Forecasting Renewable Energy Generation Scenarios Based on Multi-Agent Diverse GANs

In this paper, we propose an approach for renewable energy scenario forecasts based on an improved generative adversarial networks (GANs) with multiple generators. Basically, the proposed approach can capture almost all the renewable scenario patterns from historical measurements and generate more accurate future scenarios without the knowledge of any explicit model. More specifically, Multi-Agent Diverse GAN is proposed as a data-driven deep generative modeling framework to learn distinct patterns inherent with renewable power data, which lays the foundation of diverse scenario forecasts. Furthermore, an optimization problem is formulated and solved to adjust input noises after the GAN training process in order to generate future scenarios closer to the real situation. Experiment results show that the proposed method can capture renewable generation patterns in a more sophisticated way so as to gain diverse and more accurate scenario forecasts.