The rapid development of smart grids has brought new challenges to grid operation. In order to adapt to the requirements of rapid response of smart grids and to rapidly estimate the future operation trend of power loads, we propose a prediction method of an ultra-short-term power load interval based on the least squares support vector machine (LSSVM) model. This method predicts the interval by estimating the overall noise variance of the sample data on the basis of point prediction. which has a small calculated amount and greatly reduces the prediction time consumption. With regard to model parameter selection, the optimal training sample size and embedding dimensions are first determined using the parameter determination method of Gamma Test noise estimation, and then the optimal hyper-parameters are selected by the grid search method so that the fitting error of the LSSVM model on the training samples approximates the estimated minimum noise. To verify the validity of the proposed method in this paper, we apply the scheduling load data from a certain grid.to simulation experiments. The results show that the proposed method not only reflects the simplicity and high speed of the LSSVM but also ensures the accuracy of the prediction intervals by optimizing the model parameters.