According to the similarity of power load, this paper proposes an integrated load forecasting method based on estimating signal parameter via rotational invariance techniques (ESPRIT). First, the raw data signal is broken up into blocks through a spinning method, and then, it is separated into independent harmonic ingredients by using the least squares ESPRIT algorithm. In addition, before forecasting the power load with different models to get the final integrated forecasting load, we should cluster the ingredients for several categories by K-means clustering. ESPRIT algorithm which has high frequency resolution, is not requested to the synchronized sampling, and it can reduce the dimension data matrix. A better forecast is got by comprehensive forecasting method. Finally, MATLAB simulations indicate that the method is proved to be more stable, accurate and effective.