TY  - EJOU
AU  - Gao, Jinggeng 
AU  - Ma, Yanhong 
AU  - Niu, Wei 
AU  - Li, Chunhua 
AU  - Han, Zifen 
AU  - Guan, Xinyu 
AU  - Li, Xia 
AU  - Wang, Xiang 

TI  - A High-Risk Operational Scenario Generation Method for Renewable Energy Base Clusters Based on SR-WGAN-GP
T2  - Energy Engineering

PY  - 
VL  - 
IS  - 
SN  - 1546-0118

AB  - To address the scarcity of high-risk operating scenario samples in renewable energy base clusters and the limited capability of conventional generative models to preserve risk attributes and temporal structures, this paper proposes a risk-oriented scenario generation method based on an improved generative adversarial network. First, based on multi-source time-series data including wind power, photovoltaic power, power export demand, and adjustable resources, a window-level composite risk index is constructed. A continuity-based risk identification rule is then introduced to identify and screen representative high-risk samples, forming a training dataset for critical operating conditions. On this basis, the Wasserstein generative adversarial network with gradient penalty (WGAN-GP) is adopted as the basic framework, and a risk-consistency constraint is incorporated into the generator loss function to develop the Safety-Risk-driven WGAN-GP (SR-WGAN-GP) model. Finally, the generation performance of SR-WGAN-GP is compared with that of the conventional WGAN-GP in terms of composite risk score distribution, multi-order autocorrelation characteristics, and dimensionality-reduced sample-space distributions. The results show that the proposed model can effectively reproduce the main risk characteristics of real high-risk scenarios. The Wasserstein distance of the composite risk score distribution decreases from 0.0100 to 0.0074, representing a reduction of 26.0%. For the multi-order autocorrelation coefficients of net power export demand, the mean absolute error and root mean square error decrease by 54.48% and 47.57%, respectively, while the correlation coefficient increases from 0.9695 to 0.9843. Principal component analysis and t-distributed stochastic neighbor embedding show substantial overlap between the generated and real samples in the reduced-dimensional space, together with similar local clustering patterns. These findings demonstrate that incorporating operational risk information into the generation objective improves the preservation of target risk attributes and short-term temporal characteristics under limited high-risk samples, providing a risk-oriented scenario generation approach for scenario simulation and subsequent risk analysis of renewable energy base clusters.
KW  - Machine learning; performance evaluation; renewable energy sources; risk analysis; wind power generation

DO  - 10.32604/ee.2026.089122