With the improvement of intelligence level of demand-side user terminals, short-term load data has the characteristics of non-stationarity. Single load forecasting model and conventional combined forecasting model ignore the timing characteristics of load data, making it difficult to achieve satisfactory prediction accuracy. In view of this situation, this paper proposes a short-term load prediction method based on HHT and improved shapley value model. Through HHT transformation, the non-stationary load is reconstructed to obtain random, periodic and trend components. Through the improvement of shapley value model, the weight distribution of each prediction method of composite prediction is determined, and it is applied to the prediction of random, periodic and trend components respectively, and the final predicted value is obtained by superposition of each predicted component. The short-term prediction of non-stationary load is carried out by using three schemes: single prediction model, unimproved Shapley combination prediction model and improved Shapley combination prediction model. The results show that the prediction method proposed in this paper has higher accuracy and stability.