Linear correlation-based sparseness method for time series prediction with LS-SVR

Abstract

Fault or health trend prediction using time series is an effective way to protect the safe operation of highly reliable systems. Least squares support vector regression (LS-SVR) has been widely applied in time series prediction. However there is one of the main drawbacks of LS-SVR, which is lack of sparseness. This drawback impacts on its application if the number of training samples is large. So a new pruning method based on linear correlation is proposed, which reduces the number of support vectors by judging the linearly correlation among the sample data after they are mapped into high dimension feature space. This method can efficiently control the loss of useful information of sample data, improve the generalization capability of prediction model and reduce the prediction time simultaneously. And it also avoids the difficulty of reasonable selection of parameters. Simulation experiment results show that the computing time and prediction accuracy are both satisfied with the approach, which proves the efficiency of the proposed method.