Wind Power Scenario Generation based on the Generalized Dynamic Factor Model and Generative Adversarial Network
Young-ho Cho, Hao Zhu, Duehee Lee · 2025
Read Wind Power Scenario Generation based on the Generalized Dynamic Factor Model and Generative Adversarial Network by Young-ho Cho, Hao Zhu, Duehee Lee (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-00692.txt.
Introduction
Young-ho Cho, Hao Zhu, and Duehee Lee combine in 2025 two modeling tools to fake realistic wind power futures for grid planning. Adequacy studies need many long scenarios for scattered wind farms that capture shared weather, waveform shape, and ramp behavior. A factor model extracts the shared spatial factors well but mimics waveforms poorly, while a generative network fakes waveforms well.
Their hybrid lets the network supply the filter that pulls dynamic factors from real data, then feeds those factors into the factor model. Tests on Australian wind data beat both parent methods and other rivals at matching the real statistics of wind output.
Catalog record
| CREATOR | Young-ho Cho, Hao Zhu, Duehee Lee |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 54 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2508-00692
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Wind Power Scenario Generation based on the Generalized Dynamic Factor Model and Generative Adversarial Network Young-ho Cho, Hao Zhu, Duehee Lee Abstract For conducting resource adequacy studies, we synthesize multiple long-term wind power scenarios of distributed wind farms simultaneously by using the spatio-temporal features: spatial and temporal correlation, waveforms, marginal and ramp rates distributions of waveform, power spectral densities, and statistical characteristics. Generating the spatial correlation in scenarios requires the design of common factors for neighboring wind farms and antithetical factors for distant wind farms. The generalized dynamic factor model (GDFM) can extract the common factors through cross spectral density analysis, but it cannot closely imitate waveforms. The GAN can synthesize plausible samples representing the temporal correlation by verifying samples through a fake sample discriminator. To combine the advantages of GDFM and GAN, we use the GAN to provide a filter that extracts dynamic factors with temporal information from the observation data, and we then apply this filter in the GDFM to represent both spatial and frequency correlations of plausible waveforms. Numerical tests on the combination of GDFM and GAN have demonstrated performance improvements over competing alternatives in synthesizing wind power scenarios from Australia, better realizing plausible statistical characteristics of actual wind power compared to alternatives such as the GDFM with a filter synthesized from distributions of actual dynamic filters and the GAN with direct synthesis without dynamic factors. I Introduction Resource adequacy means to maintain power system reliability by having sufficient capacity such that, even with failures or variability of resources, the probability of not being able to meet all load is sufficiently small [ 1 ] . System operators achieve resource adequacy of a power system by ensuring there is enough generation capacity [ 2 ] . In the case of intermittent energy resources, the effective load carrying capacity (ELCC) of the intermittent resource is the equivalent capacity of highly reliable generators that would result in the same probability of not being able to meet all load [ 3 ] . For example, the ELCC of wind power can be obtained by simulating power systems with long-term wind power scenarios with realistic ramping rates and marginal distributions [ 4 ] . Furthermore, the capacity factor and reserve margin contribution of wind power to the power system reliability can also be obtained by simulating a future power system by using realistic long-term wind power scenarios [ 5 ] . To calculate the ELCC of wind power and reserve margin of future power system with high penetration level of wind power, we should develop an algorithm that can synthesize as many long-term wind power scenarios as desired. Although we can calculate the ELCC by analyzing correlation and variability of historical wind data, when the duration of wind power is limited, and when not all combinations of observed characteristics are exhibited in the data [ 6 ] , we have to synthesize long-term wind power scenarios [ 7 ] . In order to be realistic, scenarios should match several long-term statistical characteristics of actual wind power. First, wind power scenarios should have the same marginal distributions as actual wind power, which represent possible deliverable amounts of energy [ 8 ] . In addition, scenarios should have temporal characteristics so that they can have similar ramping events [ 9 ] . For distributed wind farms, scenarios should be spatially and temporally correlated with each other depending on locations [ 10 ] . For long-term forecasting, the normalized scenarios should be synthesized to be scaled up to a future expected wind power capacity [ 11 ] . There have been many studies to model long-term statistical characteristics of wind power. The distribution of wind power was modeled to assess the potential amount of wind power at the given number of wind turbines in [ 12 ] . However, better stochastic characteristics could be obtained if the distribution of ramping events were represented as a function of ramping intervals. In [ 13 ] , the trend of spectral density of total wind power was modeled to assess the variability of future total wind power as more wind farms are aggregated. The relative sizes of ramping events of scenarios decrease as more wind farms are aggregated. However, if the spectral densities of the cross correlation among geographically spread wind farms were used, spatially and temporally correlated wind farms could be modeled while representing the decreasing relative sizes of ramping events under aggregation. In [ 14 ] , the spatial and temporal correlation between pairs of wind farms was modeled based on a static Pearson correlation coefficient. However, the correlation could be better modeled if time-varying correlation coefficients among wind farms were measured. Therefore, these characteristics of wind power should be further modeled through advanced statistical models in order to be reflected into wind power scenarios. Many traditional statistical models have been used to synthesize wind power scenarios. For example, [ 15 ] generated scenarios from a multivariate normal distribution; however, representing the complex distribution of ramping events is challenging when only marginal distributions are considered. Similarly, [ 16 ] utilized an autoregressive moving average (ARMA) approach to synthesize wind speed scenarios and then convert them into wind power scenarios that capture correlated ramp events, though ARMA models are generally limited to short-term movements. While methods such as applying Fourier transforms can improve the design of scenarios by emphasizing amplitude and duration, they remain constrained by the underlying distributional assumptions. An alternative advanced statistical method is the generalized dynamic factor model (GDFM), which constructs multivariate time-series models by extracting dynamic factors from the frequency-domain cross-correlations among multiple wind power signals [ 17 ] . In the GDFM framework, wind power waveforms are decomposed into a common component—representing concurrent movements across farms—and an idiosyncratic component—capturing site-specific variations. Although [ 18 ] demonstrated that the GDFM can effectively synthesize load and wind power scenarios for applications such as transmission planning, its nonmodular design limits the generation of diverse samples and the representation of marginal distributions without additional sampling strategies. Recently, generative adversarial networks (GANs) have emerged as a powerful nonparametric alternative for scenario generation. Unlike traditional statistical or factor‐based approaches that require an explicit a priori distribution, GANs learn to replicate complex, high-dimensional data distributions directly from historical observations. In a GAN framework, the generator network produces synthetic power output signals while the discriminator network evaluates their authenticity, iteratively refining the outputs through an adversarial process [ 19 , 20 ] . Several GAN variants have been applied to renewable energy scenario generation. For instance, [ 21 ] synthesized wind and solar power scenarios while incorporating spatial correlations, whereas sequence GANs [ 22 ] and deep convolutional GANs (DCGANs) [ 23 ] have further enhanced the modeling of temporal and spatial dependencies. Moreover, recent advances—such as conditional GANs with style-based enhancements (e.g., conditional StyleGAN2 as reported in [ 24 ] )—improve controllability by incorporating external conditions (e.g., weather data and forecast error metrics) into an implicit latent space, all without requiring a preset a priori distribution. Techniques such as the Wasserstein GAN with Gradient Penalty [ 25 ] and spectral normalization [ 26 ] further stabilize training and mitigate mode collapse. Despite these advances, many existing studies have