Icing detection on wind turbine blades using a dual-attention CNN-BiLSTM model optimized by the hawkfish optimization algorithm

In cold offshore regions, blade icing on wind turbines can cause abnormal loads, elevated failure risk, and reduced power generation efficiency. Accurate detection of icing states is therefore crucial for ensuring safe operation. This study presents a detection method for icing states that combines feature engineering, deep learning, and intelligent optimization. First, the supervisory control and data acquisition (SCADA) data collected from wind turbines at an offshore wind farm in China are preprocessed. Second, a feature engineering stage is performed to construct icing-sensitive features, followed by key feature selection using recursive feature elimination with fivefold cross-validation (RFECV). To fully exploit these features, an icing detection model is constructed. The framework integrates CNN with a BiLSTM network, while dual attention mechanisms are introduced to adaptively emphasize critical channel-wise features and temporal dependencies, respectively. Finally, the hawkfish optimization algorithm (HFOA) is used for hyperparameter optimization. The proposed method achieves an accuracy of 96.57%, an F1-score of 96.61%, and a PR-AUC of 94.83%. The cross-turbine validation results further demonstrate that the proposed method maintains stable recognition performance on the target turbine. The proposed method can effectively identify blade icing states, thereby reducing power generation losses and failure risk in cold regions at sea.