Probabilistic and Explainable Machine Learning for Tabular Power Grid Data

Modeling power grid frequency stability is becoming increasingly challenging due to the integration of renewable energy sources.Machine learning approaches, such as gradient-boosted trees, have shown promise in analyzing the complex characteristics of power systems.However, these models are inherently deterministic, providing only point estimates.Meanwhile, the task of capturing the underlying uncertainty, particularly through (deep) probabilistic models, is still underexplored, despite its potential to better account for the stochastic nature of power grid dynamics.In this paper, we first compare the performance of TabNet, a deep learning architecture designed for tabular data, to XGBoost for modeling power grid frequency stability.We then present TabNetProba: a probabilistic extension of TabNet, that enables uncertainty-aware estimates comparable to NGBoost.Using these (trained) models, we leverage explainable artificial intelligence (XAI) to analyze the drivers influencing grid stability and identify sources of uncertainty in two major European synchronous areas: Continental Europe and the Nordic region.Our results demonstrate that TabNetProba achieves competitive performance with state-of-the-art methods while providing reliable uncertainty estimates.We find that load and conventional generation ramps, as well as forecast errors, are the key quantities for modeling and explaining mean stability indicators in both synchronous areas.In Continental Europe, renewable generation emerges as a key factor in explaining model uncertainty, while in the Nordic region, load and generation features dominate uncertainty estimation, allowing for more reliable and interpretable stability estimates for modern power systems.