Abstract: The operation of condenser plays a significant role in energy saving for the coal-fired power unit. A data-mining based methodology was proposed in this paper combing with deep learning methods. Firstly, Gaussian mixture model was employed to determine the benchmark of vacuum by mining the historical operational data with respect to the varying working conditions. Secondly, a predicted model of vacuum was established by long short-term memory networks and maximum index coefficient. The vacuum's anomaly can be detected by comparing the predicted value and the benchmark. Finally, the methodology was validated by an on-duty coal-fired power unit with 1000 MW capacity. Results show that the proposed method can detect vacuum's anomaly effectively and provide opportunities for energy saving enhancement.