Abstract: False data injection attacks pose a significant threat to the security of cyber-physical power system. Traditional attack detection methods fail to accurately identify and localize attacked nodes due to their lack of consideration for the topological relationships among measurement data and their inadequate feature extraction capabilities. Therefore, this paper introduces a novel detection method for locating false data injection attacks, based on a graph attention network and a multi-scale parallel fusion convolutional model. This method dynamically captures the topological relationships among measurement data through the graph attention network, enhancing the localization performance of the detection technique. It utilizes a parallel convolutional neural network, augmented with an attention feature fusion module, to extract multi-scale features, thereby improving the learning and generalization capabilities of the detection method to achieve high-precision localization. Evaluation studies conducted on the IEEE-14 and IEEE-57 bus test systems demonstrate that the proposed method precedes existing localization techniques, achieving F1 scores of 98.40% and 95.29%, respectively. Consequently, this method provides a more effective solution for the localization detection of false data injection attack.