An Iterative Cleaning Method for Abnormal Wind Power Data in Wind Farms Based on Wasserstein Distance

A wind turbine's power curve is an important indicator for evaluating the power generation performance of wind turbines, which is of great significance for the operation of wind farms and the scheduling of power systems. However, the shutdown of wind turbines, sensor failures, and power curtailment can cause a large number of outliers, which poses great challenges to wind turbine status monitoring and power prediction. Aiming at the characteristics of abnormal wind turbine data, this paper proposes an iterative cleaning method for wind farms based on wasserstein distance. Via this proposed method, the speed and power are modeled while gradually removing outliers. A neural network combined with wasserstein distance and monotonic constraints is leveraged to create a curve model and to synchronously clean up abnormal data. The curve fitted by the neural network converges to the true wind turbine power curve, which ultimately enables curve modeling while removing outliers. Finally, various experiments are conducted on numerical simulation datasets and twelve real wind turbine datasets. Qualitative and quantitative results demonstrate that the algorithm proposed can establish an accurate power curve model in the presence of a large amount of abnormal data, and significantly outperforms other baselines based on some discussed criteria.