Abstract: In order to improve the accuracy of predicting gearbox oil temperature by using supervisory control and data acquisition (SCADA) with multivariate long time series, and solve the problem of inconsistent data distribution caused by different wind power generation units under different operating environments, a method for gearbox condition monitoring was proposed based on multi-branch time series prediction and transfer learning. Firstly, the original sequence was formed by filtering input parameters with extreme gradient boosting (XGBoost) algorithm, so as to obtain the seasonal and trend sequence through decomposing. Secondly, the feature extraction module of seasonal and trend sequence was proposed to obtain the feature of seasonal and trend sequence, which was fused with the feature sequence processed by Informer model and then input into the multi-layer perceptron to map into the final prediction value, so as to construct the proposed multi-branch time series prediction network (MBFN). Finally, the health index of gearbox was constructed by using transfer learning combined with one-class support vector machine (OCSVM) model and sliding window, so as to complete the condition monitoring of gearbox. Experimental results show that the MBFN from proposed model has significantly improved the prediction accuracy of oil temperature, which is better than the conventional time series prediction model. Meanwhile, the adopted migration strategy can adapt to the distribution of different data with less data, so as to realize the condition monitoring of gearbox, while the proposed model can issue the gearbox fault warning with 18.9 days in advance.