Robust Photovoltaic Power Forecasting Against Multi-Modal Adversarial Attack via Deep Reinforcement Learning

With the increasing applications of deep learning and external multi-modal data on photovoltaic (PV) power forecasting, cyberattacks, especially false data injections, can remarkably mislead forecasting methods, threatening the secure and economic management of power grids. Developing accurate and robust PV power forecasting methods is of great importance. Current studies have yet focused on the impact of multi-modal attacks and fell short of responding to unperceivable attacks. Therefore, we proposed a novel robust PV power forecasting framework. A multi-modal adversarial attack fully utilizing the multi-modal correlations was executed in the proposed framework to simulate a potential false data injection. To recover from attacks, we adopted deep deterministic policy gradient to dynamically distribute weights for each modal to mitigate the effects of data poisoning and utilize valuable information from multi-modal inputs. Within the framework, actor and environment were pretrained to facilitate convergence and generalization. As revealed by the comparisons against other state-of-the-art methods, with input perturbance under 5%, a mere 0.053 kW increase in mean absolute error was observed, which was remarkably less than that observed with no robustness methods as 0.207 kW. The experimental results indicated the effectiveness of the proposed framework on improving the robustness of multi-modal PV power forecasting.