The sparse distributed peak samples in time series are poorly fitted in the least squares support vector regression
model. Therefore, a new least squares support vector regression model is proposed based on weighted least squares method
and used to predict peak samples in time series. In this model, the structural risk objective is optimized by the distribution
density and the amplitude of expected output. The fitting errors of model are not influenced by the distribution of samples,
and the fitting and prediction accuracy of the peak samples is improved significantly with the holistic accuracy maintained
simultaneously. The simulation results on the Lorenz time series prediction and load prediction in power system show the
effectiveness of the model.