Abstract：Aiming at the problems of difficult identification of stacked outliers and insufficient extraction of

raw data features in the data-driven ultra-short-term power prediction of wind turbines，a prediction method

combining data cleaning and parallel spatio-temporal neural network was proposed. First，a combined data cleaning method was proposed to clean the wind turbine power data；then，a parallel spatio-temporal neural network was proposed to extract the temporal features of the power and meteorological data of the target wind turbine，and the spatial features of the power data of similar wind turbines for the fusion prediction，respectively. In addition，a prediction interval accuracy indicator was defined to reflect the accuracy of prediction results under different error intervals，avoiding the drawback of traditional error indicators that obscure large local prediction errors. The analysis results indicate that the proposed method can effectively identify anomalous data and improve the ultrashort-term power prediction accuracy of wind turbines.

      Key words：wind turbine power；deep learning；data cleaning；feature fusion；evaluation index