• Proposes the SK-CGAN architecture, which is suitable for generating day-ahead and intraday scenarios from limited samples. It combines scenario trees with conditional GANs to create diverse curves that closely mirror real-world situations. • Introduces CGAN guided by discrete scenario trees, which uses scenario trees with occurrence probability based on the statistical knowledge of volatility and random errors, as conditional inputs. This method explores a wider data space than single-point forecasts, boosting sample diversity to overcome data scarcity. • Analyzes the impact of random noise in SK-CGAN. Findings show that optimally reducing noise in SK-CGAN, unlike in standard CGAN, improves the coverage of generated scenarios without widening the forecast interval. Scenario generation plays a critical role in short-term power system operations with high renewable penetration. Data-driven scenario generation typically requires extensive sample data, however, due to confidentiality constraints or limited historical records—such as those associated with extreme weather scenarios—only small datasets may be available, thereby making credible scenario generation challenging. This paper proposes a combined methodology that integrates statistical knowledge and adversarial learning for few-shot renewable scenario generation. Specifically, the framework incorporates statistical knowledge that captures historical fluctuations and power prediction errors, together with conditional generative adversarial networks (CGANs), to generate accurate and reliable day-ahead or intraday look-ahead scenarios. This approach enables exploration of more diverse regions within the data space, generates a broader range of samples, and compensates for the lack of diversity resulting from limited datasets (e.g., one month or less). Case studies are conducted on a provincial power grid in China with abundant wind power resources. Compared with the traditional CGAN, the proposed methodology, when implemented with appropriate parameter settings, improves the coverage of the generated scenarios without increasing the corresponding power interval width.
Wang et al. (Sun,) studied this question.