An attack detection model for supervisory control and data acquisition （SCADA） system of new energy plant based on improved online sequence extreme learning machine （OSELM） is proposed. First of all， ADASYN algorithm is used to balance the number of abnormal data and normal data in the data samples， so as to meet the characteristics of less abnormal data in the real SCADA system environment. Then， the balanced data is reduced by using the de-noising autoencoding network （DAE）， and the irrelevant and redundant features are eliminated to reduce the training time of the detection model. Finally， a large number of comparative experiments are carried out on the AWID data set. The results show that the proposed data reduction method can effectively reduce the data dimension and detection time. Compared with other detection classifiers based on shallow learning algorithm， the proposed method also shows better performance in detection accuracy and false alarm rate.