Study On Data Selection Method Of Historical Operation Data For Large Scale Power System

A data selection method based on similarity measurement and support vector machine (SVM) is proposed. At first, the critical clearing time (CCT) is used as the class label, and features which are strongly correlated with the class label will be extracted. Secondly, a SVM classifier is trained on the initial training instances with extracted features, and the instance which is misclassified will be removed. Thirdly, the concept of the most similar instance pair is proposed, which two instances with the minimum distance are selected, and then removes the eligible instances which is noisy and redundant instances. The proposed method which can simultaneously prune data in horizontal and vertical directions is tested by online historical data of an actual large scale power system. Experimental results demonstrate that more than 70% features and 30% instances are reduced, and the accuracy and storage reduction are also improved. This method can be well used with the good performance in large scale power system.