Model-Free Renewable Scenario Generation Using Generative Adversarial Networks

Yize Chen, Yishen Wang, Daniel S. Kirschen, Baosen Zhang

Abstract

Scenario generation is an important step in the operation and planning of power systems with high renewable penetrations. In this work, we proposed a data-driven approach for scenario generation using generative adversarial networks, which is based on two interconnected deep neural networks. Compared with existing methods based on probabilistic models that are often hard to scale or sample from, our method is data-driven, and captures renewable energy production patterns in both temporal and spatial dimensions for a large number of correlated resources. For validation, we use wind and solar times-series data from NREL integration data sets. We demonstrate that the proposed method is able to generate realistic wind and photovoltaic power profiles with full diversity of behaviors. We also illustrate how to generate scenarios based on different conditions of interest by using labeled data during training. For example, scenarios can be conditioned on weather events (e.g., high wind day) or time of the year (e.g., solar generation for a day in July). Because of the feedforward nature of the neural networks, scenarios can be generated extremely efficiently without sophisticated sampling techniques.

Introduction

As power grids increasingly rely on renewable energy sources, system operators must navigate the inherent intermittency of wind and solar power. Effective grid management—including unit commitment and storage management—requires "scenarios," which are sets of plausible future time-series data.

Traditional scenario generation methods are typically "model-based," relying on fitting historical data to predefined mathematical distributions like Gaussian or Copula models. However, these approaches often fall short when dealing with the nonlinear, bounded, and highly intermittent nature of renewable energy. Furthermore, scaling these models to account for spatial correlations across dozens of geographically dispersed sites is mathematically intensive and often requires restrictive assumptions that can lead to unrealistic scenarios.

Methodology: Generative Adversarial Networks

To overcome these limitations, we propose a "model-free" approach using Generative Adversarial Networks (GANs). Our framework utilizes two competing deep neural networks to learn the underlying distribution of renewable energy data without explicit statistical assumptions.

The GAN architecture consists of a Generator that transforms noise into power profiles and a Discriminator that distinguishes between real and generated samples.

The Adversarial Framework

  1. The Generator ($G$): This network takes random noise as input and uses de-convolutional layers to up-sample it into a structured 24-hour power profile.
  2. The Discriminator ($D$): Acting as a critic, this network uses convolutional layers to extract features from both real historical data and the synthetic data produced by the generator, attempting to classify them as "real" or "fake."

Through this adversarial process, the Generator learns to synthesize profiles that are indistinguishable from historical data. To improve training stability and capture the full diversity of renewable behaviors (avoiding "mode collapse"), we employ the Wasserstein GAN (WGAN) variant. Additionally, by incorporating labels—such as the month or specific weather events—into the training process (Conditional GANs), we can generate targeted scenarios on demand.

Key Results and Validation

We validated our approach using high-resolution 5-minute interval data from the NREL Integration Datasets, covering 24 wind farms and 32 solar plants in Washington State.

Comparison of real and generated wind/solar profiles alongside their autocorrelation coefficients, demonstrating high statistical fidelity.

Statistical and Temporal Fidelity

As shown in the results, the generated scenarios' Cumulative Distribution Functions (CDFs) and autocorrelation plots closely mirror historical data. This indicates that the GAN successfully captures both the magnitude of power production and the temporal "rhythm" of fluctuations. Power Spectral Density (PSD) analysis further confirmed that the model replicates the diurnal patterns essential for power system planning.

Spatial and Conditional Performance

The model demonstrated a remarkable ability to preserve spatial correlations between multiple sites. Even when noise was introduced to the training data, the GAN maintained the correlation matrix across 24 wind farms. Furthermore, the GAN outperformed standard Gaussian Copula methods in capturing extreme "ramp events"—sudden, sharp changes in power output that are critical for grid stability.

Conclusion

This GAN-based approach offers a scalable, efficient, and data-driven alternative to traditional scenario generation. By learning the implicit "physics" of weather patterns from raw data, it eliminates the need for complex statistical modeling. Once trained, the feedforward nature of the neural network allows for nearly instantaneous scenario generation, making it a powerful tool for real-time stochastic optimization in modern power grids.

Summary updated

Abstract

Scenario generation is an important step in the operation and planning of power systems with high renewable penetrations. In this work, we proposed a data-driven approach for scenario generation using generative adversarial networks, which is based on two interconnected deep neural networks. Compared with existing methods based on probabilistic models that are often hard to scale or sample from, our method is data-driven, and captures renewable energy production patterns in both temporal and spatial dimensions for a large number of correlated resources. For validation, we use wind and solar times-series data from NREL integration data sets. We demonstrate that the proposed method is able to generate realistic wind and photovoltaic power profiles with full diversity of behaviors. We also illustrate how to generate scenarios based on different conditions of interest by using labeled data during training. For example, scenarios can be conditioned on weather events~(e.g. high wind day) or time of the year~(e,g. solar generation for a day in July). Because of the feedforward nature of the neural networks, scenarios can be generated extremely efficiently without sophisticated sampling techniques.
BibTeX
@article{chen2018model,
  author = {Yize Chen and Yishen Wang and Daniel S. Kirschen and Baosen Zhang},
  title = {{Model-Free Renewable Scenario Generation Using Generative Adversarial Networks}},
  journal = {IEEE Transactions on Power Systems},
  year = {2018},
  doi = {10.1109/tpwrs.2018.2794541},
}
Figure