Analysis and Prediction of Chaotic Time Series Based on Deep Learning Neural Networks

• Main Authors: HUI-YU SHEN, 沈慧瑜
• Other Authors: Cheng-Hsiung Yang
• Format: Others
• Language: en_US
• Published: 2019
• Online Access: http://ndltd.ncl.edu.tw/handle/r2dann

碩士 === 國立臺灣科技大學 === 自動化及控制研究所 === 107 === Time series prediction is the task of using historical data to predict future values for a given sequence. Recently, this task has attracted the attention of researchers in the field of machine learning, with the increasing availability of a large amount of historical data and the strong predictive technology inferring random dependence between past and future values to improve time-consuming and complex traditional predictions method. Using a Long Short-Term Memory (LSTM), this is a special type of recurrent neural network that has the advantage of being able to learn the long term dependencies between the provided network inputs and outputs. In this thesis, we propose a Differencing Long Short-Term Memory (D-LSTM) architecture as an extension of recurrent neural networks. The differential is the latter value minus the previous value, which can reduce the noise of the original data to make it smooth and improve the prediction accuracy. We design a 3D nonlinear chaotic system and analyze its properties and dynamic behaviors by phase portraits, equilibrium points, Lyapunov exponents, spectral entropy etc. We study prediction result by change the initial value and the coefficient for our chaotic system. We compare D-LSTM with Adaptive Neuro Fuzzy Inference system (ANFIS) and LSTM, using Root Mean Square Error (RMSE) to measure their performance. The result shows that our model is almost better than others.