Conditional Style-Based Generative Adversarial Networks for Renewable Scenario Generation

Day-ahead scenario generationof renewable power plays an important role in short-term power system operations due to considerable output uncertainty included. In this paper, a deep renewable scenario generation model using style-based generative adversarial networks followed by a sequence encoder network, is developed to generate accurate and reliable day-ahead scenarios directly from historical data through different-level scenario style controlling and mixing, thus achieving better characterization of renewable spatial-temporal dynamics. Meanwhile, the integration of meteorological information serving as conditions enables our model to precisely capture the complex diurnal pattern and seasonality difference of renewable power. From wind and photovoltaic power perspectives, the effectiveness of the proposed model is validated on two real-world datasets reflecting region aggregation level and distributed power station level respectively. Numerical results demonstrate the superiority of model performance through both the statistical and power system scheduling analysis, compared to three benchmarks.

I. Introduction

With the continuous depletion of fossil fuels and concerns about the environmental crisis, renewable energy generation like wind power and photovoltaic (PV) power has been extensively deployed by many countries for the advantage of abundance, inexhaustibility and cleanness [1]. However, some characteristics inherent to renewable power, e.g. intermittence and volatility, introduce significant uncertainty to the short-term power system operations [2], in particular, to the decision-making process of day-ahead unit commitment and real-time economic dispatch. The renewable uncertainty, especially from wind generation units, is driven by the complex spatial-temporal interdependencies among multiple geographically adjacent wind farms [3]. Therefore, it is imperative to account for more accurate and reliable prediction for uncertainty quantification, which is the prerequisite for large-sale renewable power integration and can benefit correct and effective decisions for power systems.