The 10th International Conference on Energy, Resources, Environment and Sustainable Development (ERESD 2024)

197 / 2024-03-12 21:25:30

Transformer-enabled health condition prediction of wind turbines fusing multi-source-heterogeneous data

Health condition monitoring, machine learning, intelligent prediction

At present, lots of methods aim to monitor and predict the health condition based on vibrations with physical statistical indicators or SCADA data with a single indicator. Therefore, it is not reliable for them to predict and monitor the health condition of wind turbines under varying speeds and loads, especially across different turbines. To solve the above issue, a transformer-enabled health condition prediction  method of wind turbines fusing multi-source-heterogeneous data is proposed in this paper, which uses the transformer to train a virtue and generalized health indicator and is able to overcome the drawbacks of single indicator extracted from single data. First, we calculate the correlation between wind farm and select representative normal turbines; Secondly, the input features are constructed using massive normal SCADA data and vibrations from the representative turbine. Then, an intelligent prediction model based transformer is design to train a virtue heal indicator. Finally, a large amount of historical normal data of other turbines is used to tune the trained model for transfer health condition prediction across turbines. The proposed method was verified using real-world data from a wind farm cooperated with our group. The results show that the proposed method can accurately identify the operating status of wind turbines and outperform other methods based single-source data or single physical statistical indicator.