Machine Learning Methods Integrated Analysis for Detecting False Distribution Power System Data Injection Attack under High Proportional Renewable Energy Integration

With the increasing proportion of renewable energy in power system, the data volume exchange between the renewable energy and energy management system (EMS) become huge and various types of data appear in the distribution power system, which bring security and stability challenges to the system operation. Different from the measured data in the traditional power system, the data measured in the distribution power system with highly proportional renewable energy integration shows highly related with time and presents random and intermittent in one day, which brings the opportunity of false data injection attack (FDIA) by manipulating the measurement data in the power monitoring system that causing the no response or wrong response of power devices. In order to solve these challenges, many machine learning-based methods are proposed. However, the characteristics of these methods, such as precision, are different and present different potential applications in FDIA detection. In order to assess the performance of typical machine learning-based FDIA detection methods used in distribution power system with highly proportional renewable energy integration, the Convolutional Neural Network (CNN), Support Vector Machine (SVM), Recurrent Neural Network (RNN) and Autoencoder (AE) is analyzed and the decisive indexes, such as F1-score, precision and recall, are compared to show their potential application scenarios. The actual current data measured from a real distribution power system in northern of China is used for the FDIA methods comparison. The experiment results indicate that CNN and SVM outperformed RNN and AE in detecting FDIA, suggesting that they are better suited for time series data analysis in power system.