The detection and identification of abnormal transients in nuclear power plants is essential for implementing correct emergency measures and ensuring the safe operation of nuclear power plants. The accuracy of current abnormal transient detection methods in nuclear power plants is susceptible to the influence of the spatiotemporal correlation of operational data, and the subjective setting of abnormal detection thresholds also impacts the detection precision. Therefore, this paper proposes an online detection method for abnormal transient conditions in nuclear power plants based on spatiotemporal graph attention networks. Firstly, taking normal operating condition data as the learning target, a multi-head Graph Attention Network (GAT) is utilized to extract the spatial features of multidimensional time-series data. Secondly, multi-scale temporal convolutions and gating mechanisms are applied to capture temporal features from operational data. After feature fusion, an anomaly detection model is constructed by integrating a prediction model and a reconstruction model. Finally, the dynamic detection threshold based on Peaks-over-threshold (POT) is used to identify the abnormal transient conditions by combining the prediction error and the reconstruction error. In the application of the primary circuit system of nuclear power plant, this method shows higher accuracy than the classical baseline model.