In order to improve the forecasting accuracy of wind farm output power, a method based on the principle of similar time period selection and the principle of principal component analysis (PCA) is proposed. The forecasting model is based on the combination of PCA and multi-layer auto encoder extreme learning machine (ML-AE-ELM). Through the correlation analysis, the range of similar periods of time to be tested is determined, and the training and test samples are constructed by combining weather data, unit status and historical power, and the forecasting algorithm is used to complete the training and test of the samples, so as to obtain the forecasting results of output power and verify them. The experimental results show that compared with the common algorithm model, the proposed model has higher forecasting accuracy in different installed capacity and different working conditions of wind farms, and shows good forecasting stability and generalization ability.