Data-driven efficient reliability evaluation of power systems with wind penetration: an integrated GANs and CE method

The conventional Monte Carlo simulation may not be efficient enough for reliability evaluation of composite power systems. The cross-entropy (CE) algorithm is a promising state-of-the-art fast sampling method, while it has not been well developed in this field due to the implicit probability distributions of penetrated renewable energies. Specifically, the CE sampling requires the distributions of interest to be explicit and parametric, while some preconceived probabilistic distribution functions (PDFs) such as the Weibull distribution of wind speed make the results to deviate from the reality sometimes. In this study, a data-driven efficient approach for reliability evaluation of power systems with wind penetration is proposed utilising generative adversarial networks (GANs) and CE sampling. The distributions of wind speeds in multiple wind farms are estimated by GANs considering their spatial correlation without any prior knowledge. With the trained generative network mapping from the explicit Gaussian noise to the raw wind speed data, the CE sampling is successfully enabled to efficiently sample the system states with implicit PDFs, which are associated with wind speeds and component failures. A real wind speed dataset and the RTS testing system are utilised to verify the proposed integrated method, including the accuracy of distribution estimation and reliability evaluation result, as well as the speed-up efficiency of sampling.