Photovoltaic power prediction plays an important role in the scheduling and operation of modern power systems. Aiming at the variability and complexity of photovoltaic power generation， a short-term PV power prediction method based on new similar day selection and northern Goshawk optimization （NGO） to optimize bidirectional gated recurrent unit （BiGRU） is proposed. The main meteorological factors are selected with the Spearman correlation coefficient， and the original PV power and maximum meteorological factor are decomposed into a series of sub-signals by variational mode decomposition （VMD）. Then， according to the construction of new evaluation indicators， the data set of similar days is screened out， a group of BiGRU is used to establish a deep learning model with similar day signals as network input， and NGO is used to effectively optimize the hyperparameters of each BiGRU network. Finally， the predicted value of PV power is obtained by synthesizing the predicted results of each sub-signal. Simulation results show that the proposed hybrid deep learning method is superior to other methods in terms of prediction accuracy and computational efficiency.