A forecasting model about the daily gas load based on support vector machine theory is developed in this paper. The ways to improve the forecasting accuracy is discussed, including the normalization method, the data grouping method and the effect of different history data period. It is proved that the proper normalization method is to map the input gas load data from the small and narrow range to the big and wide one. The data grouping method is important because it relates to the gas consumer composition. It is better to obtain the period of history load data by experiment investigation than by theory analysis. As it relates to both the number of training samples and the characteristic of the nonlinear regression. For this study, the period of 5 days is better than 7 days, although the latter one is the number of a week. The high performance of the model is proved as the average error is 0.94% for 5 days forecasting in heating period. Moreover, the research about the ways to improve the forecasting accuracy is helpful to solve the similar problems.