| Popis: |
SCADA data analysis has attracted considerable research interest for monitoring wind turbine condition without additional equipment. Above all, normal behaviour models using artificial neural networks have shown promising results. However, the crucial question of how much training data is actually required to train robust and reliable models has not been addressed in the literature so far. In fact, contradictory statements ranging from a few months up to more than a whole year of training data can be found. This paper therefore empirically investigates the relationship between available training data and model performance. A small feed-forward network as well as a larger recurrent network architecture are trained with variable training length and evaluated on a healthy as well as on a turbine with gearbox-problems. It is shown that longer training periods minimize the risk of poor model performance and larger model architectures can be beneficial. Based on these findings at least one full year of data is recommended for model training. |
