Ensemble adversarial training-based robust model for multi-horizon dynamic line rating forecasting against adversarial attacks

Abstract

Dynamic line rating (DLR) forecasting is critical in the effective and economical utilization of overhead lines (OHLs) in smart grids, which facilitates the integration of renewable energy sources and reduces infrastructure upgrade costs. The forecasting techniques used for DLR rely on weather data collected from sensors as well as data communication, which can introduce a potential vulnerability to adversarial attacks. Hence, this work utilizes extreme gradient boosting (XgBoost), categorical boosting (CatBoost), and random forest as ensemble learning techniques for multi-horizon forecasting, while investigating their vulnerability by introducing adversarial attacks using two different attack models with variable data contamination and perturbations. Additionally, ensemble adversarial training (EAT)-based countermeasure is proposed for robust and accurate DLR forecasting. Experimental results indicate the outperformance of the CatBoost method compared to XgBoost and random forest models under normal conditions, while highlighting the vulnerability of all models to adversarial attacks in terms of root mean square error (RMSE) and mean absolute percentage error (MAPE). The proposed CatBoost with EAT significantly mitigates the impacts of adversarial attacks and retains accuracy under normal conditions. This research contributes to developing an accurate, cyber-resilience, and reliable forecasting methodology for line rating technology, leading towards academic and industrial developments in smart grids.