To eliminate limitation of the ID3 algorithm, an optimized algorithm for two-level information gain of attribute- value pairs is presented to establish the forecasting model of daily characteristic load decision tree. The algorithm has improved primitive ID3 algorithm in many aspects. It can prevent expansion biasing the attribute that has multi values. The relationship of attributes can be considered well by this improved algorithm. Through setting threshold value sensitization of noise can be reduced. Daily characteristic load forecasting can be implemented by the model which associate day-forecasted information such as weather, week and so on. The analytic method of histogram is adopted to disperse the data of the load rate-of-change and the data of weather combined hierarchical clustering and discretization based on entropy; after the data is pre-processed, the forecasting model of load decision tree is established by the optimized algorithm for two-level information gain of attribute-value pairs and the characteristic load can be forecasted by giving the information of date-forecasted weather and week. The forecasting results meet the requirements of utility and demonstrate high-accuracy of the proposed model. If the 24 points or 96 points load and its corresponding influent factors would be used to train model, 24 points or 96 points forecasting models will be formed. Then load of 24 points or 96 points can be forecasted. (6 pages)