Abstract: Photovoltaic(PV) power prediction is the key basis for optimizing energy distribution and stabilizing power grid operation. Nevertheless, the issue of inaccurate prediction arises in the traditional methods from the imprecise data preprocessing and the inadequate prediction algorithm. Therefore, the paper raises a bi-ensembled PV power prediction based on the consensus clustering and the integration prediction, improving the accuracy of the meteorological classification and the power prediction by the heterogeneous integration. Firstly, based on the re-marking and the projection methods, the Kmeans-GMM-AGNES-BIRCH consensus clustering framework is constructed, which combines four heterogeneous clustering algorithms. Then, the typical meteorological model is established by selecting the outlier day according to the sliding time window. Secondly, by averting the overfits with the k-fold cross-validation regulation, the integration prediction model, the GRU-RF-XGBoost-LightGBM, is constructed to comprehensively explore the underlying laws of the PV data by using the Stacking framework. The simulation is conducted using the data from a PV power station in Australia. It proves that the prediction of the bi-ensembled model is superior to the common ones, verifying the validity of the consensus clustering and integration prediction model.