Photovoltaic power generation is one of the emerging clean energy power generation methods. However, its efficiency is severely influenced by light intensity in the external environment, resulting in unstable electricity input to the power grid. Therefore, it is very important to predict the trend of change in power generation through collecting and analyzing external environmental factors. Currently, most of the existing methods use a single model to construct the prediction structure, which leads to unstable prediction results when faced with different environmental data. To address this problem, we propose an optical power prediction method based on double deep neural networks. It employs BPNN (back propagation neural networks) and LSTM (long short term memory) as the basic discriminators and combines them into a more accurate and robust optical power prediction model through the genetic algorithm. Experiments on the real datasets of northeast power grid show that compared with existing single neural network models, the proposed method has higher discrimination accuracy and more stable prediction results.