Abstract
Conventional methods for predicting wind turbine bearing temperatures often overlook differences in the scales of local variations and changes in variable contributions and temporal dependencies under varying operating conditions. These limitations hinder the extraction of local patterns and the identification of key features and temporal information, thereby reducing prediction accuracy. To address these limitations, this study proposes THDA-Transformer, a model integrating temporal heterogeneous convolution with dynamic feature–temporal attention. First, the maximum information coefficient (MIC) is used to quantify nonlinear associations between operating variables and bearing temperature. Fourteen relevant variables are selected from supervisory control and data acquisition (SCADA) data to construct multivariate time-series inputs that capture key operating conditions. Temporal heterogeneous convolution (Temporal-HetConv), comprising parallel one-dimensional convolutional branches with different kernel sizes, then extracts and fuses local fluctuations, sustained increases, and turning-point patterns across multiple receptive fields. Next, dynamic feature–temporal attention (DFTA) is incorporated before multi-head self-attention in the Transformer encoder to adaptively recalibrate hidden representations along the feature and temporal dimensions, emphasizing information that is more informative for prediction under the current operating conditions. Finally, multi-head self-attention models temporal dependencies over a broader range, while a feedforward network and a regression layer map the learned features to bearing temperature predictions. The main experiments are conducted using actual operating data from Phase I of a wind farm in northern China. For wind turbine No. 28, the proposed model achieves a mean absolute error (MAE) of 0.29 C, a root mean square error (RMSE) of 0.52 C, a mean absolute percentage error (MAPE) of 0.70%, and a coefficient of determination ( ) of 0.99. Compared with PatchTST, the best-performing benchmark model, it reduces MAE and RMSE by 45.28% and 33.33%, respectively. Ablation experiments demonstrate that Temporal-HetConv and DFTA each improve prediction performance, with their combination yielding the best results. Without retraining or parameter adjustment, the model achieves RMSEs of 0.53 C and 0.62 C and values of 0.99 and 0.98 on turbines No. 49 and No. 35, respectively. These results demonstrate its applicability to another turbine of the same model within the original wind farm and to a different turbine model in another wind farm.