Abstract: According to the theory of power load forecasting，data mining based on historical data of power load data is used in load predicting.For practical operation process，there is an error in data collection，so load forecasting curve contains bigger saw tooth.This paper presents a new outlier data mining approach.It finds the sharp angle points between two straight，which correspond to outliers of power load value，and smoothes the curve at same time outliers are treated.Experiments show that after the new outlier mining approach is applied，load forecast results have improved significantly.