An XAI framework for robust and transparent data-driven wind turbine power curve models

TL;DR

This study tackles the opacity of data-driven wind turbine power-curve models under non-stationary conditions by introducing an XAI framework that contrasts learned strategies with physics-informed baselines. Using Shapley-value attributions and a global similarity metric $R^2_{phys}$, the approach assesses physical plausibility alongside conventional RMSE to guide model selection and data processing decisions. The results show that models with physically plausible strategies generalize better in out-of-distribution settings, with case studies demonstrating both the improvement of physically meaningful behavior and the ability to explain deviations in turbine output. The work provides a practical framework and open-source Python implementation to train, select, and deploy more transparent and robust data-driven wind power curves.

Abstract

Wind turbine power curve models translate ambient conditions into turbine power output. They are essential for energy yield prediction and turbine performance monitoring. In recent years, increasingly complex machine learning methods have become state-of-the-art for this task. Nevertheless, they frequently encounter criticism due to their apparent lack of transparency, which raises concerns regarding their performance in non-stationary environments, such as those faced by wind turbines. We, therefore, introduce an explainable artificial intelligence (XAI) framework to investigate and validate strategies learned by data-driven power curve models from operational wind turbine data. With the help of simple, physics-informed baseline models it enables an automated evaluation of machine learning models beyond standard error metrics. Alongside this novel tool, we present its efficacy for a more informed model selection. We show, for instance, that learned strategies can be meaningful indicators for a model's generalization ability in addition to test set errors, especially when only little data is available. Moreover, the approach facilitates an understanding of how decisions along the machine learning pipeline, such as data selection, pre-processing, or training parameters, affect learned strategies. In a practical example, we demonstrate the framework's utilisation to obtain more physically meaningful models, a prerequisite not only for robustness but also for insights into turbine operation by domain experts. The latter, we demonstrate in the context of wind turbine performance monitoring. Alongside this paper, we publish a Python implementation of the presented framework and hope this can guide researchers and practitioners alike toward training, selecting and utilizing more transparent and robust data-driven wind turbine power curve models.

Figures (16)

• Figure 1: Power of wind (light blue, dashed) vs. wind speed and a turbine's power curve under standard conditions (dark blue), from which the power coefficient ($c_p$) can be derived. Measured data points are also displayed (grey markers). Additionally, the four distinct operational regions are marked.
• Figure 2: Proposed automated ML models strategy validation pipeline. Next to conventional error metrics, such as the root mean squared error (RMSE), we obtain the distance between model strategies as a second criterion for model selection (each marker in the decision plane represents a model).
• Figure 3: Model performance and strategy beyond input distributions for all turbines and model types. Left: model performance on full-year test set when trained and validated on the period indicated on the x-axis. Note, that models with only two weeks of training data outperformed the $Phys_{base}$ model. Right: $R^2_{phys}$ of models trained and validated on the respective periods. Note, that strategies get on average more consistent and closer to the physical model strategy with more training data available.
• Figure 4: Full-year test set performance of all models trained and validated with up to three months of data. Validation performance in three bins on the x-axis. The colour indicates models with more (green) and less (blue) physically reasonable strategies.
• Figure 5: Top: Results of model evaluation over all turbines (colours), models (shapes) and training runs (transparency) in terms of relative test set RMSE (x-axis) and strategy (y-axis). Details can be found in Table \ref{['tab:corr_phys']}. Bottom: Conditional distributions of attributions (mean as lines, range shaded) for the physics-informed model (grey) and the data-driven models with the lowest (red) and the highest (green) similarity to the physics-informed strategy (above, marked by arrows). The curves illustrate the wide range of adapted strategies and exemplify the findings presented above on the level of individual models.