Stochastic Physics-Informed Deep Generative Network Scenario Generation: Application on Responsive Residential Load Management

This paper introduces a stochastic model for the optimal residential responsive loads power scheduling considering participants' satisfaction. In this context, firstly, a physical-informed based generative adversarial network (PI-GAN) network is designed for scenario generation with a high correlation with the actual data. In this network, conventional GANs are improved to learn spatial-temporal features of the residential loads and physics-informed concepts. To realize the spatial feature of the complex and highly nonlinear time series like residential loads, a residual convolutional neural network (Res-CNN) is considered to learn the spatial features, while the fully temporal features are realized by gated recurrent neural networks (GNN). Then, generated scenarios are used to cover the uncertainty associated with residential loads and provide the optimal results for responsive loads, including shiftable and curtailable loads. The numerical results on actual data in London, England, verify the effectiveness of the proposed stochastic framework and superiority by comparison with conditional GAN and improved version of GAN in scenario generations impact of stochastic demand response program.