Power system extreme weather scenario identification: Avoiding mode collapse in meteorological small sample diverse generation

The integration of fluctuating renewable energy sources has heightened power output volatility, particularly during extreme weather, exacerbating supply-demand imbalances. Short-term power balance analysis requires accurate scenarios that account for both typical and extreme weather. However, manually identifying these conditions from large datasets is time-consuming and error-prone due to the rarity of extreme weather events. This paper proposes a data-driven framework to identify extreme weather scenarios in power systems, addressing the common issue of mode collapse in small sample generation. The framework employs a multi-generator-evaluator generative adversarial network (MGE-GAN) to improve the diversity and authenticity of generated scenarios. It first explores the causes of mode collapse and then introduces a two-stage identification approach. In the upstream stage, multiple generators enhance sample diversity, while an evaluator ensures unique information capture. In the downstream stage, a convolutional neural network-based classifier identifies extreme weather scenarios post-generation. A case study based on a real-world weather dataset is conducted. Comparisons with the cutting-edge counterparts reveal that the designed MGE-GAN effectively mitigates the mode collapse. The downstream stage further validates the diversity and authenticity of the data generated by the proposed model.