To address the issues of insufficient coverage of extreme scenarios, the tendency of high-risk boundary features to be weakened during the clustering process, and the lack of dy-namic risk-preference adjustment in park-level integrated energy system (PIES) planning, an extreme risk-oriented sce-nario generation and planning framework is proposed. First, a scenario generation model based on a Wasserstein generative adversarial network with gradient penalty (WGAN-GP) is developed to expand extreme scenarios by learning the distri-bution characteristics of historical data. Second, a five-dimensional extremeness evaluation index system is es-tablished, and a Pareto-TOPSIS-based approach is employed to comprehensively evaluate both extreme and conventional scenarios. On this basis, a dual-track clustering strategy is proposed, in which extreme and conventional scenarios are clustered separately. Extreme scenarios are independently clustered using a max–min distance analysis method to pre-serve high-risk boundary features. Furthermore, an extreme scenario proportion index is introduced to balance system security and economic performance within the planning mod-el. Case studies demonstrate that the proposed method can effectively expand scenario boundaries and improve the cov-erage of unknown extreme scenarios. Compared with the planning scheme using only historical scenarios, the proposed method significantly reduces both load shedding amount and load shedding duration under the same extreme test scenarios, thereby verifying its effectiveness in enhancing energy supply resilience and adaptability to extreme risks.