LAUSR

With the rapid development of wind power industry recently, huge data source are accumulated by the widespread supervisory control and data acquisition systems. The data-driven wind turbine power curve plays an important role in many fields, whereas it is sensitive to data quality. The invalid and unnatural data need to be reasonably eliminated. Considering the complex influences to data records, probabilistic description is effective to represent the data uncertainty. Initially, raw data are cleaned in the three-dimensional copula space. On this basis, in divisional operation regions of the variable-pitch wind turbine, the weighted mixture of Archimedes copula functions are estimated by expectation maximization to establish the joint probabilistic distributions. Then, a confidence boundary modeling procedure of power curve is presented to identify abnormal data, while an evaluation system is constructed for adaptive modeling with guaranteed performance. After outliers elimination by the boundary, a bi-directional Markov chain interpolation method is proposed to recover consecutively missing data with optimized weights. Finally, the operation data from different wind turbines are preprocessed for validation. The simulation results show that more accurate power curve can be obtained to calculate the theoretical power, which suggests effectiveness of the proposed methods and their great application potential.