, but it may face problems of inaccuracy of results due to insufficient computing power and collisions in training samples. One of the proposed solutions to the problem is cascading specialized neural network modules. Positive results were demonstrated by both the software and hardware implementation of the system based on the proposed cascade. The evaluation of the hardware implementation demonstrates the possibility of acceleration, compared with the software implementation, which may be necessary when conducting ultra-short-term forecasts. The proposed methods and algorithms have demonstrated their correctness.