Reliable Detection of Cyber Threats in Smart Grids Using a Hybrid XGBoost and LSTM Framework
Abstract
The smart grid utilizes communication, measurement, and automation techniques that help to monitor the power system but increase vulnerability to cyber attacks. FDIA and DoS are two types of cyber attack that either tamper with measurements or limit access to information and therefore need a detection technique capable of capturing the dynamics of measurements. In this study, a hybrid classification model that combines XGBoost and LSTM is introduced to classify smart-grid observations as normal state, FDIA state, and DoS state. The XGBoost algorithm will be utilized for performing static feature analysis, and the LSTM will model the temporal dependencies based on the consecutive measurement windows. The output of the two algorithms is then merged using the weighted decision fusion method. The proposed scheme is evaluated based on data created based on IEEE 39-Bus New England power system test case, after applying preprocessing techniques such as missing values interpolation, min-max normalization, ten-sample windows, and 80% training and 20% testing ratio. Experimental results report an overall accuracy of 82.12%, with class-wise F1-scores between 0.81 and 0.82 and a false-positive rate of 5.1%. The proposed system is further combined with a real-time monitoring interface using Flask, email alerts through SMTP, and logging events. This monitoring system has achieved an average latency of 0.23 seconds per sample, 1.5 seconds average alert time, ability to handle up to 50 simultaneous sessions, and 99.7% average uptime during the experiment.