On the Prediction of Chaotic Time Series using Neural Networks

Authors : Josue Alexis MARTİNEZGARCİA, Astrid Maritza GONZALEZZAPATA, Ericka Janet RECHYRAMİREZ, Esteban TLELOCUAUTLE

Pages : 94-103

Doi:10.51537/chaos.1116084

View : 19 | Download : 11

Publication Date : 2022-07-30

Article Type : Research Paper

Abstract :Prediction techniques have the challenge of guaranteeing large horizons for chaotic time series. For instance, this paper shows that the majority of techniques can predict one step ahead with relatively low root-mean-square error insert ignore into journalissuearticles values(RMSE); and Symmetric Mean Absolute Percentage Error insert ignore into journalissuearticles values(SMAPE);. However, some techniques based on neural networks can predict more steps with similar RMSE and SMAPE values. In this manner, this work provides a summary of prediction techniques, including the type of chaotic time series, predicted steps ahead, and the prediction error. Among those techniques, the echo state network insert ignore into journalissuearticles values(ESN);, long short-term memory, artificial neural network and convolutional neural network are compared with similar conditions to predict up to ten steps ahead of Lorenz-chaotic time series. The comparison among these prediction techniques include RMSE and SMAPE values, training and testing times, and required memory in each case. Finally, considering RMSE and SMAPE, with relatively few neurons in the reservoir, the performance comparison shows that an ESN is a good technique to predict five to fifteen steps ahead using thirty neurons and taking the lowest time for the tracking and testing cases.
Keywords : Chaotic time series, Neural network, Echo state network, Long short term memory, RMSE, Prediction technique