Multivariate Chaotic Time Series Prediction: Broad Learning System Based on Sparse PCA

The sparse principle component analysis (SPCA) comprehensively considers the maximal variance of principal components and the sparseness of the load factor, thus making up for the defects of the traditional PCA. In this paper, we are committed to propose a novel approach based on broad learning system with sparse PCA named as SPCA-BLS for chaotic time series prediction. We also develop the incremental learning algorithms to rapidly rebuild the network without full retraining if the network is considered to be expanded. The core of the SPCA-BLS is that we achieved the dimensionality reduction and the features extraction of high-dimensional data and the dynamical reconstruction of the network without the entire retraining. The method has been simulated on an artificial and an actual data sets and the experimental results in regression accuracy confirm the characteristics and effectiveness of the SPCA-BLS.