A dynamic threshold method for wind turbine fault detection based on spatial-temporal neural network

Strong coupling of wind turbines (WTs) makes the supervisory control and data acquisition (SCADA) data spatially relevant and time-dependent. Aiming at identifying the fault state of wind turbines accurately by comprehensively using the change information in the spatial and temporal scale of SCADA data, a novel fault detection method of WTs based on spatial-temporal features' fusion of SCADA data by a deep autoencoder (DAE) and a gated recurrent unit (GRU) is proposed in this paper. The spatial feature extraction capability of the encoder layer and the temporal feature extraction capability of GRU are combined with the symmetric framework of DAE to achieve the spatial-temporal feature extraction of the whole model. Mahalanobis distance (MD) is adopted to convert the multidimensional output into a one-dimensional performance index. To avoid the influence on the fault detection results due to the changeable operational state of WTs, a dynamic threshold regression method based on the gray wolf optimization algorithm and support vector regression model is developed to identify fault data instances. Experiments with SCADA data from real wind farms verify the effectiveness of the proposed method.