Aiming at the problem of frequent failures for the wind turbines working in harsh environment for a long time, an anomaly warning method with convolutional neural network (CNN) and gate recurrent unit (GRU) with attention mechanism is proposed. Firstly, the clustering fast search and find of density peaks (CFSFDP) algorithm and the local outlier factor (LOF) algorithm are jointly employed to preprocess the abnormal data in the wind turbine condition supervisory control and data acquisition (SCADA) system. Then the input and output parameters of the model are selected by evaluating the feature importances with the extreme gradient boosting (XGBoost) algorithm. The selected parameters are then input into the CNN-GRU-Attention model for training and prediction to establish the prediction model of the wind turbine under normal operation conditions. Based on this model, the state evaluation index of wind turbine is established by using the time-lapse sliding window, and the upper warning threshold is determined by the interval estimation method in statistics to realize the abnormal condition early warning. Finally, fault warning experiments are carried out with the real historical fault data of a wind power unit, which show that the CNN-GRU-Attention model can effectively identify and warn the abnormal working conditions in advance to ensure the stable operation of the wind turbine.