Designing an optimal recurrent neural network for time series forecasting
Игорь Александрович Ботыгин, Yuri Volkov, Vladislav S. Sherstnev, Anna I. Sherstneva · 2025
Experiments on designing an optimal recurrent neural network for one-stage and multi-stage forecasting of time series have been carried out. Such parameters of the recurrent neural network as the number of neurons in hidden layers, the number of hidden layers, the number of training epochs, the mini-sample size were manipulated. It is shown that the method of optimization stochastic gradient descent is more optimal, as it allows obtaining more accurate results with the same architecture of the neural network in the case of a large number of neurons in hidden layers. However, the method of adaptive moment estimation optimization with a smaller number of neurons allows to achieve higher prediction accuracy and training speed. A tendency to stop decreasing the root mean square error starting from the hundredth epoch of training was revealed. This fact indicates the beginning of the process of retraining of the recurrent neural network, which leads to the impossibility to form a correct result on unfamiliar data. At the same time, for multistage forecasting, increasing the number of training epochs leads to a slight increase in accuracy, but the forecasting accuracy decreases as the dimensionality of the output vector increases.