Anomaly Detection in Smart Grids using Machine Learning

Smart grid data can be analyzed for detecting abnormalities in many different areas such as cybersecurity, fault detection, electricity theft, etc. There is a strong case for the use of machine learning in anomaly detection. The raw grid data requires feature extraction. Anomalies can be defined as instances or changes in the smart grid data that are out of character concerning the average trend. A typical grid architecture results can vary significantly, depending on trends or changes in power, voltage, current, or consumption. This paper develops an anomaly detection model for a real-world smart grid system implemented on a hardware-based testbed. By detecting abnormal activities, one can improve the system behavior in data communication flow. It will also identify if there are parameter changes that indicate the presence of cyber-attacks. Our proposed anomaly detection model is build based on Isolation Forest (IF) to isolate outliers from standard observations through multiple decision trees. The performance of the proposed detection method was verified using the simulation results on a hardware-based testbed. Feature selection was optimized by principal component analysis and the model was further analyzed for performance with dickey-fuller test.