The traditional PV power generation prediction has the problem that the prediction accuracy is not high due to incomplete and inaccurate feature extraction of meteorological factors. In order to fully explore the influence of meteorological factors on PV output and effectively utilize the advantage of deep learning technology in non-linear fitting, this study proposes a short-term forecasting method for PV power that is based on the full mining of meteorological factors and is realized through the BiLSTM network. Based on outliers and standardized processing of the original data, KNN is used to fully explore the key factors affecting PV output among meteorological factors such as external temperature, humidity, and pressure. And then multivariate data sequences are reconstructed. On the basis of exploring the reasonable setting scheme of the hyper parameters such as the time steps of the input layer, the number of model layers, and the dimensions of each layer, a BiLSTM network model is built to realize the high-precision prediction of short-term power of PV power generation. Simulation results show that the proposed KNN-BiLSTM method has higher prediction accuracy than the classical methods such as KNN, DBN, BiLSTM, and PCA-LSTM.