Abstract: False Data Injection Attack (FDIA) threatens the security of power systems by tampering with grid measurement information. For the new power system FDIA, the principle of the attack is studied and a detection method based on CNN-BiGRU-XGBoost is proposed. The method uses Convolutional Neural Networks (CNN) and Bidirectional Gated Recurrent Unit (BiGRU) to extract spatio-temporal features, utilizes eXtreme Gradient Boosting (XGBoost) for classification, and introduces Multi-Head Attention and Optuna methods to optimize the model performance. Simulation experiments are conducted in IEEE-14 node and 39 node systems, and the results show that the method in this paper possesses better accuracy and balance than common methods, which verifies the effectiveness and robustness of the method.