Robust Load Forecasting towards Adversarial Attacks via Bayesian Learning

Abstract

Electric load forecasting is an essential problem for the power industry, which has a significant impact on power system operation. Currently, deep learning is proved to be an effective tool for load forecasting. However, those learning-based models are vulnerable towards adversarial attacks, which raises concerns about the robustness of load forecasting models. In this study, we propose a Bayesian training method to enhance the robustness of deep learning based load forecasting models towards adversarial attacks. We theoretically prove that the proposed method can improve the load forecasting robustness against various attacking objectives without compromising the prediction performance. An approximated training scheme is proposed to accelerate the training process so as to make the method better applied in practice. The experimental results show that such an approximation still yields higher robustness compared to four recently proposed benchmark robust forecasting methods while maintaining the prediction performance under no attack.