To improve the accuracy of mid-long term load forecasting, a variable weight combination forecasting method based on golden section algorithm is proposed, which avoids the shortcoming of single grey model and exponential smoothing method. Taking the minimum sum of absolute relative error between fitted value and true value as objective function, the golden section algorithm is used to select the optimal smoothing factor and the weight of grey model and self-adaptive cubic exponential smoothing method. With the power load data updated by metabolism, the new smoothing factors and weights of each single method are selected by golden section algorithm repeatedly. Then establish the new variable weight combination forecasting model. Results of simulation verify the feasibility of the proposed variable weight combination forecasting method. Compared with a single grey model, cubic exponential smoothing method or equal weight combination method, the accuracy of the mid-long term load forecasting is effectively improved.