Feature selection techniques for modelling tower fatigue loads of a wind turbine with neural networks

Abstract

The rapid development of the wind industry in recent decades and the establishment of this technology as a mature and cost-competitive alternative have stressed the need for sophisticated maintenance and monitoring methods. Structural health monitoring has risen as a diagnosis strategy to detect damage or failures in wind turbine structures with the help of measuring sensors. The amount of data recorded by the structural health monitoring system can potentially be used to obtain knowledge about the condition and remaining lifetime of wind turbines. Machine learning techniques provide the opportunity to extract this information, thereby improving the reliability and cost-effectiveness of the wind industry as well. This paper demonstrates the modelling of damage-equivalent loads of the fore–aft bending moments of a wind turbine tower, highlighting the advantage of using the neighbourhood component analysis. This feature selection technique is compared to common dimension reduction/feature selection techniques such as correlation analysis, stepwise regression, or principal component analysis. For this study, recordings of data were gathered during approximately 11 months, preprocessed, and filtered by different operational modes, namely standstill, partial load, and full load. The results indicate that all feature selection techniques were able to maintain high accuracy when trained with artificial neural networks. The neighbourhood component analysis yields the lowest number of features required while maintaining the interpretability with an absolute mean squared error of around 0.07 % for full load. Finally, the applicability of the resulting model for predicting loads in the wind turbine is tested by reducing the amount of data used for training by 50 %. This analysis shows that the predictive model can be used for continuous monitoring of loads in the tower of the wind turbine.