Abstract:With the help of the supervisory control and data acquisition (SCADA),the wind turbine gearbox realizes the fault alarm by monitoring whether the oil temperature of the gearbox exceeds the threshold value,the judgment accuracy is not high and the problem is not discovered in time.In view of that,the long short-term memory network model (LSTM) was used to integrate SCADA data to predict the oil temperature state of the gearbox.The LSTM model was trained through the data under the normal operating state of the gearbox,and the residual between the predicted value and the real value was calculated,according to the principle of normal distribution,the upper and lower warning thresholds were set.In order to simplify the training complexity of the model,the parameters closely related to the gearbox oil temperature were selected as the input items of the LSTM model.In order to reduce the poor prediction accuracy caused by improper setting of hyperparameters of the LSTM model,a combined model of the moth flame optimization (MFO) algorithm and LSTM was proposed,with preserving the powerful global search capability of the MFO algorithm,it could avoid the trap of local search,the LSTM model was iteratively optimized by MFO,and the suitable model was finally constructed.Finally,the SCADA data of a wind turbine is verified that the method can effectively warn the fault of the gearbox,and compared with other methods,the accuracy is higher,the warning is more timely,and the iterative effect is better.