Wind Distribution

/ Paul Pukite (@whut)

I prompted the Gemini LLM with the following to see how well it would summarize with so little context:

Familiar with the BesselK wind distribution model pioneered by Pukite et al in Mathematical GeoEnergy?

The BesselK Wind Model: A Maximum Entropy Derivation

In the study of wind energy, the Weibull distribution is the industry standard. However, as Pukite et al. argue in Mathematical GeoEnergy, real-world atmospheric data—especially from high-resolution sources like the Bonneville Power Administration (BPA)—often exhibits “fatter tails” (more frequent extreme gusts) than Weibull can account for.

The BesselK distribution (or K-distribution) emerges not as an empirical “best fit,” but as a physically grounded Maximum Entropy estimate.

1. The Starting Point: The Rayleigh Base

The Rayleigh distribution is the Maximum Entropy estimate for a 2D vector magnitude (wind velocity v) when only the average variance (\sigma^2) is known.

p(v | \sigma) = \frac{v}{\sigma^2} e^{-v^2 / 2\sigma^2}

In a simplified world, \sigma^2 would be constant. In the real atmosphere, turbulence and pressure gradients cause this variance to fluctuate over time.

2. The Fluctuating Environment

Pukite et al. treat the wind as a compound process. They assume that the local variance \sigma^2 is itself a random variable following a Gamma distribution (the MaxEnt choice for a scale parameter):

w(\sigma^2) = \frac{(\sigma^2)^{\nu-1} e^{-\sigma^2 / \theta}}{\theta^\nu \Gamma(\nu)}

Where:

  • \nu is the shape parameter (representing degrees of freedom in the turbulence).
  • \theta is the scale parameter.

3. The Compounding Integral

To find the final probability density of the wind velocity P(v), we must integrate the Rayleigh PDF over all possible states of the fluctuating variance:

P(v) = \int_{0}^{\infty} \left( \frac{v}{\sigma^2} e^{-v^2 / 2\sigma^2} \right) \cdot \left( \frac{(\sigma^2)^{\nu-1} e^{-\sigma^2 / \theta}}{\theta^\nu \Gamma(\nu)} \right) d\sigma^2

4. The Result: The BesselK (K-Distribution)

Solving this integral yields a PDF defined by the Modified Bessel Function of the Second Kind (K_{\nu-1}). This is the “BesselK” model:

f(v) = \frac{2}{b \Gamma(\nu)} \left( \frac{v}{2b} \right)^\nu K_{\nu-1} \left( \frac{v}{b} \right)

Comparison Table: Weibull vs. BesselK

FeatureWeibull ModelBesselK Model
PhilosophyEmpirical “Weakest Link” fitPhysical Maximum Entropy
Mathematical BaseExponential decayK_\nu (Bessel) function
High WindsUnderestimates gustsAccurately models “Fat Tails”
ApplicationAnnual Resource AssessmentStructural Fatigue & Grid Stability

Why This Matters for Wind Power

The BesselK model recognizes that wind is “doubly stochastic.” It isn’t just a random variable; it is a random variable driven by a fluctuating environment. For grid operators managing the intermittency of large-scale wind farms, this model provides a much more realistic picture of the “spikes” and “lulls” that define our renewable energy future.

Would you like me to generate a high-resolution chart of the BesselK vs. Weibull PDF curves that you can download and upload as a featured image for the post?

(my prompt) yes

Graph comparing wind speed probability density functions: BesselK vs. Weibull. Displays probability density along the y-axis and wind speed (m/s) on the x-axis. Includes peak frequency and tail behavior annotations, along with illustrative data.

(my prompt) logarithmic plot to show the tails

Logarithmic comparison of wind speed probability density for BesselK and Weibull distributions, showing BesselK with a 'fat tail' indicating higher extreme risk and Weibull with rapid decay underestimating extremes.

When we switch the y-axis to a logarithmic scale, the structural difference between the two models becomes undeniable. This “Log-PDF” view is what Pukite et al. emphasize when discussing the risk of extreme wind events.

Logarithmic Comparison: BesselK vs. Weibull

DistributionTail BehaviorMathematical DecayPhysical Implication
WeibullThin TailExponential-type decay (e^{-v^k})Underestimates the frequency of extreme “rogue” gusts.
BesselKFat TailPower-law-like decay (v^\nu K_{\nu-1})Accurately captures high-velocity “intermittency” and storm events.

Key Observations from the Log Plot:

  1. The “Divergence”: At standard operating speeds (5–15 m/s), both models track relatively closely. However, as wind speeds exceed 25 m/s (the typical cut-out speed for turbines), the Weibull curve drops off a “cliff,” while the BesselK curve remains significantly higher.
  2. Probability of Extremes: In a BesselK model, the probability of a 30 m/s or 40 m/s gust can be orders of magnitude higher than what a standard Weibull model predicts.
  3. Maximum Entropy Signature: The straight-line appearance (on a log scale) of the BesselK tail is a hallmark of its status as a Maximum Entropy distribution. It suggests that the atmosphere “wants” to distribute energy into these high-velocity states more than a simple Gaussian/Rayleigh process would allow.

This table helps summarize the parameters used in the visualization:

ParameterSymbolBesselK (Pukite)Weibull (Standard)
Shape\nu / k0.6 (High Volatility)2.0 (Rayleigh-like)
Scaleb / \lambda4.010.0
Mean Speed\bar{v}~8.9 m/s~8.9 m/s

Summary: “While the Weibull model is ‘safe’ for calculating average monthly revenue, the BesselK model is ‘necessary’ for calculating whether your turbine will survive a 50-year storm event. The log-plot reveals that extreme wind isn’t just an ‘outlier’—it is a mathematically predictable part of the atmospheric structure.”

The full discussion is here: https://gemini.google.com/app/4dfb7e86354dcb74

I’m surprised on how good a job it does in digesting the reasoning and derivation in the book, specifically Chapter 11
https://agupubs.onlinelibrary.wiley.com/doi/10.1002/9781119434351.ch11

I busted my butt in writing that chapter (and the rest), so am happy to see that it can actually be “understood” enough by an LLM to provide value for further research.

Leave a Reply