To address the issues of frequent random disturbances during the operation of wind turbine key components and the difficulty in identifying early-stage faults, this study proposes a wind turbine key component anomaly early-warning method based on a bidirectional long short-term memory (BiLSTM) network integrated with single-head attention (SA),whose hyperparameters are optimized by the RIME optimization algorithm (RIME).First, the supervisory control and data acquisition (SCADA) data of wind turbines are preprocessed, and a combined method of interquartile range and data binning is adopted to eliminate abnormal data. Second, the gray relational analysis method is applied, with the main bearing temperature and generator stator temperature set as the reference sequences respectively; relevant operational parameters with high correlation are selected as comparison sequences to construct the multi-dimensional input feature set of the model. Subsequently, the Single-Head Attention mechanism is introduced to capture long-term dependencies in the time series, and the RIME algorithm is used to optimize the hyperparameters of the BiLSTM model, thereby establishing a RIME-BiLSTM-SA multi-input single-output anomaly early-warning model. Furthermore, the exponentially weighted moving-average (EWMA) method is employed to analyze the residuals between the predicted values and actual values of the model, so as to determine the early-warning threshold. Finally, the proposed method is validated using measured SCADA data from a wind farm in East China. The results show that the RIME-BiLSTM-SA early-warning model proposed in this study can generate valid early-warning signals 12 hours prior to the occurrence of main bearing temperature anomalies and 32.5 hours prior to the occurrence of generator stator temperature anomalies.