As an important index reflecting the production efficiency and energy saving of power plants, the authenticity and accuracy of the reported data of power plant power consumption rate are very important. Therefore, a method of abnormal data identification based on predictive model is proposed. Firstly, the NARX neural network is improved by Adaboost to build the forecasting model of the power consumption rate. Then, the present power consumption rate is forecasted in a dynamic way by continuously introducing the reported value. When the time series of the power consumption rate suddenly changes, the residual time series obviously increase or decrease. Then the outliers of each residual vector group are obtained by using the isolation forest algorithm. Finally, the method is used to identify the actual generation data injected with false data, and the effectiveness of the proposed method is verified.