Practical Methods of Defective Input Feature Correction to Enable Machine Learning in Power Systems

In this research work, three practical correction methods are proposed to mitigate the impact of defective input features in power system data measurement for machine learning (ML) applications. A well-trained ML tool may become ineffective due to defective input features, which may originate from measurement issues, such as monitor malfunction, cyberattack, communication failure, or others. It is crucial to correct defective input features to enable ML tools with desirable performances. This letter first introduces the mechanism of three correction methods, i.e., statistical-value-based method, minimal-error-based method, and DNN-based adaptive method. Then, the methods are validated via a deep neural network (DNN) case for power system stability enhancement. Validation results demonstrate that the adaptive method achieves the best performance, enabling the well-trained ML tool with a similar accuracy level to the case of no data defects. Although actual measurements may have various data issues challenging ML applications in power systems, the proposed methods show promise for addressing these challenges.