风电机组功率曲线的集值回归

Precise wind power curves are pivotal for monitoring the status of wind turbines and predicting wind power, which are important parts of utilizing wind energy in power systems. However, the data sets for training wind power curve models have a critical issue. A considerable proportion of the data sets is abnormal due to communication failure and other factors. Using the data sets with abnormal data will significantly deteriorate the fitting performance. This paper resolves the above issue by proposing a unified way to achieve abnormal data detection and curve fitting. Instead of regression with scalar output, set-valued regression of the wind power curve is considered, giving a set of wind power for a given wind speed. Interval neural network is adopted as the model for set-valued regression. A chance-constrained optimization problem is formulated to train an interval neural network. The obtained interval neural network can specify a subset with the normal data area, which can be used to give the threshold for abnormal data detection. Besides, the center points of the interval can be used as the fitted wind power curve. Since the formulated chance-constrained optimization problem is intractable, a sample-based sigmoidal approximation method is proposed to approximately solve it. The convergence and probabilistic feasibility of the approximation are given. Finally, experimental validations have been conducted to compare the proposed method with several existing methods.