Data Augmentation for Data-Driven Methods in Power System Operation: A Novel Framework Using Improved GAN and Transfer Learning

Power system dispatching and operation rely on massive operation data, especially for data-driven methods. However, power systems often face challenges with insufficient or imbalanced data, significantly affecting the accuracy of power system analysis. To address these challenges, this research introduces a novel data augmentation framework specifically designed to efficiently generate high-quality target data to enhance data-driven methods in power system operation. First, this framework decomposes the task of generating data considering multiple objectives into feature generation and combination, enabling efficient data generation with broad applicability. Then, an improved generative adversarial network integrated with transfer learning is proposed, which significantly improves the data generation performance even with limited data. Furthermore, a method based on the least absolute shrinkage and selection operator (LASSO) is presented to select and combine the critical features from massive variables considering different objectives, improving the method's applicability especially for complex power systems. Through data generation, large amounts of target data can supplement the original dataset, thereby improving the performance of data-driven methods. The case studies on the IEEE 39-bus system and the realistic 300-bus system demonstrated the efficacy of the proposed method, showcasing its capability to significantly enhance power system operation.