The deterministic load prediction is difficult to meet the variability decision in power demand. A short-term load interval forecasting based on Gaussian kernel density estimation with optimal window width is presented. The deterministic load prediction is carried out by least squares support vector machine, then, on the premise of characteristic statistics of historical load relative error, the kernel density estimation method is used to select the Gauss kernel function and the optimal window width is used to establish the density function for the relative error in each region. Taking the load data in one area of Zhejiang as an example, the results of load intervals forecasting are given under different confidence levels. The proposed forecasting with optimal window width is compared with the load forecasting with fixed window width, the interval coverage of the proposed intervals forecasting is obviously improved and the interval width is reduced under the same confidence level.