Abstract: A short-term photovoltaic power prediction model based on bi-directional long short-term memory（BiLSTM）considering data decomposition and evolutionary predation strategy is proposed. Firstly, for a large number of high-frequency components and complex frequency components of the original PV power, the SCS method（SVD-CEEMD-SVD, SCS）is developed to fuse the complementary ensemble empirical modal decomposition（CEEMD）with the singular value decomposition（SVD）of matrix operations through the data decomposition theory, which can realize the secondary noise reduction of PV power data. In addition, a combined prediction model with evolutionary predation strategy（EPPS）and BiLSTM is established to better exploit the intrinsic features of the proposed model for improving the prediction accuracy. Finally, the effectiveness of the model in filtering out the PV power noise and improving the prediction accuracy is verified by taking an actual PV power plant in a region of Shandong as an example.