Quantifying the Influences on Probabilistic Wind Power Forecasts

Jens Schreiber, Bernhard SickarXiv 2018 · arXiv

Abstract

Probabilistic wind power forecasts (PF) are essential for managing grid stability in an era of increasing renewable energy integration. However, these models often function as "black boxes," providing little insight into how specific weather features drive predicted distributions. This paper applies variance-based sensitivity analysis to three popular machine learning models—Gradient Boosting Regression Trees (GBRT), Monotone Quantile Regression Neural Networks (MQRNN), and Support Vector Regression (SVR)—to quantify the influence of weather inputs across various terrains. Our findings reveal that model architecture, rather than terrain complexity, is the primary driver of sensitivity patterns, with GBRT demonstrating superior interpretability and accuracy.

Introduction

As renewable energy becomes a cornerstone of the global energy mix, grid operators require more than point forecasts; they need to understand the uncertainty inherent in wind power generation. Probabilistic forecasts address this by predicting a range of possible outcomes (a probability distribution). Despite their utility, most research focuses on accuracy metrics rather than interpretability. This study aims to transform these "black-box" models into "gray boxes" by identifying which input features drive forecasts and how those inputs interact across different predicted quantiles.

Methodology

The study evaluates three non-parametric machine learning models adapted for quantile regression:

  1. GBRT (Gradient Boosting Regression Tree): An ensemble method utilizing multiple decision trees.
  2. MQRNN (Monotone Quantile Regression Neural Network): A neural network architecture designed to prevent quantile crossing.
  3. SVR (Support Vector Regression): A kernel-based method adapted for quantile estimation.

To analyze these models, the authors employ Variance-Based Decomposition (VBD), a global sensitivity analysis technique. Two primary indices are calculated for each input feature:

  • First-order index ($S_i$): Measures the direct contribution of a single feature to the output variance.
  • Total-order index ($S_{T_i}$): Measures the contribution of a feature including all its interactions with other variables.

Crucially, this analysis is applied to each predicted quantile individually, revealing how feature importance shifts from low-power to high-power scenarios.

Datasets and Experimental Setup

The analysis utilizes data from 28 wind parks in Germany (2015–2016), categorized by terrain type:

  • Non-complex terrain (NCT): 11 flat-land parks.
  • Complex terrain (CT): 8 parks in forested or hilly areas.
  • Offshore (OS): 4 ocean-based parks.

Inputs include Numerical Weather Prediction (NWP) data—such as wind speed at 10m and 100m, air pressure, and humidity—alongside turbine technical meta-features.

Key Results and Discussion

Dominance of Wind Speed

As expected, Wind Speed at 100m (WS100m) emerged as the most critical feature across all models and scenarios. In the sensitivity table below, Row 9 (WS100m) consistently shows the highest indices, particularly for the GBRT model.

Table 1: Sensitivity indices (Si and STi) for various features across models and terrains. Row 9 represents Wind Speed at 100m, the dominant predictor.

Model Architecture vs. Terrain

A significant finding is that model architecture has a more profound impact on sensitivity results than terrain type. Models of the same type exhibited similar sensitivity patterns regardless of whether the wind park was offshore or in complex hilly terrain.

Quantile Dependency and Feature Interactions

The influence of features is not static across the probability distribution. For instance, in GBRT, the sensitivity to wind speed peaks at median quantiles and tapers off at the extremes. Furthermore, GBRT showed very low interaction between features, meaning its predictive power is derived directly from individual variables. In contrast, SVR exhibited massive feature interactions, suggesting it may over-complicate the relationship between weather and power, leading to its higher error rates (CRPS 0.073 compared to GBRT's 0.061).

Figure 1: Comparison of first-order sensitivity indices (S1) across quantiles for GBRT (left), MQRNN (middle), and SVR (right). Note the distinct patterns and varying levels of "noise" from secondary features.

Conclusion

By applying variance-based sensitivity analysis, this research provides a framework for interpreting probabilistic wind power models. GBRT stands out as the most effective model, offering both the highest accuracy and the cleanest interpretability due to its low feature interactions. Future work will explore multi-task learning to stabilize forecasts and investigate the potential for generalized models that use a universal set of essential features across all terrain types.

Summary updated

BibTeX
@misc{schreiber2018quantifying,
  author = {Jens Schreiber and Bernhard Sick},
  title = {{Quantifying the Influences on Probabilistic Wind Power Forecasts}},
  year = {2018},
  eprint = {1808.04750},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1808.04750},
}
Figure