Deep learning-based distributionally robust joint chance constrained distribution networks PV hosting capacity assessment

As distributed photovoltaic (PV) penetration in distribution networks (DNs) is increasing, it is essential to assess the PV hosting capacity (PVHC) to ensure the safe operation of DNs. This paper proposes a data-driven distributionally robust joint chance constrained (DRJCC) distribution networks PVHC assessment framework. Firstly, the spatiotemporal attention, projection, supervision, and Transformer architecture-based generative adversarial blocks are introduced to develop an augmented time series generative adversarial network (ATS-GAN), which, by integrating both supervised and unsupervised learning during the joint training process, better captures the spatiotemporal characteristics of PV and load power. Subsequently, leveraging the ATS-GAN, a Wasserstein metrics-based ambiguity set of PV and load power probability distributions is constructed, centered on the distributions induced by the generator neural network. Secondly, the DRJCC PVHC assessment model is proposed. A combination of the Bonferroni inequality and conditional value-at-risk approximation is adopted to transform the multivariate DRJCC model into a tractable conic formulation for efficient computation. Numerical results demonstrate that the proposed method effectively captures the spatiotemporal characteristics and uncertainties of multivariate distributions under multiple constraints, significantly reducing the conservatism typically associated with distributionally robust individual chance constraints. • Proposing an augmented time-series GAN to model spatiotemporal uncertainties. • Proposing a deep learning-based ambiguity set for PV-load uncertainty distributions. • Proposing a deep learning-based DRJCC model for precise PV hosting capacity assessment. • Proposing a Bonferroni-CVaR DRJCC approximation method for efficient computation.