This paper presents a new time-varying power system short-term load model based on the α stable distribution that can satisfactorily simulate the features of the high capacity impact load. The particle filter is applied to estimate the time-varying parameters of the short-term load model， with the latest parameters obtained by estimation used for load forecasting. The particle filter has overcome the drawbacks of the conventional filters which can only work for the constant linear Gaussian system while incapable of working on the non-linear， non-Gaussian and time-varying system. The relative error analysis of the short-term load forecasting results on an actual power network validates the effectiveness of the proposed method.