Due to extreme weather or wind turbine (WT) fault, WTs often collects abnormal data, which often interferes with the real-time control strategy of WT. To detect the abnormal data in real time, a detection framework suitable for wind power data is proposed, integrating the semi-supervised learning mechanism into the Robust Random Cut Forest algorithm. To do so, the normal data around the wind power curve are firstly selected and used to establish the structure model of normal data, considering the magnitude orders and distribution of different features. In each sample, the new sample data are inserted into the model, of which the complexity change is compared with a dynamic threshold, so as to judge whether the new sample data are abnormal. To reduce the dependence on the selection of the labeled normal data in modelling, it is presented a real-time model updating strategy based on self-training idea in semi-supervised learning. The experimental results show that the detection accuracy of the proposed method can reach 95% with only 1000 groups of the labeled normal data, and the detection time of a single sample is only 50 ms, which can detect abnormal data in real time for facilitating control strategy and other work.