Analysis of Recurrent Neural Networks for Henon Simulated Time-Series Forecasting

Abstract

Forecasting of chaotic time-series has increasingly become a challenging subject. Non-linear models such as recurrent neural networks have been successfully applied in generating short term forecasts, but perform poorly in long term forecasts due to the vanishing gradient problem when the forecasting period increases. This study proposes a robust model that can be applied in long term forecasting of henon chaotic time-series whilst reducing the vanishing gradient problem through enhancing the models ability in learning of long-term dependencies. The proposed hybrid model is tested using henon simulated chaotic time-series data. Empirical analysis is performed using quantitative forecasting metrics and comparative model performance on the generated forecasts. Performance evaluation results confirm that the proposed recurrent model performs long term forecasts on henon chaotic time-series effectively in terms of error metrics compared to existing forecasting models.