ABSTRACT Graph neural networks (GNNs) have shown promise in modelling complex graph‐structured data within power distribution systems. However, their application remains limited by the absence of uncertainty quantification techniques, which are crucial for risk‐aware decision‐making. This paper presents a deep ensemble framework for GNNs that generates uncertainty estimates for n‐1 contingency criterion classification predictions, an important task for distribution system operators (DSOs) to evaluate network reliability. Through evaluation of ensemble sizes up to 20 members using bootstrap sampling methodology, we demonstrate that ensemble configurations achieve substantial performance improvements over single models. Across the full 0%–100% range of confidence‐based filtering thresholds, the ensemble maintains consistently higher accuracy than a single GNN. The framework provides confidence levels for predictions, enabling data‐driven risk management for DSOs. This capability establishes ensemble GNNs as a reliable tool for uncertainty‐aware decision‐making in power system operations with validated performance guarantees.
Nooten et al. (Thu,) studied this question.