Abstract: In the operation of wind farms, accurate and timely fault detection is the key to reducing the operation and maintenance costs of wind turbines. However, the existing detection methods have not fully explored the potential spatio-temporal correlations between functional units, limiting the improvement in detection accuracy. An online anomaly detection method of wind turbines is proposed based on the hierarchical spatio-temporal graph neural network to improve the accuracy of fault detection. Based on the physical structure of wind turbines, functional units are divided into multiple subgraphs to construct a hierarchical spatio-temporal graph neural network, which can fully analyze the correlation strength between various sensor nodes and functional units in wind turbines through graph attention and multi-head attention mechanisms. Additionally, with respect to the temporal correlation of data for the supervisory control and data acquisition（SCADA） system, a dynamic graph neural network and time attention mechanism are designed, which make the normal behavior prediction model capture the temporal correlation, achieving effective integration of spatio and temporal characteristics. Finally, actual data from a wind farm in Shanghai, China confirm that the proposed method is highly effective.