Abstract: In the intrusion detection of power grid industrial control system, common artificial intelligence-based algorithms require complete data, and the lack of data significantly affects the accuracy of intrusion detection. Aimed at the above problems, an intrusion detection model that improves and combines generative adversarial network (GAN) and long short-term memory network (LSTM) is proposed. The upstream data missing problem was addressed by an improved GAN. The generator combined multi-head self-attention mechanism and gated recurrent unit neural network for imputation (GRUI) to handle potential correlations among incomplete temporal features and generate imputed data. In addition, a temporal prompt matrix was introduced to assist the discriminator to identify fake data. The improved loss function was used to further improve the quality of data generation. The detection rate of attack samples was improved through the improved porous long short-term memory network (PLSTM) combined with the attention mechanism. The experimental results show that the proposed method effectively solves the problem of inaccurate classification caused by missing data in the intrusion detection of the power grid industrial control system and data offset after repair. After repairing the data with a missing rate of 50%, the accuracy can still reach 94.4%, the false positive rate is as low as 5.9%.