Since the time scales of meteorological factors in short-term photovoltaic power prediction are different, time scales are usually ignored in the analysis of the correlation between time scales and photovoltaic power, leading to errors in the prediction models. To improve the prediction accuracy of photovoltaic power, the CEEMD-BiLSTM-RFR prediction model was constructed. Firstly, the photovoltaic power was decomposed by complementary empirical mode decomposition (CEEMD) to get the modalities on different time scales. Secondly, the relationship between each photovoltaic component and meteorological factors was analyzed by Pearson correlation coefficient. Strongly correlated components were predicted by the random forest regression (RFR) prediction model. Weakly correlated components perform prediction through bidirectional long short-term memory neural network (BiLSTM). Finally, the results of each component prediction were combined to obtain the final prediction result. It is verified by using the measured data of a photovoltaic station in Bengbu, Anhui Province, in July. The results show that the proposed prediction model CEEMD-BiLSTM-RFR has better prediction accuracy than the traditional prediction model.