Data-driven scenario generation of renewable energy production based on controllable generative adversarial networks with interpretability

Efficient and reliable scenario generation is of paramount importance in the modeling of uncertainties and fluctuations of wind and solar based renewable energy production for power system planning and operation in the presence of highly penetrated renewable sources. This paper proposes a data-driven method for renewable scenario creation by embedding interpretable manifold space in controllable generative adversarial networks (GAN). Without the need for laborious probabilistic modeling and sampling procedures, the proposed machine learning-based model can adaptively understand the inherent stochastic and dynamic characteristics of renewable resources. The generation of renewable patterns can be deliberately modified by embedding characteristic features with interpretability in latent input space. To address the controllable generation, the mutual information maximization and imitation learning sampling techniques are developed and incorporated into the existing GAN networks. The proposed approach is verified by the real-time series data of wind and solar energy generation profiles. The numerical results demonstrate that the proposed solution can achieve the controllable generation of scenarios covering various statistical characteristics and even create new generation patterns that are different from previous scenarios.