Abstract: Cognitive radio technology is an effective method to achieve dynamic sharing of spectrum resources and improve spectrum utilization. MLP and CNN based deep learning methods are mostly based on the assumption of Euclid data domain, but the channel state information (CSI) in the complex cognitive radio network cannot meet this characteristic, so the existing power control algorithm based on deep learning exists poor scalability, poor generalization ability and other issues. To address these issues, a power control method based on message passing graph neural network (MPGNN) is proposed. This method constructed a cognitive radio channel model graph, and designed a cognitive graph convolutional neural network (CGCNet) based on CSI. The simulation results show that compared with the existing deep learning methods, the algorithm in this paper can achieve higher performance requirements after being trained in an unsupervised manner, and has good scalability and robustness.