Probabilistic Wind Power Modelling via Heteroscedastic Non-Stationary Gaussian Processes

Domniki Ladopoulou, Dat Minh Hong, Petros DellaportasarXiv 2025 · DOI · arXiv

Predicting how much power a wind turbine will generate isn't just about measuring wind speed; it’s about understanding the complex, "moody" relationship between the atmosphere and the machine. A new paper by Ladopoulou et al. introduces a sophisticated probabilistic framework that captures the nuances of wind power data better than traditional models, offering a more reliable way to manage renewable energy on the grid.

TL;DR

The authors propose a Heteroscedastic Non-Stationary Gaussian Process (GP) framework that simultaneously accounts for changing correlation structures and input-dependent noise in wind power data. By using a Generalised Spectral Mixture (GSM) kernel and neural-network-parameterized noise, the model significantly improves uncertainty quantification and predictive accuracy over standard stationary and non-GP baselines.

What Problem Does This Paper Address?

Wind power is notoriously variable. When looking at SCADA (Supervisory Control and Data Acquisition) data—the 10-minute averages recorded by turbines—two major challenges emerge:

  1. Non-stationarity: The "smoothness" of the relationship between wind speed and power changes depending on the regime. The physics of a turbine starting up (cut-in) are different from when it’s running at full capacity (rated).
  2. Heteroscedasticity: This is a fancy way of saying the "noise" or volatility isn't constant. In transitional wind speeds (5–10 m/s), power output fluctuates wildly, creating high noise. At very high speeds where the turbine is capped, the output is much more stable (low noise).

Standard Gaussian Processes (GPs) usually assume the world is "stationary" (the same patterns repeat everywhere) and "homoscedastic" (the noise level is a single constant). When these assumptions fail, models produce unreliable uncertainty estimates—either being too confident when they should be cautious, or vice versa.

Methods

The researchers built a GP model designed to "flex" with the data. The core innovations are:

  • Generalised Spectral Mixture (GSM) Kernel: Unlike standard kernels that have fixed parameters, the GSM kernel allows the amplitude, frequency, and length-scales (which determine smoothness) to be functions of the input wind speed. This allows the model to adapt its "shape" to different operating regimes.
  • Neural Network Parameterization: To make these functions learnable, the authors used compact neural networks to output the kernel parameters and the noise variance ($\sigma^2(x)$) for any given wind speed.
  • Smoothness Regularization: To prevent the model from becoming too erratic (overfitting), they added penalty terms to the loss function. This ensures that the changes in noise and smoothness happen gradually, reflecting physical reality.
  • MAP Inference: They used Maximum A Posteriori (MAP) inference with the AdamW optimizer, providing a balance between Bayesian rigor and computational efficiency.

Datasets

The model was evaluated using real-world SCADA data from the Kelmarsh wind farm in the UK.

  • Source: A Senvion MM92 turbine.
  • Scope: Over 1.7 million records, though the authors used a representative subset of 7,000 points for training to maintain computational feasibility (GPs are mathematically "heavy").
  • Preprocessing: Crucially, the data was filtered to remove "non-operational" periods like maintenance or sensor faults, ensuring the model learned the actual physics of power generation.

Key Insights and Ideas

  • Joint Modelling is Key: The biggest "aha" moment is that you can't just fix one problem. Models that were only non-stationary or only heteroscedastic performed worse than the proposed model that tackled both.
  • GPs vs. Deep Learning: While Deep Learning (like BNNs) is popular, this paper shows that for structured physical problems like wind power, GPs are often more data-efficient and provide more "interpretable" uncertainty.
  • The "Transitional Gap": The model correctly identified that uncertainty is highest during the "ramp-up" phase of the power curve. This is vital for grid operators who need to know not just the expected power, but the risk of a sudden drop or surge.

Key Results

The proposed Heteroscedastic GSM-GP outperformed all baselines, including RBF-GPs, Spectral Mixture GPs, and Bayesian Neural Networks (BNNs).

  • Probabilistic Accuracy: It achieved the lowest CRPS (Continuous Ranked Probability Score) and NLPD (Negative Log-Predictive Density), indicating that its predicted probability distributions were the most accurate.
  • Calibration: In the 95% predictive interval tests, the model's intervals actually contained nearly 95% of the test data, whereas simpler models were often "off" by significant margins.
  • Point Predictions: Even in simple metrics like RMSE (Root Mean Square Error), the flexible GP outperformed the others, proving that better uncertainty modelling actually leads to better "best guesses" too.

Limitations and Future Work

The authors are transparent about a few hurdles:

  • Symmetry Constraints: GP posteriors are inherently symmetric (Gaussian). However, real wind power is often "skewed" (e.g., you can't produce more than the turbine's maximum capacity). The model occasionally underestimates the probability of extreme values because of this.
  • Scalability: Standard GPs scale cubically ($O(N^3)$) with the number of data points. To use this on years of data from an entire wind farm, the authors suggest moving toward Sparse or Variational GP approximations.
  • Input Dimensionality: This study focused primarily on wind speed. Future iterations could incorporate wind direction, turbulence intensity, and ambient temperature to further refine the "power curve."

Paper Citation: Ladopoulou, D., Hong, D. M., & Dellaportas, P. (2025). Probabilistic Wind Power Modelling via Heteroscedastic Non-Stationary Gaussian Processes. arXiv:2505.09026v2.

Summary updated

Abstract

Accurate probabilistic prediction of wind power is crucial for maintaining grid stability and facilitating the efficient integration of renewable energy sources. Gaussian process (GP) models offer a principled framework for quantifying uncertainty; however, conventional approaches typically rely on stationary kernels and homoscedastic noise assumptions, which are inadequate for modelling the inherently non-stationary and heteroscedastic nature of wind speed and power output. We propose a heteroscedastic non-stationary GP framework based on the generalised spectral mixture kernel, enabling the model to capture input-dependent correlations as well as input-dependent variability in wind speed-power data. We evaluate the proposed model on 10-minute supervisory control and data acquisition (SCADA) measurements and compare it against GP variants with stationary and non-stationary kernels, as well as commonly used non-GP probabilistic baselines. The results highlight the necessity of modelling both non-stationarity and heteroscedasticity in wind power prediction and demonstrate the practical value of flexible non-stationary GP models in operational SCADA settings.
BibTeX
@misc{ladopoulou2025probabilistic,
  author = {Domniki Ladopoulou and Dat Minh Hong and Petros Dellaportas},
  title = {{Probabilistic Wind Power Modelling via Heteroscedastic Non-Stationary Gaussian Processes}},
  year = {2025},
  eprint = {2505.09026},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2505.09026},
}
Figure