📊 statistics

Stochastic weather generators for high-frequency wind vector time series

This paper develops and evaluates machine learning models based on vector-quantized variational autoencoders to generate realistic, high-frequency surface wind vector time series for Lamont, Oklahoma, successfully capturing complex diurnal volatility patterns while noting limitations in reproducing extreme wind speed distributions.

Original authors: Mingshi Cui, Kevin Eng, Justin T. Greene, Zern Ke, Abolfazl Sodagartojgi, Zhiqiu Xia, Gemma E. Moran, Michael L. Stein

Published 2026-06-10

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mingshi Cui, Kevin Eng, Justin T. Greene, Zern Ke, Abolfazl Sodagartojgi, Zhiqiu Xia, Gemma E. Moran, Michael L. Stein

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a computer to act like a weather god, specifically one who controls the wind at a single spot in Oklahoma. But here's the catch: you don't just want the computer to guess the wind speed once an hour. You want it to predict how the wind behaves every single minute, capturing the sudden gusts, the lulls, and the way the wind direction spins like a top.

This paper is about building that computer brain. The researchers took 30 years of minute-by-minute wind data from a weather station in Lamont, Oklahoma, and used advanced machine learning to create a "stochastic wind generator." Think of this generator as a digital wind simulator that can create endless, realistic days of wind data that look and feel just like the real thing.

Here is a breakdown of how they did it and what they found, using simple analogies:

1. The Challenge: The Wind is Chaotic

Wind isn't a smooth, predictable river; it's more like a chaotic crowd of people running in different directions.

The Problem: Old weather models were like looking at a photo of the crowd once an hour. They missed all the running, shoving, and sudden stops.
The Goal: The researchers wanted a model that captures the "minute-by-minute" chaos. They focused on the month of June to avoid the confusion of changing seasons, essentially asking: "If we only look at June, can we teach a computer to mimic the wind perfectly?"

2. The Tool: A "Digital Translator" (Time VQ-VAE)

To handle this chaos, the team used a special type of AI called a Time Vector-Quantized Variational Autoencoder (Time VQ-VAE).

The Analogy: Imagine you have a massive library of wind patterns. To store them efficiently, you can't write down every single gust. Instead, you create a dictionary of "wind words."
- Step 1 (Tokenizing): The AI looks at the wind data and translates it into a sequence of these "wind words" (discrete tokens). It's like turning a complex symphony into a simple sheet of musical notes.
- Step 2 (The Generator): They trained a "translator" (a Bidirectional Transformer) to learn the rules of this language. It learns which "wind words" usually follow others.
- Step 3 (The Output): When they want new data, the AI writes a new sequence of "wind words" and then translates them back into actual wind speeds and directions.

3. The Secret Ingredient: Weather States

The researchers realized that sometimes the wind goes crazy because of specific events, like a thunderstorm or a sudden pressure drop.

The Analogy: Imagine you are predicting traffic. If you just look at the cars, you might miss that a parade is happening. But if you add a "Parade Flag" to your model, you can predict the traffic jam better.
The Result: They created 16 different "weather states" (like "Heavy Rain + Pressure Drop") and fed this information into the AI. They found that giving the AI these "flags" helped it generate more realistic days, especially when the wind was acting wild.

4. The Two Modes: Solo Days vs. Continuous Seasons

They tested the AI in two ways:

Solo Mode: Generating one random day at a time, completely independent of the day before.
Chain Mode: Generating a whole month where each day depends on the one before it (using the last hour of the previous day to start the next).

The Finding: The "Chain Mode" worked better at creating a continuous story, but the AI sometimes struggled to make the transition between days feel perfectly smooth.

5. Did It Work? The "Turing Test" for Wind

How do you know if the fake wind is good? You need a referee.

The Test: They trained a "detective" AI (a classifier) to look at real wind data and fake wind data and guess which is which.
The Result: The best generator was very hard to catch. The detective could only guess correctly about 75% of the time (which is much better than random guessing, but not perfect). This means the fake wind looked very similar to the real wind.
The Filter: Interestingly, the detective got better at spotting the fakes if they filtered out the slow, smooth parts of the wind and only looked at the fast, jittery parts. This told the researchers: "Our AI is good at the big picture, but it's a little too smooth when it comes to the tiny, crazy jitters."

6. The Weakness: The "Super-Gust" Problem

The biggest flaw in their digital wind god is that it is afraid of extremes.

The Analogy: Imagine a painter who is great at painting a calm sunset but refuses to paint a violent storm because they've never seen one in their training photos.
The Reality: The AI learned the "average" wind perfectly. However, when it came to the most extreme wind speeds (the top 0.01% of gusts), the AI rarely generated winds as strong as the real ones. It tended to stay within the "safe zone" of what it had seen before.
Why? The training data had very few extreme events (like tornadoes), so the AI didn't learn how to create them. It's a classic case of the AI playing it safe.

Summary

The researchers built a sophisticated machine that can generate realistic, minute-by-minute wind data for a specific location in Oklahoma.

What it's good at: Capturing the daily rhythm of the wind, the average speed, and the general "mood" of the weather.
What it struggles with: Creating the most violent, extreme wind gusts and perfectly smoothing the transition from one day to the next.

The paper concludes that while this tool is a huge step forward for simulating wind for things like wind energy or wildfire spread, we still need better ways to teach computers how to imagine the "impossible" extreme storms that haven't happened often enough in the training data.

