Typical data mining methods(ANN and SVM) are applied in short-term load forecasting widely. Howere, These methods have some deficiences including being trapped in local optimization easily and ensure the model hardly and so on. In order to overcome shortcomings, A method of combination of fuzzy clustering and random forest(RF) for load forcasting is proposed in this paper. On the other hand, various feature of the periodical load and the similarity of input samples are considered in the proposed method. Input samples are clustered depending on similarity. Then load forecasting model is established based on random forest algorithm and similar data are selected as training samples. The final results rely on the historical loads in Anhui for hourly load forcasting. And the results show that the proposed method is better than traditional support vector machine and BP neural network.