A dynamic spatiotemporal graph generative adversarial network for scenario generation of renewable energy with nonlinear dependence

Scenario generation plays an important role in addressing the uncertainty associated with renewable energy sources (RES) outputs in power systems. Existing generative adversarial network-based methods primarily focus on the representation of static linear dependence in scenarios, almost ignoring the dynamic nonlinear dependence among sources, which can easily lead to insufficient recognition of spatiotemporal features and ultimately ineffective scenarios. In this paper, a novel dynamic spatiotemporal graph generative adversarial network (DSTG-GAN) model is proposed to generate renewable scenarios with consistently high quality on complex nonlinear dependency characteristics. In this model, a multi-modal data fusion block is firstly developed to make comprehensive use of time-series, graph and explicit nonlinear dependence information. Then, a spatial–temporal feature extraction method using Chebyshev graph convolution with time-varying graph structures and multi-head attention mechanism is presented to simultaneously capture the deep dynamic and static features in RES data. Comparative experiments are conducted to validate the superior performance of DSTG-GAN model in capturing comprehensive spatiotemporal dependence of RES, and experimental results show the contributions of different blocks to improve the quality of generated scenarios.