Traditional power grid industrial control systems are mainly isolated from external networks through tools such as firewalls, but with the application of new technologies such as cloud computing and the Internet of Things, the degree of interconnection between networks has continued to deepen, and the difficulty of security protection has greatly increased. How to effectively detect network intrusion behavior has become very important. Compared with traditional intrusion detection technology, convolutional neural networks have a better ability to extract intrusion features. This study proposes a power grid industrial control system intrusion detection algorithm based on convolutional neural networks. The KDD99 dataset is processed for model training, and a cascade convolution layer is added to optimize the network structure. Under the premise of small parameter scale, the real-time requirements of the model are guaranteed. Compared with the traditional SVM algorithm and the k-means algorithm, the intrusion detection accuracy of the proposed algorithm in this study is improved, the false detection rate is reduced, and the intrusion behavior to the power grid industrial control system can be effectively detected.