Prediction of Wind Turbine Blades Icing Based on CJBM With Imbalanced Data

Abstract

Supervisory control and data acquisition (SCADA) is widely used in wind farms as an effective data acquisition system for wind turbines (WTs). However, in practical engineering applications, it is difficult for us to have adequate conditions to collect enough WT blade icing data, which leads to data imbalance and uneven distribution in the feature space. Using the classical synthetic minority oversampling technique (SMOTE) to balance the data may increase the overlap of positive and negative samples, or produce some redundant samples without useful information. A center jumping boosting machine (CJBM) method is proposed that combines an improved clustering-based oversampling (γ mini density peaks clustering SMOTE, γMiniDPC-SMOTE) and light gradient boosting machine (LightGBM) for blade icing prediction. First, to solve the problem of imbalanced and uneven distribution of WT data, a <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">${\gamma }$ </tex-math></inline-formula> MiniDPC-SMOTE method is proposed, which divides icing samples into multiple clusters, then increases icing samples, and alleviates uneven distribution in feature space. Second, calculating the intercept distance <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">${d}_{c}$ </tex-math></inline-formula> based on the binary search method and the adaptive selection of DPC parameters based on the step phenomenon of <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\gamma $ </tex-math></inline-formula> parameters and verified by <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\gamma $ </tex-math></inline-formula> -step of two WT icing data are proposed. Then, for the problem of low operating efficiency of the model under a large amount of imbalanced data, LightGBM is used for model training and icing prediction. Finally, validation was performed on two SCADA datasets. The results showed that the accuracy, precision, recall, F1-measure, and running times increased significantly, proving the superiority of the CJBM.

Topics & Concepts