Abstract: Against the backdrop of intensifying global climate change leading to frequent extreme hydrological events, this study proposes a deep learning-based method for identifying and simulating abrupt wet-dry transition characteristics in runoff sequences. The innovative approach addresses critical limitations in existing methodologies, including low identification efficiency and insufficient simulation accuracy when confronting extreme inflow events, aiming to break through the technical constraints of conventional hydrological analysis methods. By establishing a coupled framework integrating feature extraction, scenario recognition and scenario generation, this method not only achieves precise identification of rapid transitions, but also enables extreme inflow scenario simulation under small-sample conditions through an enhanced Generative Adversarial Network (GAN). In the case study of Three Gorges Reservoir, this method demonstrates significant technical superiority: the identification accuracy of extreme inflow events has been increased to 98.2%, while successfully capturing historical rapid transition patterns between high and low flow periods. The framework proves capable of generating statistically authentic extreme runoff sequences that effectively expand the extreme inflow scenario library. These research outcomes provide reliable data support for reservoir flood control operations and drought emergency decision-making, substantially enhancing the predictability and reliability of watershed water security management.