Accurate fault prediction for smart meters is a key part of timely processing of abnormal meters and automatic verification of removing meters. To solve the highly imbalanced multi-classification task of smart meter fault prediction, a cost-sensitive ensemble tree model is proposed. In the stage of data preprocessing, hierarchical clustering and other techniques are used for reducing the dimension of features and alleviate the effect of the overfitting. Through optimizing the cost-sensitive objective function based on class prior, the proposed model can effectively overcome the bias caused by the imbalanced data. Experimental results on real-world dataset indicate that the proposed model performs well in fault prediction task.
