Abstract: The partial least square（PLS） method is firstly used to decide which variables have great impacts on the output power.Then the auto-associative kernel regression（AAKR） method is used to construct the multivariable power curve model.Compared with the traditional power curve model with only wind speed as input,the AAKR method has a much higher modeling accuracy.Moving window method is introduced to detect the abnormal changes of the predicting power residuals to give accurately and timely alarms for wind turbine.With a test 1.5 MW wind turbine SC ADA data,the method proposed in this paper successfully detects the abnormal condition of the wind turbine caused by the pitch system failure and is proved to be effective.