Technical Summary: Stochastic Weather Generators for High-Frequency Wind Vector Time Series

Problem Statement
Surface wind vectors exhibit substantial variability on minute-to-minute time scales, a resolution critical for applications in wind energy, wildfire spread, and aviation. While extensive literature exists on estimating marginal wind distributions or modeling wind at coarse resolutions (e.g., hourly), capturing the complex temporal dynamics of minute-by-minute wind vectors remains a formidable challenge. Traditional parametric models, such as linear autoregressive processes or simple cosine functions for diurnal cycles, fail to capture the extreme volatility and intricate intraday variations observed in high-frequency data. Furthermore, existing methods often struggle to balance the fidelity of temporal dynamics with the granularity of temporal resolution. This paper addresses the gap in developing stochastic generators capable of producing realistic, high-frequency synthetic wind vector time series that preserve both global temporal structures and short-term stochastic volatility.

Methodology
The authors propose a deep generative modeling framework based on Time Vector-Quantized Variational Autoencoders (Time VQ-VAE) combined with MaskGIT (a bidirectional transformer). The methodology is structured in three stages:

Tokenization (VQ-VAE): Continuous wind vector data (Easterly and Northerly components) are encoded into discrete latent tokens. A VQ-VAE with a learned codebook ( $K=512$ vectors) maps high-dimensional time series inputs ($1440$ minutes $\times$ $2$ components) to a discrete latent space. This approach avoids restrictive distributional assumptions and explicit frequency decomposition, which the authors found problematic for preserving high-frequency wind dynamics.
Generator Training (MaskGIT): A bidirectional transformer is trained to predict the distribution of these discrete tokens. Unlike autoregressive models, MaskGIT generates an entire day's sequence simultaneously by predicting randomly masked tokens. This allows for better capture of global consistency and accelerated sampling.
Iterative Decoding: Synthetic days are generated via an iterative unmasking process. The model progressively refines the sequence over multiple steps, unmasking the most confident tokens at each iteration until the full sequence is reconstructed.

Experimental Design and Conditioning Strategies
The study utilizes over 30 years of high-quality minute-by-minute data from the ARM facility in Lamont, Oklahoma, restricted to the month of June to isolate diurnal dynamics from seasonal shifts. The authors evaluate six distinct generator configurations based on two variables:

Temporal Conditioning: "Independent" (generating single days without context) vs. "Consecutive" (generating a sequence of days conditioned on the final hour of the preceding day).
Weather State Incorporation:
- No Weather: Unconditional generation.
- Features: Weather states (derived from pressure changes and precipitation) are treated as an additional input channel during tokenization.
- Embedded: Weather states are mapped to learnable embeddings and incorporated into the transformer's conditioning stage.

Key Results and Evaluation
The generators were evaluated using a comprehensive suite of formal and informal methods:

Discriminative Evaluation: A bidirectional LSTM classifier was trained to distinguish between real and synthetic data. The best-performing model (Consecutive generation with "Embedded" weather states) achieved the lowest classification accuracy, indicating the highest realism. However, the discriminator performed significantly better on high-pass filtered data (focusing on high-frequency variations), suggesting that while macro-trends are well-captured, high-frequency stochastic details remain a challenge.
Energy Scores: Using the multivariate Continuous Ranked Probability Score (CRPS), synthetic data generally produced slightly higher (worse) energy scores than training data, with deviations being more pronounced during daytime hours.
Graphical and Statistical Analysis:
- Diurnal Patterns: The best generators accurately mimic diurnal changes in wind volatility and the general shape of wind vector distributions.
- Extreme Values: A consistent limitation across all models is the inability to reproduce the most extreme wind speeds observed in the data. The generators tend to produce maximum wind speeds that do not exceed the maximum observed in the training set, a known limitation of generative models concentrating within the support of training data.
- Stochastic Volatility: Quantile regression analysis of wind speed changes revealed that the "Embedded" generator successfully replicates the scale and structure of stochastic volatility (dependence on current state and recent history). In contrast, the "Features" approach suffered from systematic over-smoothing, underestimating the magnitude of high-frequency turbulence.

Significance and Claims
The paper claims that the proposed Time VQ-VAE framework represents a significant step forward in modeling high-frequency wind vectors, successfully capturing complex diurnal structures and stochastic volatility that standard time series models miss. The authors emphasize that:

Discrete Tokenization: Using discrete latent tokens via VQ-VAE is essential for preserving both global structure and high-frequency variability without the smoothing artifacts associated with frequency-domain decomposition or continuous latent spaces.
Conditioning on Weather States: Incorporating discrete weather states as learned embeddings ("Embedded" approach) yields superior results compared to treating them as raw features, likely due to more effective conditioning of the transformer on nonlinear interactions between atmospheric regimes and turbulence.
Limitations: The authors modestly acknowledge that their nonparametric approaches struggle to extrapolate beyond the support of the training data, specifically failing to generate the most extreme wind events. They note that the transition between consecutive days in the "Consecutive" models is not yet fully realistic, and the models do not yet capture long-range recurring patterns spanning multiple days (e.g., the repetitive patterns observed in June 2020).

The work concludes that while the best generators capture many complex features of observational data, reproducing extreme tails and ensuring seamless multi-day continuity remain open challenges requiring further methodological development, potentially involving robust loss functions or broader data sources.