Ultra-short-term photovoltaic power generation prediction is beneficial to power grid dispatching management, improve power system operation efficiency and economy. In view of the defect of traditional long and short time memory (LSTM) neural network that tends to ignore important timing information when processing long sequence input, this paper proposes a power prediction model combining Attention mechanism and LSTM network. In this paper, the Pearson correlation coefficient method is firstly used to analyze the historical data set of the experiment. Irrelevant variables are eliminated and dimensionality reduction is carried out on the data set to simplify the structure of the prediction model. On this basis, Attention mechanism and LSTM network are combined as prediction models. The Attention mechanism assigns different weights to the input characteristics of LSTM, which makes the prediction model more effective in processing input of long time series. The model proposed in this paper is trained and compared with the measured data of a photovoltaic power station. The prediction model proposed in this paper can make full use of historical data, be more sensitive to the key information in the input sequence of a long time, and have higher prediction accuracy.