GAN-based data-driven risk assessment for carbon emission in renewable power systems

With the growing momentum toward carbon neutrality, evaluating the carbon emission capacity of power systems has become increasingly vital. This study introduces a novel risk assessment framework for carbon-power accommodation, leveraging advanced machine learning techniques. To overcome the limitations posed by sparse spatiotemporal data, semi-supervised learning and remote learning approaches are employed, while Generative Adversarial Networks (GANs) are utilized to enrich the dataset. A Graph Neural Network (GNN) is developed to model the attribution between power systems and carbon emissions by incorporating both grid topology and power flow data. Furthermore, a fine-tuned, multi-scale risk quantification strategy enhances the representation of imbalance risks. The Bayesian Dirichlet equivalent uniform (BDeu) metric is used to manage the coupling of multi-categorical influencing factors. A case analysis of the Zhejiang power grid validates the framework's applicability, emphasizing how renewable energy penetration, energy structure, and seasonal changes impact carbon emission capacity.