Design of a prediction system based on the dynamical feed-forward neural network

Analysis and prediction of time series play a significant role in scientific fields of meteorology, epidemiology, and economy. Efficient and accurate prediction of signals can give an early detection of abnormal variations, provide guidance on preparing a timely response and avoid presumably adverse impacts. In this paper, a prediction system is designed based on the dynamical feed-forward neural network. The trajectory information in the reconstructed phase space, which is topologically equivalent to the dynamical evolution of the system, is applied to establish the prediction model. Moreover, an integer constrained particle swarm optimization algorithm is employed to select the optimal time delay, which is the parameter of our system. Simulation results for applications on the Lorenz system, stock market index, and influenza data indicate that our proposed method can produce efficient and reliable predictions.

Automated summary

Analysis and prediction of time series are crucial in scientific fields such as meteorology, epidemiology, and economy. This paper proposes a prediction system based on a dynamic feed-forward neural network, utilizing trajectory information in the reconstructed phase space to establish the prediction model. An integer constrained particle swarm optimization algorithm is employed for selecting the optimal time delay parameter. Simulation results on various datasets validate the efficiency and reliability of the proposed method.