Fault diagnosis of wind turbine blade icing based on feature engineering and the PSO-ConvLSTM-transformer.

In winter, wind turbines often face the risk of blade icing, and traditional methods – which mostly rely on manual observation and sensors – are limited by the experience of testers and costs; therefore, they are not widely used in practice. With the development of artificial intelligence, data-driven methods have garnered increased attention. However, it is challenging to mine the effective hidden information within data. In this regard, this paper proposes a wind turbine blade icing detection method based on feature engineering and the PSO-ConvLSTM-Transformer. Firstly, feature engineering is carried out with the help of expert experience to construct the mechanism variables related to icing; secondly, the ConvLSTM-Transformer prediction model is constructed to mine the timing information between supervisory control and data acquisition (SCADA) data, and finally, the model hyper-parameters are optimized by using the particle swarm optimization (PSO) algorithm to improve the diagnostic performance and generalizability of the model. The proposed model is evaluated using two wind turbine blade icing datasets. The proposed PSO-ConvLSTM-Transformer has significant advantages over multiple baseline models. This study can inform wind farm operations in terms of optimizing maintenance strategies, ensuring the safe and efficient operation of wind turbines in cold weather conditions. • Constructing ConvLSTM-Transformers deep learning model for wind turbine icing detection. • Explored the impact of domain knowledge-based icing mechanism features on deep neural networks. • Tested the effect of different optimization algorithms on the performance of deep learning models. [ABSTRACT FROM AUTHOR]