Long-term scenario generation of renewable energy generation using attention-based conditional generative adversarial networks

Long-term scenario generation of renewable energy is regarded as an important part of the optimal planning of renewable energy systems. This study proposes a scenario generation method for generating long-term correlated scenarios of wind and photovoltaic outputs from historical renewable energy data. The generation of scenarios was divided into two processes: long-term yearly sequence generation and intraday scenario generation of wind-solar energy. In the long-term yearly sequence generation process, the k-means clustering algorithm and Markov chain Monte Carlo simulation method were developed to capture the seasonal and long-term features of wind and photovoltaic energies. Furthermore, an attention-based conditional generative adversarial network (ACGAN) was proposed to capture short-term features. An attention structure and conditional classifiers were developed to capture features in the generated scenarios. To accelerate the convergence process and improve the quality of the generated scenarios, a gradient penalty was included in the ACGAN model. Numerical case studies were conducted to verify the validity of the proposed method using a real-world dataset.