Abstract: A short-term one-step photovoltaic power prediction model based on Encoder Decoder structure and time embedding was proposed to address the safety operation issues of the power grid caused by intermittent and fluctuating photovoltaic power generation. LSTM units were used to extract the characteristics of photovoltaic power generation, a multi-head attention mechanism was introduced to enhance important information attention in the input sequence. The difference embedding and time embedding were added in the Decoder layer, combined with the output of the Encoder layer for one-step prediction. Average interpolation, downsampling and z-score standardization were introduced to process raw data, and different models were compared under different meteorological types. The experimental results show that, in terms of MSE, MAE and RMSE, the proposed model achieves 34.7%, 27.5% and 16.6% lower error values, respectively, compared with other models under cloudy weather conditions, with higher prediction accuracy and enhanced robustness.