The quality of nuclear power plant operation data is very important for equipment performance monitoring. As the main step of data cleaning, the accuracy of data anomaly detection is affected by the spatiotemporal characteristics of multidimensional time series. At the same time, the existing technology has limitations in distinguishing sensor data anomalies from equipment state anomalies. Therefore, this paper proposes an online anomaly detection method for nuclear power plant operation data based on spatiotemporal graph neural network. Firstly, the graph attention network is used to extract the spatial features of multi-sensor data, and the multi-scale feature fusion and attention mechanism are used to extract the temporal features of historical data. Secondly, after feature fusion, the bidirectional gated recurrent unit is used for time series prediction, and the variational autoencoder is used for data reconstruction. Finally, the adaptive threshold selection method is used to determine the anomaly detection threshold, and the data anomaly is identified by combining the prediction error and the reconstruction error. In addition, the abnormal data are classified according to multivariate correlation and tolerant window, and the abnormal data are corrected online with reference to the predicted value. In the application of nuclear power plant primary loop system, this method shows higher efficiency and accuracy than the classical baseline model.