Abstract: Large-scale power transformers have complex structure and high equipment cost, and is a key component of power system, whose safety and reliability are closely related to the operation status of power transformers. Therefore, the transformer status evaluation has become a common operation and maintenance business. However, the current status evaluation work relies heavily on guidelines and expert experience, which is of high labor cost and vulnerable to subjective influence; while the existing evaluation models commonly apply standard algorithms and perform poorly in production environment. This paper proposes a new evaluation model to solve the existing problems in data quality, sample distribution, application requirements and model performance in the large-scale power transformer status evaluation. Firstly, the invalid samples are eliminated and a cross weight method is designed to label the raw valid data. Secondly, the processed status data are distinguished according to their integrity, and then feature extraction and high-dimensional mapping are performed, then the dataset is split into multiple complete training datasets. Thirdly, the SMOTE-BORDERLINE algorithm is applied to synthesize positive samples and provide multiple complete balanced training datasets. Finally, multiple SVM component learners modified with cost sensitive requirements are trained in parallel, which are then integrated into an ensemble learner by weighted voting method. The model proposed in this paper effectively utilizes ensemble learning method to improve generalization ability with the impact of the imbalanced datasets and the cost sensitivity. It is verified a good performance in the production environment. Compared with traditional methods, it significantly reduces the false prediction rate and the missing rate of abnormal status samples.