A Comprehensive Review of Robust AI Techniques for Securing Deep Learning Models in Smart Grid Systems

Deep learning models now support load forecasting, fault detection, state estimation and renewable energy management in smart grids. Small, deliberately crafted changes to sensor data, known as adversarial attacks, can push these models to wrong answers and put grid reliability at risk. We reviewed twenty papers on adversarial attacks and defences in smart-grid and related energy systems, seventeen of them published from 2021 onwards, along with foundational work on FGSM, PGD and ensemble adversarial training. Attack success rates above 90% were reported against undefended detectors and adversarial training improved robustness considerably. Yet a defence trained against one attack often fails against another, most studies defend a single model and ensemble defences are rarely tested for sustainable grids. From these findings we draw four research gaps and outline a CNN–LSTM ensemble framework with adversarial training.