Abstract:The existing intrusion detection methods for cyber-physical power systems have some problems such as not paying enough attention to data quality, especially in dealing with discrete data in data sets.In order to address the above issues, an intrusion detection method for cyber-physical power system based on entity embedding and convolutional neural network is proposed, which maps the discrete features in the data set to continuous vectors through entity embedding technology to better represent the discrete features and generate high-quality new data.After that, it is combined with the standardized continuous features as a new dataset to train the convolutional neural network to establish the intrusion detection model of the cyber-physical power system.The experimental evaluation results on the KDD Cup 99 dataset show that the proposed scheme improves the attack detection accuracy by 6.2% and 6.04% compared with the one-hot encoding and the traditional sequential encoding, respectively, and further reduces the false positive rate and false negative rate.