基于 Boosting 集成框架的新能源发电功率异常值检测方法

New energy power data makes an important part of the power big data, and its good quality provides important protections for the power forecasting, the load forecasting, the power grid planning and operation, the economic dispatch, and the demand-side response. The outliers in the new energy power data account for a minority, and the new energy power data belongs to the imbalanced data. Mostly a single method model is used in the traditional outlier detection process, and its detection accuracy is relatively low. And in the face of imbalanced data sets with a long-tailed distribution, the traditional single model is likely to cause over-fitting of the multi-type data due to its same-weight training mode, resulting in great reduction in the detection accuracy. Aiming at the shortcomings of the traditional single-model outlier detection methods, a new method based on the Boosting ensemble learning is proposed. The overall framework adopts a three-layer progressive training mode, in which the base classifier makes a preliminary judgment on the original data, the detection results of which constitute a balanced data with the same abnormal and normal quantities to be trained on the secondary classifier. The final classifier retrains on the divergent samples. The test results on the real wind power data show that the accuracy of the model based on the Boosting ensemble learning is greatly improved compared with several common single models, effectively solving the problem of low detection accuracy of the traditional single models on imbalanced data. Compared with the Bagging and the Stacking ensemble models, the model proposed has the best precision, recall rate and F1Score.