16ENERGY Enhancing Storm-Induced Power Outage Prediction Using Synthetic Event Generation with GANs: Addressing Data Scarcity and Imbalance for Improved Model Performance

Storm-induced power outages pose significant challenges for utility companies, leading to extensive damage to power systems and substantial financial losses. The development of accurate Outage Prediction Models (OPMs) has become increasingly crucial for mitigating these impacts and ensuring the resilience of power systems. However, the effectiveness of OPMs is often constrained by the limited availability of high-quality data from recent storm-caused outages, particularly severe weather events such as hurricanes or major Nor'easters. This data scarcity presents a significant barrier to the application of advanced predictive models, such as Artificial Neural Networks (ANNs) and Deep Neural Networks (DNNs), which require large and diverse datasets to perform optimally. Furthermore, the imbalance in data representing various levels of storm severity exacerbates this issue, leading to suboptimal performance of OPMs in predicting outages from the most destructive storms. This study addresses this data scarcity issue by generating synthetic storm events using Generative Adversarial Networks (GANs). By extracting the variable features of extreme storms—primarily weather parameters—and feeding them into a GAN model, we create synthetic events that can be combined with other regional data (environmental and infrastructure parameters) used in OPM prediction. These synthetic events not only expand the training dataset, allowing for the use of more sophisticated predictive models, but also offer a means to simulate future storm scenarios under climate change conditions. This capability enables utility companies to better evaluate the potential impacts of future storms and assess the effectiveness of their resiliency measures. The quality of the synthetic events is assessed through visualization techniques and by integrating them into a high-performing OPM previously trained on real storm events in the electric grid service territory of Connecticut, USA. After identifying high-quality synthetic events, a new model is trained using both real and synthetic data. This model's performance is then compared against the original model to evaluate the benefits of incorporating synthetic events into outage prediction.