Large Scale Chaotic Time Series Prediction Based On Feature Extraction Methods

Chaotic time series are common in all aspects of social life.With the development of science and technology and the popularization of sensor application,the observation scale and dimensions of chaotic time series in various fields have increased significantly.The increase of data scale not only increases the difficulty of data feature extraction,but also makes prediction more difficult.In order to effectively obtain the feature information of large-scale time series and achieve accurate prediction,we should extract the feature information from large-scale chaotic time series and build an efficient prediction model.In this regard for large-scale chaotic time series,we analyses the data from three aspects of evolution information extraction,time-series feature capture and information extreme exploitation so as to effectively extract feature information and improve the prediction accuracy.Aiming at the complex dynamic information in chaotic time series,phase space reconstruction is performed.Although the uniform phase space reconstruction can obtain the nonlinear evolution information in the chaotic time series,the reconstructed data has a larger scale and it is difficult to identify the key features.In a bid to solve the problem of difficult parameter selection in the process of phase space reconstruction,we introduce sparse principal component analysis for automatic key variable selection and feature extraction,and apply a broad learning system adapted to large-scale data for time series prediction.For the unique dynamic time series information,an effective time series feature extraction model is constructed.Because of the powerful unsupervised feature learning ability,restricted Boltzmann machine has shown excellent performance in time series feature extraction.Meanwhile,it is more adaptive to input data than traditional time series prediction models.However,for the application of time series data,restricted Boltzmann machine lacks a unique mechanism to capture time series information.It only provides a feature extraction mechanism for dynamic modeling and cannot complete the task of regression prediction alone.In view of the above problems,we construct a recurrent restricted Boltzmann machine to achieve regression prediction.This model introduces a leaky integral reservoir recurrent structure to capture dynamic information.This recurrent structure can not only make up for dynamic characteristics,but also has short-term historical information memory capability,which is more suitable for time series applications.On this basis,we establish a cross-layer connection to remedy the information lost in the recurrent process and achieve feature reuse.For the sake of comprehensive utilization of linear and nonlinear features,we build a model for maximum information exploitation,improve the recurrent structure to capture linear information while extracting time series information and construct an enhancement layer for non-linearity through non-linear random mapping to supplement the nonlinear information.In order to make full use of the extracted features,we also construct cross-layer connections at each layer of the neural network to make up for the missing information during the hierarchical connection process and achieve feature reuse.