A WASSERSTEIN-GRADIENT FLOW APPROACH TO ENHANCING POWER FLOW DATASET QUALITY

Abstract

The application of artificial intelligence (AI) methods in power grid analysis necessitates the utilization of power flow datasets for model training. Presently, power flow data sources predominantly stem from offline simulations and real-time data collection. However, the accumulated online and offline power flow datasets have limitations that impede their direct suitability for AI model training. Online power flow data, collected during actual grid operations, offers a substantial volume of sample data. Nevertheless, this data distribution lacks uniformity and contains numerous redundant samples, falling short of the comprehensive coverage and clear boundaries required for effective analysis. On the other hand, offline power flow data, characterized by extreme operational scenarios, is manually curated and often situated at the stable boundaries of grid operations. While it possesses strong sample typicality and clear boundaries, the dataset's volume is limited and fails to represent the full spectrum of typical working conditions in grid operations. Addressing this challenge involves supplementing datasets to align with the distribution characteristics of offline analysis data. By doing so, the resulting dataset can fulfill both comprehensive coverage and clear boundary requirements. However, the methodologies for dataset adjustment that consider distribution characteristics remain underexplored, hindering the full exploitation of offline analysis data's distribution traits. This study delves into the development of advanced dataset adjustment methods that consider distribution characteristics. It aims to bridge the gap between online and offline power flow data, enabling the creation of comprehensive and boundary-clear datasets suitable for AI-driven power grid analysis. The proposed approach not only enhances the efficacy of AI methods in grid analysis but also offers a unique perspective on utilizing distribution characteristics in dataset adjustment. By addressing this gap in research, we contribute to the improved applicability of AI techniques in power grid analysis, optimizing grid performance and reliability.