Deep learning method can help to learn the deep features of power load data and improve the accuracy of prediction, but it also brings problems, such as large amounts of super parameters and poor interpretability of the model. To solve these problems, this paper introduces the deep forest model for short-term load forecasting. Based on the multi-Grained Cascade forest model, the multi-granularity window scanning method is improved with adjusted window size and sliding step size, so that the model can extract the periodicity characteristics of power load data in different time scales. In addition, the calculation method of deep forest output layer is improved with changing the output result from discrete class vector to continuous predicted value, improving the accuracy of the model. Finally, the feasibility and effectiveness of the proposed method are verified with the measured data of northeast China power grid. The experimental results show that the improved deep forest algorithm can achieve higher accuracy with higher prediction accuracy, and has faster learning speed than the deep neural network.