Abstract: A photovoltaic power forecasting model based on dual-attention-GRU network and similar sequences modification is proposed. On the basis of the Encoder-Decoder framework,feature attention and temporal attention are introduced,which can effectively solve the problem of GRU network’s distraction from input features and time series. The DA-GRU forecasting results can be further improved by using the future power values of similar power sequences to modify the forecasting results. The example is verified by DKASC data,and the performance of the model under different forecasting steps is compared. The results show that DA-GRU has the best performance under different evaluation indexes compared with other traditional models,and the similar sequences modification method can further improve its forecasting accuracy.