Abstract: A BiGRU ultra-short-term photovoltaic power forecasting method based on SOM clustering and secondary decomposition was proposed in this paper. To reduce the influence of different weather conditions on the characteristics of photovoltaic power output,SOM clustering was used to classify the input data. Then,a secondary decomposition method combining singular spectrum analysis and variational modal decomposition was adopted to decompose the original signal aiming to reduce the volatility of the original signal and the complexity of photovoltaic data feature mapping. Finally,the BiGRU network was built by time series modeling with the decomposed signal as input. The training strategy combined the signal characteristics at different times significantly improves the accuracy of the power prediction. Compared with several other classical methods,the proposed method can effectively improve the forecasting performance of photovoltaic power.