Investigation of Artificial Intelligence Vulnerability in Smart Grids: A Case from Solar Energy Forecasting

The increasing integration of renewable energy sources, such as solar photovoltaic (PV), into the power grid has heightened the significance of accurate solar radiation forecasting for grid stability and energy management. Deep learning-based models have shown promise in improving the accuracy of solar radiation forecasts, but their vulnerability to adversarial attacks remains a largely unexplored area of research. This paper investigates the vulnerability of deep learning-based solar radiation forecasting models to imperceptible adversarial attacks, focusing on false data injection within a restricted input data region. Leveraging Bayesian optimization, we strategically craft subtle perturbations in the input data to target the local optimum with the largest error from the true value while ensuring the perturbations remain imperceptible. Our experiments validate the potency of these attacks, highlighting the critical need for improved model robustness and security in applications vital to energy infrastructure. We explore the transferability of these attacks across different models and rigorously evaluate their resilience under various environmental conditions.