Abstract: In order to improve the prediction accuracy of PV power, a hybrid PV power prediction model based on variational mode decomposition, fuzzy entropy, convolution neural network and bidirectional long short-term memory network: VMD-FE-CNN-BiLSTM is proposed in this paper. In view of the randomness and strong fluctuation of photovoltaic power generation, VMD is used to decompose the original photovoltaic sequence data into multiple sub-sequences, so as to reduce the influence of random fluctuation components and noise interference on the prediction model. Fuzzy entropy（FE） is used to reorganize each sub-sequence, and the features and trends of different components are extracted by using local connection and weight sharing of one-dimensional CNN, and the features output by CNN are fused and input into BiLSTM model; BiLSTM model is used to establish the time characteristic relationship between historical data, and the prediction results of photovoltaic power generation are obtained. Simulation and experimental results show that compared with BiLSTM, CNN-BiLSTM, EEMD-CNN-BiLSTM and VMD-CNN-BiLSTM, the proposed VMD-FE-CNN-BiLSTM model has higher accuracy and stability in PV power prediction, and meets the requirements of short-term PV power prediction.